[{"data":1,"prerenderedAt":1602},["ShallowReactive",2],{"i-kinnu:logo":3,"i-kinnu:origami-folding":8,"tile-technology-artificial-intelligence-machine-learning":12,"i-lucide:chevron-right":1597,"i-lucide:menu":1600},{"left":4,"top":4,"width":5,"height":5,"rotate":4,"vFlip":6,"hFlip":6,"body":7},0,27,false,"\u003Cg fill=\"none\">\u003Cpath d=\"M0.046875 1.05555C0.046875 1.03541 0.048197 1.01579 0.0507438 0.996728C0.0987149 0.438619 0.586845 0 1.18194 0H25.4398C26.451 0 26.9575 1.171 26.2424 1.85585L15.7301 11.9243L1.31574 0.903476C1.17475 0.79568 1.01137 0.761884 0.859586 0.784111L26.2936 25.1441C27.0086 25.829 26.5022 27 25.4909 27H1.18194C0.555061 27 0.046875 26.5133 0.046875 25.9129V1.05555Z\" fill=\"currentColor\"/>\u003C/g>",{"left":4,"top":4,"width":9,"height":10,"rotate":4,"vFlip":6,"hFlip":6,"body":11},1000,236,"\u003Cg fill=\"none\">\u003Cpath fill-rule=\"evenodd\" clip-rule=\"evenodd\"\n    d=\"M193.68 38.2238C195.994 38.2238 197.87 40.0989 197.87 42.412V231.812C197.87 234.125 195.994 236 193.68 236H4.19013C1.87603 236 2.02305e-07 234.125 0 231.812V42.412C-2.02305e-07 40.0989 1.87603 38.2238 4.19013 38.2238H193.68ZM111.76 89.0072C111.685 87.9474 110.572 87.2905 109.608 87.7376L96.8872 93.641C95.7786 94.1554 95.702 95.7016 96.7545 96.3225L101.579 99.167C94.7045 109.365 90.5733 122.892 90.5732 137.642C90.5733 154.323 95.8569 169.439 104.416 179.945C105.301 181.032 106.9 181.196 107.987 180.311C109.075 179.426 109.238 177.828 108.353 176.741C100.621 167.25 95.6522 153.305 95.6521 137.642C95.6522 123.661 99.6138 111.051 105.963 101.754L110.456 104.403C111.508 105.024 112.826 104.21 112.74 102.991L111.76 89.0072ZM9.63194 136.286C9.14864 136.286 8.75684 136.678 8.75684 137.161C8.7569 137.644 9.14868 138.035 9.63194 138.035H17.2161C17.6993 138.035 18.0912 137.644 18.0912 137.161C18.0912 136.678 17.6994 136.286 17.2161 136.286H9.63194ZM22.6813 136.286C22.198 136.286 21.8062 136.678 21.8062 137.161C21.8063 137.644 22.1981 138.035 22.6813 138.035H30.2655C30.7487 138.035 31.1406 137.644 31.1406 137.161C31.1406 136.678 30.7488 136.286 30.2655 136.286H22.6813ZM35.7464 136.286C35.2631 136.286 34.8713 136.678 34.8713 137.161C34.8713 137.644 35.2631 138.035 35.7464 138.035H44.4973C44.9805 138.035 45.3724 137.644 45.3724 137.161C45.3724 136.678 44.9806 136.286 44.4973 136.286H35.7464ZM49.9977 136.286C49.5144 136.286 49.1226 136.678 49.1226 137.161C49.1226 137.644 49.5144 138.035 49.9977 138.035H57.5819C58.0651 138.035 58.4569 137.644 58.457 137.161C58.457 136.678 58.0651 136.286 57.5819 136.286H49.9977ZM63.0783 136.286C62.595 136.286 62.2032 136.678 62.2032 137.161C62.2033 137.644 62.5951 138.035 63.0783 138.035H70.6625C71.1457 138.035 71.5375 137.644 71.5376 137.161C71.5376 136.678 71.1457 136.286 70.6625 136.286H63.0783ZM76.1277 136.286C75.6444 136.286 75.2526 136.678 75.2526 137.161C75.2527 137.644 75.6445 138.035 76.1277 138.035H83.7119C84.1951 138.035 84.5869 137.644 84.587 137.161C84.587 136.678 84.1951 136.286 83.7119 136.286H76.1277ZM102.266 136.286C101.782 136.286 101.39 136.678 101.39 137.161C101.391 137.644 101.782 138.035 102.266 138.035H109.85C110.333 138.035 110.725 137.644 110.725 137.161C110.725 136.678 110.333 136.286 109.85 136.286H102.266ZM115.338 136.286C114.855 136.286 114.463 136.678 114.463 137.161C114.463 137.644 114.855 138.035 115.338 138.035H122.923C123.406 138.035 123.798 137.644 123.798 137.161C123.798 136.678 123.406 136.286 122.923 136.286H115.338ZM128.403 136.286C127.92 136.286 127.528 136.678 127.528 137.161C127.528 137.644 127.92 138.035 128.403 138.035H135.988C136.471 138.035 136.863 137.644 136.863 137.161C136.863 136.678 136.471 136.286 135.988 136.286H128.403ZM141.468 136.286C140.985 136.286 140.593 136.678 140.593 137.161C140.593 137.644 140.985 138.035 141.468 138.035H149.053C149.536 138.035 149.928 137.644 149.928 137.161C149.928 136.678 149.536 136.286 149.053 136.286H141.468ZM154.541 136.286C154.058 136.286 153.666 136.678 153.666 137.161C153.666 137.644 154.058 138.035 154.541 138.035H162.125C162.609 138.035 163 137.644 163.001 137.161C163.001 136.678 162.609 136.286 162.125 136.286H154.541ZM167.614 136.286C167.131 136.286 166.739 136.678 166.739 137.161C166.739 137.644 167.131 138.035 167.614 138.035H175.198C175.681 138.035 176.073 137.644 176.073 137.161C176.073 136.678 175.681 136.286 175.198 136.286H167.614ZM180.671 136.286C180.188 136.286 179.796 136.678 179.796 137.161C179.796 137.644 180.188 138.035 180.671 138.035H188.255C188.739 138.035 189.13 137.644 189.131 137.161C189.131 136.678 188.739 136.286 188.255 136.286H180.671Z\"\n    fill=\"currentColor\" />\n  \u003Cpath fill-rule=\"evenodd\" clip-rule=\"evenodd\"\n    d=\"M444.85 38.2277C447.164 38.2277 449.04 40.1028 449.04 42.4159V132.928C449.04 135.241 447.164 137.116 444.85 137.116H255.36C253.046 137.116 251.17 135.241 251.17 132.928V42.4159C251.17 40.1028 253.046 38.2277 255.36 38.2277H444.85ZM361.96 125.388C361.618 125.046 361.064 125.046 360.722 125.388L354.534 131.572C354.192 131.914 354.192 132.468 354.534 132.81C354.876 133.151 355.43 133.151 355.772 132.81L361.96 126.624C362.301 126.283 362.301 125.73 361.96 125.388ZM371.047 116.311C370.705 115.969 370.15 115.969 369.809 116.311L364.446 121.671C364.104 122.012 364.104 122.567 364.446 122.908C364.788 123.249 365.342 123.25 365.684 122.908L371.047 117.548C371.388 117.207 371.388 116.652 371.047 116.311ZM380.124 107.246C379.782 106.904 379.227 106.904 378.885 107.246L373.523 112.606C373.181 112.948 373.181 113.502 373.523 113.844C373.864 114.185 374.419 114.185 374.761 113.844L380.124 108.483C380.465 108.142 380.465 107.587 380.124 107.246ZM385.736 65.8841C385.891 64.6727 384.622 63.7845 383.536 64.3434L371.069 70.7636C370.124 71.2504 369.96 72.5334 370.752 73.2424L381.2 82.5938C382.11 83.4081 383.561 82.8672 383.717 81.6557L384.393 76.3725C391.143 77.1933 398.567 80.7709 404.771 86.9711C411.124 93.3213 414.726 100.952 415.43 107.827C415.573 109.221 416.819 110.236 418.214 110.093C419.609 109.95 420.624 108.703 420.481 107.309C419.644 99.1317 415.435 90.4514 408.362 83.3817C401.466 76.489 393.038 72.3185 385.038 71.338L385.736 65.8841ZM389.2 98.1733C388.859 97.8319 388.304 97.8318 387.962 98.1733L382.6 103.534C382.258 103.875 382.258 104.429 382.6 104.771C382.941 105.112 383.496 105.112 383.838 104.771L389.2 99.4108C389.542 99.0693 389.542 98.5149 389.2 98.1733ZM398.262 89.1047C397.92 88.7633 397.365 88.7632 397.024 89.1047L391.661 94.4649C391.319 94.8065 391.319 95.3608 391.661 95.7024C392.002 96.0436 392.557 96.0438 392.899 95.7024L398.262 90.3421C398.603 90.0007 398.603 89.4463 398.262 89.1047ZM416.431 70.9616C416.089 70.6202 415.534 70.6201 415.193 70.9616L409.83 76.3218C409.488 76.6634 409.488 77.2177 409.83 77.5592C410.172 77.9005 410.726 77.9007 411.068 77.5592L416.431 72.199C416.772 71.8575 416.772 71.3032 416.431 70.9616ZM425.508 61.891C425.166 61.5496 424.611 61.5495 424.27 61.891L418.907 67.2512C418.565 67.5928 418.565 68.1471 418.907 68.4887C419.249 68.8299 419.803 68.8301 420.145 68.4887L425.508 63.1284C425.849 62.787 425.849 62.2326 425.508 61.891ZM434.569 52.8146C434.227 52.4731 433.673 52.4731 433.331 52.8146L427.968 58.1748C427.626 58.5163 427.627 59.0706 427.968 59.4122C428.31 59.7534 428.864 59.7537 429.206 59.4122L434.569 54.052C434.91 53.7105 434.91 53.1562 434.569 52.8146ZM443.638 43.7479C443.296 43.4065 442.742 43.4064 442.4 43.7479L437.037 49.1081C436.695 49.4496 436.696 50.004 437.037 50.3455C437.379 50.6868 437.933 50.687 438.275 50.3455L443.638 44.9853C443.98 44.6438 443.979 44.0895 443.638 43.7479Z\"\n    fill=\"currentColor\" />\n  \u003Cpath fill-rule=\"evenodd\" clip-rule=\"evenodd\"\n    d=\"M684.066 38.2277C687.798 38.2281 689.667 42.7391 687.027 45.3773L596.473 135.889C595.687 136.675 594.621 137.116 593.51 137.116H506.335C504.021 137.116 502.145 135.241 502.145 132.928V42.4159C502.145 40.1028 504.021 38.2277 506.335 38.2277H684.066ZM514.603 124.566C514.261 124.224 513.707 124.224 513.365 124.566L507.178 130.751C506.836 131.093 506.836 131.646 507.178 131.988C507.519 132.329 508.073 132.329 508.415 131.988L514.603 125.803C514.945 125.462 514.945 124.908 514.603 124.566ZM523.689 115.491C523.348 115.15 522.794 115.15 522.452 115.491L517.09 120.852C516.748 121.193 516.748 121.747 517.09 122.088C517.431 122.43 517.985 122.43 518.327 122.088L523.689 116.728C524.031 116.386 524.031 115.833 523.689 115.491ZM532.102 65.8295C530.707 65.6872 529.46 66.7017 529.318 68.0957C529.175 69.4896 530.189 70.7355 531.584 70.8787C538.463 71.5825 546.096 75.1826 552.45 81.5329C558.723 87.8037 562.312 95.3226 563.079 102.13L557.738 102.392C556.518 102.452 555.865 103.855 556.607 104.827L565.115 115.969C565.76 116.814 567.051 116.751 567.611 115.847L574.992 103.928C575.635 102.889 574.848 101.555 573.628 101.615L568.161 101.882C568.161 101.878 568.162 101.874 568.161 101.871C567.324 93.6931 563.114 85.0124 556.041 77.9425C548.968 70.873 540.283 66.6668 532.102 65.8295ZM532.766 106.421C532.425 106.079 531.871 106.079 531.529 106.421L526.166 111.781C525.825 112.123 525.825 112.676 526.166 113.018C526.508 113.359 527.062 113.359 527.403 113.018L532.766 107.657C533.108 107.316 533.108 106.762 532.766 106.421ZM541.843 97.3445C541.501 97.003 540.948 97.003 540.606 97.3445L535.243 102.705C534.901 103.046 534.902 103.6 535.243 103.941C535.585 104.283 536.139 104.283 536.48 103.941L541.843 98.5809C542.185 98.2393 542.185 97.686 541.843 97.3445ZM550.92 88.2778C550.578 87.9363 550.025 87.9363 549.683 88.2778L544.32 93.638C543.978 93.9796 543.978 94.5329 544.32 94.8745C544.662 95.2161 545.215 95.2161 545.557 94.8745L550.92 89.5142C551.262 89.1727 551.262 88.6193 550.92 88.2778ZM569.066 70.1405C568.724 69.799 568.17 69.7991 567.829 70.1405L562.466 75.5008C562.124 75.8423 562.124 76.3956 562.466 76.7372C562.808 77.0788 563.361 77.0788 563.703 76.7372L569.066 71.377C569.407 71.0354 569.407 70.4821 569.066 70.1405ZM578.143 61.0699C577.801 60.7284 577.247 60.7285 576.906 61.0699L571.543 66.4302C571.201 66.7717 571.201 67.3251 571.543 67.6666C571.885 68.0082 572.438 68.0082 572.78 67.6666L578.143 62.3064C578.484 61.9648 578.484 61.4115 578.143 61.0699ZM587.219 51.9896C586.878 51.6481 586.324 51.6481 585.982 51.9896L580.62 57.3498C580.278 57.6914 580.278 58.2447 580.62 58.5863C580.961 58.9279 581.515 58.9279 581.857 58.5863L587.219 53.2261C587.561 52.8845 587.561 52.3312 587.219 51.9896ZM596.288 42.9249C595.947 42.5833 595.392 42.5833 595.05 42.9249L589.689 48.2851C589.347 48.6267 589.347 49.18 589.689 49.5216C590.03 49.863 590.584 49.8631 590.926 49.5216L596.288 44.1613C596.63 43.8198 596.63 43.2664 596.288 42.9249Z\"\n    fill=\"currentColor\" />\n  \u003Cpath fill-rule=\"evenodd\" clip-rule=\"evenodd\"\n    d=\"M850.814 38.2277C854.547 38.2281 856.416 42.739 853.777 45.3773L763.223 135.889C762.437 136.674 761.371 137.116 760.26 137.116H673.176C669.443 137.116 667.574 132.605 670.213 129.966L760.768 39.4544C761.554 38.6692 762.62 38.2277 763.731 38.2277H850.814ZM761.338 121.8C760.855 121.8 760.463 122.191 760.463 122.674V131.13H762.213V122.674C762.213 122.191 761.821 121.8 761.338 121.8ZM761.338 108.971C760.855 108.971 760.463 109.363 760.463 109.846V118.301H762.213V109.846C762.213 109.363 761.821 108.971 761.338 108.971ZM761.338 96.1402C760.855 96.1406 760.463 96.5321 760.463 97.0149V105.47H762.213V97.0149C762.213 96.532 761.821 96.1404 761.338 96.1402ZM782.263 71.887C781.043 71.951 780.395 73.3571 781.139 74.3257L784.474 78.6631C779.115 82.951 771.242 85.7443 762.35 85.7444C753.366 85.7442 745.421 82.8944 740.059 78.5305C738.972 77.6461 737.373 77.8099 736.488 78.8961C735.602 79.983 735.766 81.582 736.853 82.467C743.231 87.6574 752.348 90.8207 762.35 90.8209C772.209 90.8208 781.205 87.746 787.568 82.6884L790.833 86.9341C791.577 87.9025 793.103 87.6391 793.479 86.4767L797.791 73.138C798.118 72.127 797.33 71.1017 796.268 71.1566L782.263 71.887ZM761.338 70.4847C760.855 70.4851 760.463 70.8767 760.463 71.3594V79.8147H762.213V71.3594C762.213 70.8766 761.821 70.485 761.338 70.4847ZM761.338 57.656C760.855 57.6564 760.463 58.048 760.463 58.5307V66.986H762.213V58.5307C762.213 58.0479 761.821 57.6563 761.338 57.656ZM761.338 44.8293C760.855 44.8297 760.463 45.2212 760.463 45.704V54.1592H762.213V45.704C762.213 45.2211 761.821 44.8295 761.338 44.8293Z\"\n    fill=\"currentColor\" />\n  \u003Cpath\n    d=\"M995.759 38.2277C999.53 38.228 1001.42 42.5171 998.752 45.0253L959.55 81.9005L905.796 41.5363C905.271 41.1418 904.662 41.0182 904.096 41.0994L997.485 130.319C1000.15 132.828 998.262 137.116 994.491 137.116H905.298C902.96 137.116 901.065 135.333 901.065 133.134V42.0941C901.065 42.0204 901.07 41.9483 901.079 41.8786C901.258 39.8345 903.079 38.2277 905.298 38.2277H995.759Z\"\n    fill=\"currentColor\" />\n  \u003Cpath\n    d=\"M505.873 0C506.657 4.57042e-05 507.307 0.195499 507.823 0.587023C508.338 0.969046 508.596 1.53802 508.596 2.29251C508.596 2.76034 508.467 3.19015 508.209 3.58162C507.951 3.96344 507.497 4.26401 506.848 4.48361V4.54114C507.65 4.67487 508.205 4.96191 508.51 5.4012C508.816 5.83087 508.969 6.31772 508.969 6.86193C508.969 7.74056 508.672 8.41851 508.08 8.89604C507.497 9.38304 506.733 9.62731 505.787 9.62738C504.861 9.62738 504.158 9.42172 503.68 9.0111C503.212 8.60054 502.935 8.08005 502.849 7.44993L503.881 7.10571L503.924 7.24028C504.035 7.54934 504.211 7.82925 504.454 8.07986C504.731 8.36635 505.166 8.50986 505.758 8.50989C506.465 8.50989 506.943 8.32772 507.191 7.9648C507.449 7.6019 507.579 7.20078 507.579 6.7615C507.579 6.2173 507.378 5.80683 506.977 5.52992C506.585 5.25295 505.93 5.10026 505.013 5.07161V4.15402C505.901 4.12537 506.489 3.92484 506.776 3.55237C507.062 3.18009 507.206 2.82242 507.206 2.47876C507.206 1.62801 506.752 1.17539 505.845 1.12237L505.658 1.11749C505.467 1.11752 505.242 1.14605 504.985 1.2033C504.736 1.25105 504.511 1.3274 504.31 1.43245L504.081 2.56457L503.05 2.44951L503.322 0.687461C503.666 0.49653 504.068 0.33454 504.526 0.200875C504.985 0.0671945 505.434 0 505.873 0Z\"\n    fill=\"currentColor\" />\n  \u003Cpath\n    d=\"M905.727 2.30616L904.638 2.4066L904.466 1.26083H901.428V3.72497C901.533 3.71544 901.643 3.71034 901.757 3.71034H902.086C902.755 3.71034 903.386 3.78668 903.979 3.93949C904.58 4.09229 905.068 4.38363 905.44 4.8132C905.822 5.23335 906.014 5.84949 906.014 6.66106C906.014 7.64468 905.722 8.38068 905.14 8.86776C904.557 9.36434 903.783 9.6127 902.818 9.61275C901.91 9.61275 901.213 9.40711 900.725 8.99648C900.248 8.59544 899.96 8.08007 899.865 7.44993L900.911 7.10571C901.007 7.49723 901.203 7.8271 901.499 8.09449C901.795 8.37131 902.211 8.50985 902.746 8.50989C903.395 8.50989 903.869 8.33787 904.165 7.99405C904.461 7.65981 904.609 7.22507 904.609 6.69031C904.609 5.87861 904.337 5.3625 903.792 5.14279C903.248 4.91361 902.612 4.79958 901.886 4.79955C901.695 4.79955 901.489 4.80365 901.27 4.8132C901.059 4.82275 900.854 4.83701 900.653 4.85611L900.224 4.44071V0.143343H905.569L905.727 2.30616Z\"\n    fill=\"currentColor\" />\n  \u003Cpath fill-rule=\"evenodd\" clip-rule=\"evenodd\"\n    d=\"M765.49 6.04576H766.966L766.837 7.14862H765.49V9.48404H764.185V7.14862H759.857L759.713 6.04576L762.909 0.143343H765.49V6.04576ZM760.96 6.04576H764.185V1.26083H763.541L760.96 6.04576Z\"\n    fill=\"currentColor\" />\n  \u003Cpath d=\"M4.80573 6.47481H6.41154V7.60693H1.81068V6.47481H3.50235V1.27546H1.81068V0.143343H4.80573V6.47481Z\"\n    fill=\"currentColor\" />\n  \u003Cpath\n    d=\"M254.359 0C255.353 0 256.055 0.239186 256.466 0.716715C256.877 1.18447 257.083 1.68072 257.083 2.20573C257.083 2.85516 256.849 3.44346 256.38 3.96875C255.912 4.49397 255.348 4.96638 254.689 5.38657C254.039 5.79717 253.437 6.15968 252.883 6.47481H256.423L256.538 5.42948L257.599 5.51529L257.426 7.60693H251.407L251.292 6.58987C252.582 5.73032 253.638 4.98523 254.46 4.35489C255.281 3.71509 255.693 3.05632 255.693 2.37832C255.693 1.53787 255.166 1.11749 254.115 1.12237L254.115 1.11749C253.924 1.11754 253.695 1.14604 253.427 1.2033C253.16 1.25104 252.916 1.32238 252.697 1.41783L252.467 2.47876L251.45 2.3637L251.707 0.60165C252.118 0.401088 252.563 0.253475 253.041 0.15797C253.519 0.0529708 253.958 1.99446e-05 254.359 0Z\"\n    fill=\"currentColor\" />\u003C/g>",{"tile":13,"orbsWithOnlyMarkdownPages":637},{"id":14,"data":15,"type":16,"maxContentLevel":19,"version":20,"orbs":21},"9aa45ec4-8224-42b9-a187-4dee7e236a4f",{"type":16,"title":17,"tagline":18},9,"Machine Learning","Training like a brain",3,2,[22,84,154,253,356,448,547],{"id":23,"data":24,"type":20,"version":26,"maxContentLevel":19,"summaryPage":27,"introPage":34,"pages":41},"efeebd88-df18-427e-a761-59c3b12ebbf5",{"type":20,"title":25},"What is machine learning?",1,{"id":28,"data":29,"type":19,"maxContentLevel":19,"version":26},"7bcecddc-bbad-4f65-b366-e6c9828cb880",{"type":19,"summary":30},[31,32,33],"Machine learning is an approach to AI models which mimics human learning","An AI is said to 'learn' when it changing its behavior based on an input","Machine learning is often faster and more effective than traditional symbolic programming",{"id":35,"data":36,"type":37,"maxContentLevel":19,"version":26},"2126018d-eed4-42c9-875f-2e9e347043bc",{"type":37,"intro":38},10,[39,40],"What is the alternative to manual programming?","How did this alternative change the world?",[42,47,61],{"id":43,"data":44,"type":26,"maxContentLevel":19,"version":26},"b7466057-3b69-49b4-9f70-b8047e8c1ac8",{"type":26,"contentRole":20,"markdownContent":45,"audioMediaId":46},"Last time, we saw how symbolic AI has a major limitation. It relies on manual programming – and manual programming can only take us so far. In order to build more complex AIs, we need to use a different approach.\n\nBut what's the alternative to manual programming?\n\nTo answer that question, we can look at a human brain. Your own brain (obviously) wasn't manually programmed. Instead, you filled it with facts and ideas via a process of gradual learning. Taste an apple? That taste was added to your brain. Ride a bike? That skill was automatically stored.\n\nOver the last few decades, many scientists have taken the same approach to AI. They call this approach **machine learning**, and it's literally changed the world.","36eca252-558b-4f4a-aee3-e92f7dd6106c",{"id":48,"data":49,"type":26,"maxContentLevel":19,"version":26,"reviews":52},"33e34d80-c1b6-46af-9c10-e1e712f4652e",{"type":26,"contentRole":20,"markdownContent":50,"audioMediaId":51},"We'll get into the complexities of machine learning later, but here's a simple definition. An AI is said to 'learn' when you give it an input – like a big list of numbers, or some photos, or videos – and it changes its behavior afterwards. As we said, this is how humans learn as well. We experience an input, then we change.\n\nLet's run with that photo example. Imagine you've built an Artificial Intelligence that struggles to tell the difference between cats and dogs. You decide to show it a pile of animal photos and tell it over and over: \"this one’s a cat, this one’s a cat, this one’s a dog, this one’s a dog...\"\n\n![Graph](image://45419589-20de-4180-8724-3e3f8ab53ff2 \"Cats and dogs.\")\n\nYour AI starts to notice patterns in the photos, like the shape of the ears or the whiskers. And by the end, it's pretty good at recognizing cats and dogs on its own.","e21cdc57-0223-4607-a6a0-29876cb2f612",[53],{"id":54,"data":55,"type":56,"version":26,"maxContentLevel":19},"a7672cee-1b2a-4c12-8ad4-549f4641bd46",{"type":56,"reviewType":26,"spacingBehaviour":26,"activeRecallQuestion":57,"activeRecallAnswers":59},11,[58],"What do we call it when a computer receives an input, then changes its behavior?",[60],"Machine learning",{"id":62,"data":63,"type":26,"maxContentLevel":19,"version":26,"reviews":66},"e0961c14-1ff2-4be0-83c1-d96473a68c3a",{"type":26,"contentRole":20,"markdownContent":64,"audioMediaId":65},"Learning wouldn't be the only way to get an AI to recognize cats. Alternatively, a scientist could write symbolic rules (\"fluffy tail\", \"pointy ears\", \"long whiskers\", etc.), then manually build these rules into the AI's design.\n\nBut it's often a lot faster, and a lot more effective, to take a learning approach instead.\n\nLet's think about another example: an AI that's designed to translate text. Instead of manually programming all the words from every language, you can train the AI on a pile of texts which have already been translated. As it works through those texts, and compares the languages, it can learn all the different words.\n\nThis is exactly what Google did in 2016. They built an AI, then trained it on millions of translated texts. Whenever you load up Google Translate, you're interacting with an impressive, self-taught AI.\n\n![Graph](image://cd3c914a-7b0d-41a2-82eb-1a294c3cfeb8 \"Google Translate. Image (CC BY-SA 4.0) \u003Chttps://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons\")","aa7c5ab6-415f-4623-8abb-cf0f9bab7781",[67,76],{"id":68,"data":69,"type":56,"version":26,"maxContentLevel":19},"a633e54d-a520-4349-baab-de0c43fda69b",{"type":56,"reviewType":20,"spacingBehaviour":26,"binaryQuestion":70,"binaryCorrect":72,"binaryIncorrect":74},[71],"In 2016, how did Google build the AI model which powers Google Translate?",[73],"They trained it on millions of translated texts",[75],"They programmed it with the help of native speakers",{"id":77,"data":78,"type":56,"version":26,"maxContentLevel":19},"8305fa05-379a-4555-8af3-b246823b028f",{"type":56,"reviewType":20,"spacingBehaviour":26,"binaryQuestion":79,"binaryCorrect":81,"binaryIncorrect":82},[80],"When you want an AI to perform a task, which approach is generally faster?",[60],[83],"Manual programming",{"id":85,"data":86,"type":20,"version":26,"maxContentLevel":19,"summaryPage":88,"introPage":95,"pages":101},"8094f498-508a-47da-8d93-835aed730456",{"type":20,"title":87},"Generalization",{"id":89,"data":90,"type":19,"maxContentLevel":19,"version":26},"bec272d3-6246-4006-823a-303718d5e862",{"type":19,"summary":91},[92,93,94],"A model's ability to handle data it hasn't seen before is known as generalization","Manually programmed AI models typically struggle with generalization","Machine learning models, on the other hand, are good at generalization",{"id":96,"data":97,"type":37,"maxContentLevel":19,"version":26},"52d7a44d-af94-4e12-a96f-0eabf2eaeb2c",{"type":37,"intro":98},[99,100],"What is generalization, and why is it so important?","Are AI models capable of generalization or not?",[102,107,120],{"id":103,"data":104,"type":26,"maxContentLevel":19,"version":26},"7ace9664-80bf-4524-a3aa-dab28a6ea170",{"type":26,"contentRole":20,"markdownContent":105,"audioMediaId":106},"So, we know that machine learning will often be faster and easier than manual programming. But there's also another important benefit. If it's done correctly, this approach can actually allow a model to 'think outside the box'.\n\nImagine that you manually programmed an AI to close your window when it’s rainy, and to open your window when it’s sunny. Then one day it snows, and the AI doesn’t know what to do.\n\nIt's another big drawback with symbolic AI. However much time you spend on the programming, you might struggle to think of every possible scenario. And if you forget a scenario – if you don't programme a rule – your AI model won't know how to respond.\n\n![Graph](image://13600504-58cb-43d8-8df4-a8cd12a12a2c \"Snow. Image via Pexels\")","4e13d5e2-408f-4228-a5b4-2a688f7f829e",{"id":108,"data":109,"type":26,"maxContentLevel":19,"version":26,"reviews":112},"20602a37-9529-4b55-a555-d8ea261ed2e2",{"type":26,"contentRole":20,"markdownContent":110,"audioMediaId":111},"Compare that to machine learning, where an AI model has watched thousands of videos of people opening and closing their windows.\n\nNone of these videos have snow in them – but the AI has noticed some patterns. People usually close their windows when something falls from the sky. Especially when that something is cold and wet.\n\nBased on these patterns, the AI is able to extrapolate: the snow appears to be cold and wet, so it must be time to close the window.","ab12f439-5b21-4bc6-b607-629feb63ca2c",[113],{"id":114,"data":115,"type":56,"version":26,"maxContentLevel":19},"03c6b68d-6d2f-4294-bb46-2d0f71dffbde",{"type":56,"reviewType":20,"spacingBehaviour":26,"binaryQuestion":116,"binaryCorrect":118,"binaryIncorrect":119},[117],"Which approach is better for preparing a model to deal with unexpected scenarios?",[60],[83],{"id":121,"data":122,"type":26,"maxContentLevel":19,"version":26,"reviews":125},"dbba8a52-9411-4e55-a8a1-7b459b2d61bf",{"type":26,"contentRole":20,"markdownContent":123,"audioMediaId":124},"That snow scenario is an example of **generalization**. In the context of AI, this term refers to a model's ability to handle data it hasn't seen before, as opposed to just handling the data it originally learned from.\n\nHumans are pretty good at this. Imagine showing a child some pictures of dogs: big dogs, small dogs, furry dogs, short-haired dogs. Later, in the park, they see a breed of dog that wasn't in the photos. But they still know it's a dog, because they have a 'generalized' understanding of the data.\n\n![Graph](image://00510760-a931-406e-a9e4-b5a4b162fb17 \"Image via Pexels\")\n\nMachine learning allows an AI model to gain a generalized understanding too. This is extremely useful – as we said, it's hard to predict (and manually programme) every possible scenario that an AI model might face.","8b1760d8-4872-41f9-8ef9-1d2b8b764498",[126,145],{"id":127,"data":128,"type":56,"version":26,"maxContentLevel":19},"4fb75b36-77c9-4875-890f-79bad100dc81",{"type":56,"reviewType":19,"spacingBehaviour":26,"collapsingSiblings":129,"multiChoiceQuestion":133,"multiChoiceCorrect":135,"multiChoiceIncorrect":136,"multiChoiceMultiSelect":6,"multiChoiceRevealAnswerOption":6,"matchPairsQuestion":140,"matchPairsPairs":142},[130,131,132],"ea05be6e-cdeb-44f6-9f31-8aec11f44810","2e5c3c78-0e9c-4d53-996b-4757282d0c67","40b646f9-87d1-42a7-affa-48dfb7102234",[134],"What do we call a model's ability to handle data it hasn't seen before?",[87],[137,138,139],"Overfitting","Specialization","Adaptation",[141],"Match the pairs below:",[143],{"left":87,"right":144,"direction":19},"Ability to handle new data",{"id":146,"data":147,"type":56,"version":26,"maxContentLevel":19},"1528186c-723c-4a4e-b608-d32b1d5cc56b",{"type":56,"reviewType":20,"spacingBehaviour":26,"binaryQuestion":148,"binaryCorrect":150,"binaryIncorrect":152},[149],"True or false: humans are generally pretty good at generalization.",[151],"TRUE",[153],"FALSE",{"id":155,"data":156,"type":20,"version":26,"maxContentLevel":19,"summaryPage":158,"introPage":165,"pages":171},"fbc37b82-52c0-40e6-97b4-bd980149aa8f",{"type":20,"title":157},"Parameters and loss",{"id":159,"data":160,"type":19,"maxContentLevel":19,"version":26},"5c8f3e76-fe4a-4e1a-802e-e6aabc7c7b3d",{"type":19,"summary":161},[162,163,164],"When a machine is 'learning', it simply means that numerical parameters are changing up and down","Loss function describes the gap between an AI's output, and a real-life piece of data","The smaller the loss function, the better the machine has learned",{"id":166,"data":167,"type":37,"maxContentLevel":19,"version":26},"b4298f4e-9655-4d43-80b1-b6aa1ccf56c4",{"type":37,"intro":168},[169,170],"What does it actually mean when we say that a machine is 'learning'?","Why do larger datasets make machine learning more effective?",[172,197,202,216],{"id":173,"data":174,"type":26,"maxContentLevel":19,"version":26,"reviews":177},"bd6960df-d290-4801-b7dc-1fa3ae9ce945",{"type":26,"contentRole":20,"markdownContent":175,"audioMediaId":176},"Like everything in AI, machine learning is essentially a computer programme which mimics a human-like process. But how does this programme work?\n\nYou can think of it as a template with some numbers attached. AI scientists refer to these numbers as **parameters**.\n\n![Graph](image://61864a07-3264-42a2-b193-74302e343166 \"A (very) simple template with parameters.\")\n\nWhen data is fed into this template, those parameters automatically adjust. The numbers will either get higher or lower, depending on the nature of the data. Just imagine someone sitting there, adjusting dials up and down, as the different bits of data come through.\n\nWe'll look at an example in a second. But in simple terms: when we say that a machine is 'learning', we really just mean that these parameters are changing up and down.\n\n![Graph](image://a198586f-bd2c-491c-8382-823794434830 \"Parameters changing up and down.\")","13469af6-18a5-4823-ae2f-d35a29ee73c3",[178],{"id":179,"data":180,"type":56,"version":26,"maxContentLevel":19},"6d524319-b4e0-487f-abf1-b926a6d71ca9",{"type":56,"reviewType":19,"spacingBehaviour":26,"collapsingSiblings":181,"multiChoiceQuestion":185,"multiChoiceCorrect":187,"multiChoiceIncorrect":189,"multiChoiceMultiSelect":6,"multiChoiceRevealAnswerOption":6,"matchPairsQuestion":193,"matchPairsPairs":194},[182,183,184],"0133be01-b811-4728-b69c-1fef94437099","7a6b611e-6b5f-4bdd-bd95-10c32e397c86","dc19d651-ce4e-4af1-8819-4ac1d8132d15",[186],"Machine learning is essentially a template with parameters attached. In simple terms, what are parameters?",[188],"Numerical values",[190,191,192],"Algorithms","Strings of logic","Symbolic labels",[141],[195],{"left":188,"right":196,"direction":19},"Parameters",{"id":198,"data":199,"type":26,"maxContentLevel":19,"version":26},"fb68e3fb-43a0-4e25-868e-36444433f942",{"type":26,"contentRole":20,"markdownContent":200,"audioMediaId":201},"Here's that example we promised. Let's imagine you're building an AI model which can predict the price of a pizza based on its size.\n\nYour template has parameters for all the different sizes: 10-inch pizza, 11-inch pizza, 12-inch pizza, and so on. These parameters all have prices attached, which you've estimated yourself.\n\n![Graph](image://9300acc9-f0fc-4285-aa7c-32325ceb510f \"Estimated pizza prices.\")\n\nYou then give the AI some menus from local pizza restaurants. As each menu comes through, it tweaks its parameters, and the prices of each pizza gradually go up or down.\n\nFor example, if one menu showed a 10-inch pizza for $8.50, the AI would tweak that original parameter down a couple of notches. If the next menu showed a 10-inch pizza for $9.50, it tweaks the number up again. Every time it does this, it gets closer and closer to an average price for a 10-inch pizza in your area.\n\n![Graph](image://9e5222e3-c4b8-416a-b8b9-24657cb27570 \"Adjusting prices up and down.\")","93622f33-5ccf-43ca-868c-ddf003c3a325",{"id":203,"data":204,"type":26,"maxContentLevel":19,"version":26,"reviews":207},"251e4dea-38a8-46b1-9211-e6c5e7314fef",{"type":26,"contentRole":20,"markdownContent":205,"audioMediaId":206},"By the end, the AI will have a new set of numbers: it dropped from $10.00 to $9.00 for a 10-inch, from $11.00 to $10.00 for an 11-inch, and from $12.00 to $11.00 for a 12-inch.\n\n![Graph](image://e5bfef35-995a-49ee-a5cb-f098c43a3d8d \"Pizza prices.\")\n\nIn other words, this AI has 'learned' some average prices for pizzas being sold in your area. And if you asked it to estimate the cost of, say, a 14-inch pizza, it would hopefully give you a decent answer. In this case, maybe $13.00.\n\nYou can check if that number is accurate by looking in another pizza menu. You glance through the options, and find that this particular pizza restaurant is selling 14-inch pizzas for $13.25. That's not *exactly* the same as your AI's estimate, but all things considered, it's pretty close.","9a0cee73-47a3-454e-a06a-ea543d9932a9",[208],{"id":209,"data":210,"type":56,"version":26,"maxContentLevel":19},"17b68b2f-7243-4c22-97bf-a479489f3406",{"type":56,"reviewType":211,"spacingBehaviour":26,"clozeQuestion":212,"clozeWords":214},4,[213],"When a machine is learning, its parameters are just changing up and down.",[215],"parameters",{"id":217,"data":218,"type":26,"maxContentLevel":19,"version":26,"reviews":221},"720ea9d7-67fe-41d6-a584-8c3a9b897d77",{"type":26,"contentRole":20,"markdownContent":219,"audioMediaId":220},"There's actually a name for this 'gap' between the AI's prediction ($13.00) and the real-life data ($13.25). Scientists call it the **loss function**, and it's an important part of machine learning.\n\nThe smaller the loss function, the better your machine has learned. In an ideal scenario, the loss function would be non-existent – your AI would have predicted $13.25 in the first place.\n\nTo reduce the loss function for this particular AI, you could give it more menus to learn from. In theory, with every piece of data it encounters, it will get closer and closer to the 'truth'. It's not always that simple, but as a general rule, larger datasets usually lead to more effective AI models.\n\n![Graph](image://9a205d1e-fff1-4b1c-93e4-71d7062907ab \"Pizza by Valerio Capello, English Wikipedia (CC BY-SA 3.0) \u003Chttp://creativecommons.org/licenses/by-sa/3.0/>, via Wikimedia Commons\")","143bde96-0315-4b96-9c2c-2707d389c731",[222,237,246],{"id":130,"data":223,"type":56,"version":26,"maxContentLevel":19},{"type":56,"reviewType":19,"spacingBehaviour":26,"collapsingSiblings":224,"multiChoiceQuestion":225,"multiChoiceCorrect":227,"multiChoiceIncorrect":229,"multiChoiceMultiSelect":6,"multiChoiceRevealAnswerOption":6,"matchPairsQuestion":233,"matchPairsPairs":234},[131,132,127],[226],"What do we call the gap between an AI's predictions, and real-life data?",[228],"Loss function",[230,231,232],"Error margin","Deviation factor","Discrepancy measure",[141],[235],{"left":228,"right":236,"direction":19},"Gap between output and real-life data",{"id":238,"data":239,"type":56,"version":26,"maxContentLevel":19},"44a8ec4c-5410-4862-a820-bcb5e75e4162",{"type":56,"reviewType":20,"spacingBehaviour":26,"binaryQuestion":240,"binaryCorrect":242,"binaryIncorrect":244},[241],"Which of these is a general rule for machine learning?",[243],"Larger datasets lead to smaller loss functions",[245],"Larger datasets lead to larger loss functions",{"id":247,"data":248,"type":56,"version":26,"maxContentLevel":19},"2749d721-ea48-4906-8679-d3e21e48f3ed",{"type":56,"reviewType":26,"spacingBehaviour":26,"activeRecallQuestion":249,"activeRecallAnswers":251},[250],"An AI predicts that the sun will rise at 06:43. The sun actually rises at 06:47. What is the loss function?",[252],"4 minutes",{"id":254,"data":255,"type":20,"version":26,"maxContentLevel":19,"summaryPage":257,"introPage":264,"pages":270},"db74e4ac-2ec6-41ba-a8b1-d01d3b016868",{"type":20,"title":256},"Supervised learning",{"id":258,"data":259,"type":19,"maxContentLevel":19,"version":26},"f1427bab-de36-4f39-a27a-a61975472270",{"type":19,"summary":260},[261,262,263],"Supervised learning is when you train your AI using a labeled set of data","Labeled data is essentially just a list of input-output pairs","The AI 'learns' by identifying rules and relationships between these input-output pairs",{"id":265,"data":266,"type":37,"maxContentLevel":19,"version":26},"2d972447-8d79-4d14-952c-5961d08daa4e",{"type":37,"intro":267},[268,269],"What is the meaning of supervised learning?","When should this approach be used?",[271,302,307,332],{"id":272,"data":273,"type":26,"maxContentLevel":19,"version":26,"reviews":276},"d747844d-c0a3-4c5f-85af-70b01172e141",{"type":26,"contentRole":20,"markdownContent":274,"audioMediaId":275},"When you want an AI to learn from a dataset, like in that pizza example we talked about last time, there are a few different approaches you can take.\n\nOne of these approaches is **supervised learning**. This is when you train your AI using a labeled set of data. When we say 'labeled', we mean that the data has all been carefully arranged into something called input-output pairs.\n\nThat's what we did with those pizzas last time. The input side of the pair was the size of each pizza (e.g. 10 inch) and the output side of the pair was the price of each pizza (e.g. $9.00).\n\nIn this example, our labeled data is essentially just a list of these size-price (input-output) pairs.","28f65da0-7464-4740-9cf7-3a9fb31b3914",[277,295],{"id":278,"data":279,"type":56,"version":26,"maxContentLevel":19},"03c9194d-374d-4729-8424-6a0b3b4c8d4b",{"type":56,"reviewType":19,"spacingBehaviour":26,"collapsingSiblings":280,"multiChoiceQuestion":284,"multiChoiceCorrect":286,"multiChoiceIncorrect":288,"multiChoiceMultiSelect":6,"multiChoiceRevealAnswerOption":6,"matchPairsQuestion":292,"matchPairsPairs":293},[281,282,283],"f23209c6-ee2a-4ec9-878f-ceb80bd6ebae","223d5086-2ea3-46ba-9777-723c9e202ef2","ba8ca3b7-3f63-4310-854b-13d0f5acd9f1",[285],"What is the definition of supervised learning?",[287],"Training AI using labeled data",[289,290,291],"Training AI using unstructured data","Training AI using punishments and rewards","Manually programming an AI",[141],[294],{"left":287,"right":256,"direction":19},{"id":296,"data":297,"type":56,"version":26,"maxContentLevel":19},"3da65ec5-757a-42d8-9733-69a99e0fb260",{"type":56,"reviewType":26,"spacingBehaviour":26,"activeRecallQuestion":298,"activeRecallAnswers":300},[299],"In supervised learning, the data needs to be manually arranged into what?",[301],"Input-output pairs",{"id":303,"data":304,"type":26,"maxContentLevel":19,"version":26},"bd9b040d-d22d-4727-99b7-447bfe16bcc6",{"type":26,"contentRole":20,"markdownContent":305,"audioMediaId":306},"An input-output pair will hopefully have some kind of relationship or rule. In our pizza example, the rule went like this: the output number (price) was always *one less* than the input number (size).\n\n![Graph](image://974fb9ca-d2dd-4247-aaee-ca0d4c526f2c \"Pizza prices.\")\n\nIn a lot of cases, we won't actually know that relationship or rule ourselves. Instead, we want the AI to look at the input-output pairs, and search for patterns that help it to establish the relationship or rule by itself.\n\nOnce it's established the rule, it can start using it. For example, we could give an input number to our pizza AI: \"5-inch\". Using the rule above, the AI could predict the output number: \"$4.00\".\n\nIn plainer language, we could ask the AI, \"How much do you think a 5-inch pizza would cost?\" and it could reply to you, \"It will probably cost $4.00\".","6853f56a-ae06-46dd-8c63-e6628433c3fa",{"id":308,"data":309,"type":26,"maxContentLevel":19,"version":26,"reviews":312},"fbc759b5-3588-4de1-ba9e-0c4e9c64df77",{"type":26,"contentRole":20,"markdownContent":310,"audioMediaId":311},"Here's another example. We could give our AI a bunch of emails to look at. Half of these emails are labeled 'spam', and the other half are labeled 'not spam'. This time, the input is the email itself, and the output is either 'spam' or 'not spam'.\n\nUsing this labeled data, the AI can now learn some rules or patterns that distinguish spam from non-spam. For example, it might notice that an input email with more spelling mistakes, or strange punctuation, is typically linked to an output label saying 'spam'.\n\nLater, whenever we give this AI a random email, it can use this rule to sort it. Lots of spelling mistakes? Or unusual punctuation? The email is sent to spam.\n\n![Graph](image://4cabceb9-902b-487f-adfd-c2aac2e32121 \"Spam or not spam? Image: gmcgcc, CC BY-SA 4.0 \u003Chttps://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons\")","b627c77b-04cd-4b02-a3ea-658d8dc39c47",[313],{"id":314,"data":315,"type":56,"version":26,"maxContentLevel":19},"364a4be5-5793-4a6e-8346-87b0c79dc841",{"type":56,"reviewType":316,"spacingBehaviour":26,"matchPairsQuestion":317,"matchPairsPairs":319,"matchPairsShowExamples":6},6,[318],"A supervised AI establishes a rule. The height of a dog (input) is double the length of its tail (output). Using this rule, match the pairs below:",[320,323,326,329],{"left":321,"right":322,"direction":19},"This dog is 50 cm tall","The dog's tail is 25 cm long",{"left":324,"right":325,"direction":19},"This dog is 80 cm tall","The dog's tail is 40 cm long",{"left":327,"right":328,"direction":19},"This dog is 20 cm tall","The dog's tail is 10 cm long",{"left":330,"right":331,"direction":19},"None of these","The dog's tail is 100 cm long",{"id":333,"data":334,"type":26,"maxContentLevel":19,"version":26,"reviews":337},"952f96f8-7681-4bdd-826f-c1360522060f",{"type":26,"contentRole":20,"markdownContent":335,"audioMediaId":336},"Supervised learning is great in scenarios like that email example – when there's a clear pair of inputs and outputs.\n\nHere are some other examples: when you give your AI some text in one language (input) then get a translated version (output). When you play a piece of music to your AI (input) then get the title of that song (output).\n\nAlso, that cat and dog example we mentioned earlier on. You can give it a photo of an animal (input) then get the name of that animal (output).\n\n![Graph](image://af27fc29-bc2d-4547-96eb-053ed754f436 \"Cats and dogs.\")\n\nThis isn't the only approach to machine learning. We'll look at some other types later. But it's a very common approach right now, with thousands of different uses.","12c316c2-e8e3-4121-960d-2d57fac1c02e",[338],{"id":339,"data":340,"type":56,"version":26,"maxContentLevel":19},"763fb240-ffe9-4675-b248-a309c3869276",{"type":56,"reviewType":316,"spacingBehaviour":26,"matchPairsQuestion":341,"matchPairsPairs":343,"matchPairsShowExamples":6},[342],"Supervised learning works best for data with clear pairs of input and outputs. Match the pairs below:",[344,347,350,353],{"left":345,"right":346,"direction":19},"Movie budget (input)","Star rating (output)",{"left":348,"right":349,"direction":19},"Medical symptoms (input)","Disease diagnosis (output)",{"left":351,"right":352,"direction":19},"Size of house (input)","Market price (output)",{"left":354,"right":355,"direction":19},"Photo of animal (input)","Species identification (output)",{"id":357,"data":358,"type":20,"version":20,"maxContentLevel":19,"summaryPage":360,"introPage":367,"pages":372},"ffc93e38-1ccc-45ba-9345-e5171e03b3ac",{"type":20,"title":359},"Unsupervised learning",{"id":361,"data":362,"type":19,"maxContentLevel":19,"version":26},"9469463b-a890-4408-879d-e2d6bc9c97e9",{"type":19,"summary":363},[364,365,366],"Unsupervised learning involves giving AI unlabelled data to find patterns","Clustering is a technique in unsupervised learning, which sorts data into similar groups","Association is another technique which searches for relationships between data ",{"id":368,"data":369,"type":37,"maxContentLevel":19,"version":26},"0180b0a0-e72f-4348-8ce3-7d2f9d7b70f1",{"type":37,"intro":370},[371,269],"What is the meaning of unsupervised learning?",[373,397,402,424,429],{"id":374,"data":375,"type":26,"maxContentLevel":19,"version":26,"reviews":378},"ef5c19ab-dba0-46b8-8508-c39d9f89fb66",{"type":26,"contentRole":20,"markdownContent":376,"audioMediaId":377},"The main alternative to supervised learning is (can you guess it?) **unsupervised learning**.\n\nThis time, you won't be giving your AI a set of nicely labeled pairs. Instead, you'll be giving it an unstructured pile of raw, unlabelled data.\n\nThis data might actually follow some interesting rules and patterns. But as a human, you don't know what they are. The data is too messy, too large, too confusing. There's certainly nothing as intuitive here as a simple input-output pair.\n\nSo you ask your AI \"are there any patterns in here?\" and see if it can learn anything useful.","c1dcb7d1-cbef-4ea3-8877-78c1c7ab6c75",[379,390],{"id":281,"data":380,"type":56,"version":26,"maxContentLevel":19},{"type":56,"reviewType":19,"spacingBehaviour":26,"collapsingSiblings":381,"multiChoiceQuestion":382,"multiChoiceCorrect":384,"multiChoiceIncorrect":385,"multiChoiceMultiSelect":6,"multiChoiceRevealAnswerOption":6,"matchPairsQuestion":387,"matchPairsPairs":388},[278,282,283],[383],"What is the definition of unsupervised learning?",[289],[386,290,291],"Training AI using labelled data",[141],[389],{"left":289,"right":359,"direction":19},{"id":391,"data":392,"type":56,"version":26,"maxContentLevel":19},"0d277827-3c7a-43ea-9c95-956576de4602",{"type":56,"reviewType":26,"spacingBehaviour":26,"activeRecallQuestion":393,"activeRecallAnswers":395},[394],"What is the main purpose of unsupervised learning?",[396],"Identifying patterns in messy datasets",{"id":398,"data":399,"type":26,"maxContentLevel":19,"version":26},"9d2b8a4a-6eeb-425e-ba99-e2ee963ebc28",{"type":26,"contentRole":20,"markdownContent":400,"audioMediaId":401},"Typically, an unsupervised learning model will use a technique called **clustering**. This technique involves sorting data into groups based on apparent similarities and differences.\n\nIt's like giving a child a handful of marbles, then asking them to sort them. They may start to sort them by size, or color, or opacity, or weight, or whatever other patterns they come up with. An AI can do the same with data, seeking patterns, then learning from them.\n\nHere's a real world example. Imagine that you run a gym. You ask an AI to look for patterns in a database of all your members.\n\nIt finds some interesting clusters: people who live in the east of the city seem more interested in taking yoga classes, whereas people who live in the west of the city are more interested in taking spin classes. Why? You have no idea. But later, when you open a new gym in the east, you make sure it specializes in yoga.\n\n![Graph](image://450f2747-75d6-477c-9fbb-473fd3bd4f66 \"Clustered data. Image: BogdanShevchenko (CC BY-SA 4.0) \u003Chttps://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons\")","cbfdbd0f-6c04-4074-9fa9-9d7d766f948d",{"id":403,"data":404,"type":26,"maxContentLevel":19,"version":26,"reviews":407},"99393225-f569-47d5-bfc2-cd86cec2076c",{"type":26,"contentRole":20,"markdownContent":405,"audioMediaId":406},"Another technique that a model might use is **association**. This involves looking for connections between pieces of data. Often, these connections are causal: if X happens, then Y will happen next.\n\nAs another example, imagine that you run a streaming service like Netflix. You ask an AI to look for patterns in a database of all your members. It finds an association: after watching a scary horror movie, it's common for people to calm down with their favorite rom com.\n\nThis knowledge helps you adjust your service – you make sure that rom coms are typically suggested after horrors.","3908f885-4b23-4887-afd1-42c4d8eda2d1",[408,417],{"id":409,"data":410,"type":56,"version":26,"maxContentLevel":19},"6b77d05a-deb5-4f9a-9c84-129543e6e90e",{"type":56,"reviewType":20,"spacingBehaviour":26,"binaryQuestion":411,"binaryCorrect":413,"binaryIncorrect":415},[412],"Which unsupervised technique sorts data into groups based on similarities?",[414],"Clustering",[416],"Association",{"id":418,"data":419,"type":56,"version":26,"maxContentLevel":19},"b2c8d412-b089-4b30-928b-16d398aa67f7",{"type":56,"reviewType":20,"spacingBehaviour":26,"binaryQuestion":420,"binaryCorrect":422,"binaryIncorrect":423},[421],"Which unsupervised technique looks for relationships between pieces of data?",[416],[414],{"id":425,"data":426,"type":26,"maxContentLevel":19,"version":20},"8053e1f2-4130-48b5-96c6-8171dec183d9",{"type":26,"contentRole":20,"markdownContent":427,"audioMediaId":428},"Unsupervised learning is incredibly powerful for exploratory analysis. It can identify all kinds of rules and patterns that you never would have thought of on your own. But remember: these rules will sometimes be bizarre and useless.\n\nYour AI might notice that people born in November like movies about dogs, while people born in September prefer movies which feature cats – but only if those cats are black. For whatever reason, this pattern might genuinely be present in your data.\n\n![Graph](image://c370f4f2-a295-4f26-87e7-aece84df2053 \"Black cat by Dmitry Makeev (CC BY-SA 4.0) \u003Chttps://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons\")\n\nBut is it useful? That's for you to decide.","e0818c8e-4a63-440e-8b17-d72ff0f35847",{"id":430,"data":431,"type":26,"maxContentLevel":19,"version":26,"reviews":434},"5430bc56-3176-49fa-9813-0548c5c8a544",{"type":26,"contentRole":20,"markdownContent":432,"audioMediaId":433},"It's worth pointing out that unsupervised learning only works on massive datasets. And it's more computationally complex than supervised learning – it takes up a lot more power.\n\nThere's a trade off here. Unsupervised learning is more work for a computer, but it's much less work for a human. Supervised learning, on the other hand, is easier for the computer, but it takes time and effort on the human side to sort and label the data.\n\nA scientist will generally choose between them based on the ultimate goal of their project. If they're looking to find some unexpected patterns in a giant dataset, they'll use unsupervised. If they just want to find some simple relationships between inputs and outputs, they'll use supervised learning instead.","9ce62ffd-a232-497c-9b73-d08d012804d0",[435],{"id":436,"data":437,"type":56,"version":26,"maxContentLevel":19},"ccab711e-c983-4d28-9281-c40dc037c6bd",{"type":56,"reviewType":19,"spacingBehaviour":26,"multiChoiceQuestion":438,"multiChoiceCorrect":440,"multiChoiceIncorrect":444,"multiChoiceMultiSelect":447,"multiChoiceRevealAnswerOption":6},[439],"What are some of the main drawbacks of unsupervised learning?",[441,442,443],"Only works on massive datasets","Takes a lot of computing power","Sometimes comes up with strange patterns",[445,446],"Only works on small datasets","Takes a lot of manual setup",true,{"id":449,"data":450,"type":20,"version":26,"maxContentLevel":19,"summaryPage":452,"introPage":459,"pages":464},"1f594859-93e2-4a4d-9efa-4183ac18ae31",{"type":20,"title":451},"Reinforcement learning",{"id":453,"data":454,"type":19,"maxContentLevel":19,"version":26},"4c4e2898-d816-4425-a635-e50e9209422a",{"type":19,"summary":455},[456,457,458],"Reinforcement learning uses rewards and punishments to train an AI","These rewards and punishments are actually just positive and negative values","Reinforcement learning is used for dynamic, real world situations that can't easily be summed up in a database",{"id":460,"data":461,"type":37,"maxContentLevel":19,"version":26},"fbd3eca3-da91-42fb-ad97-d910fe152dd3",{"type":37,"intro":462},[463,269],"What is the meaning of reinforcement learning?",[465,482,495,510],{"id":466,"data":467,"type":26,"maxContentLevel":19,"version":26,"reviews":470},"b8beb943-9690-45eb-a225-5c321bd3b6e9",{"type":26,"contentRole":20,"markdownContent":468,"audioMediaId":469},"Along with supervised learning and unsupervised learning, another popular approach to machine learning is something called **reinforcement learning**.\n\nThis one is pretty different from the others, as it doesn't require a dataset. Instead, you're going to be putting your AI in a closed environment, where it can learn via rewards and punishments.\n\nThese rewards and punishments are just positive values and negative values which are assigned to different actions. Whenever the AI performs an action, the corresponding value will tell the AI whether to perform that action again.\n\nFundamentally, it's just like operant conditioning. But instead of using rewards and punishments to train rats and monkeys, it's being used to train an AI.","34fc387d-6b08-4c10-a2dd-9a7b2adb10ba",[471],{"id":283,"data":472,"type":56,"version":26,"maxContentLevel":19},{"type":56,"reviewType":19,"spacingBehaviour":26,"collapsingSiblings":473,"multiChoiceQuestion":474,"multiChoiceCorrect":476,"multiChoiceIncorrect":478,"multiChoiceMultiSelect":6,"multiChoiceRevealAnswerOption":6,"matchPairsQuestion":479,"matchPairsPairs":480},[278,281,282],[475],"What is the definition of reinforcement learning?",[477],"Training AI using rewards and punishments",[386,289,291],[141],[481],{"left":477,"right":451,"direction":19},{"id":483,"data":484,"type":26,"maxContentLevel":19,"version":26,"reviews":487},"28f721fb-e62d-4c0c-8795-19b615695fff",{"type":26,"contentRole":20,"markdownContent":485,"audioMediaId":486},"Reinforcement learning would be a great approach if you wanted to teach an AI model how to play a game of chess.\n\n![Graph](image://37632363-4caa-4564-9930-115482039eab \"Chess by User:Cburnett (CC BY-SA 3.0) \u003Chttp://creativecommons.org/licenses/by-sa/3.0/>, via Wikimedia Commons\")\n\nYou'd need to build an environment where the AI can play chess matches over and over and over. Each win is associated with a positive value; each loss is associated with a negative value.\n\nBy the time it's played thousands of matches, this AI might have learned which sets of moves are most likely to lead to a positive, winning outcome. It might also have learned which moves to avoid, or which tactics are too risky to pull off.\n\nReinforcement learning can also be used in other scenarios, like teaching a robot how to walk. Assign a positive value to a successful step, and a negative value if the robot falls to the floor.","bdb2f374-c7a6-46d9-b803-127b30c4642c",[488],{"id":489,"data":490,"type":56,"version":26,"maxContentLevel":19},"20e1af39-480d-4f38-9aef-64ad5836a0c9",{"type":56,"reviewType":26,"spacingBehaviour":26,"activeRecallQuestion":491,"activeRecallAnswers":493},[492],"In the context of reinforcement learning, rewards and punishments are really what?",[494],"Positive values and negative values which are assigned to different actions",{"id":496,"data":497,"type":26,"maxContentLevel":19,"version":26,"reviews":500},"f1c7424b-c44b-4e80-9fd3-ce897028b392",{"type":26,"contentRole":20,"markdownContent":498,"audioMediaId":499},"In general, you'd use reinforcement learning if you wanted your AI to learn how to deal with dynamic, real world situations. The kinds of situations that can't easily be summed up in a database.\n\nAnother example would be training a self-driving car. Instead of getting that AI to look for patterns in data about collisions and traffic codes, it might work better to build an artificial environment where the AI can simulate driving around, and learn through trial and error.\n\n![Graph](image://27294a6d-78b9-4ed8-aae7-9cee59c6c4b3 \"Self-driving car by Grendelkhan (CC BY-SA 4.0) \u003Chttps://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons\")\n\nHaving said that, there might also be value in getting an AI model to learn some basic principles using supervised learning, then to graduate it to reinforcement learning later. This approach was actually how AlphaGo was trained to play the game of Go.","8b5159bf-47b1-47de-bd5f-55078ddfe1bf",[501],{"id":502,"data":503,"type":56,"version":26,"maxContentLevel":19},"67d1d681-1bfc-4f9e-918e-00db4f37d037",{"type":56,"reviewType":20,"spacingBehaviour":26,"binaryQuestion":504,"binaryCorrect":506,"binaryIncorrect":508},[505],"What kind of situations is reinforcement learning best suited for?",[507],"Real world situations which can't be summed up with data",[509],"Real world situations with lots of relevant data",{"id":511,"data":512,"type":26,"maxContentLevel":19,"version":26,"reviews":515},"3f813f3f-3f88-4d22-a6bb-2706480c773c",{"type":26,"contentRole":20,"markdownContent":513,"audioMediaId":514},"AlphaGo initially learned from a dataset: it observed thousands of recordings of real-life moves by expert human Go players. This was a labeled dataset – the model could see when a set of moves (input) resulted in a victory (output).\n\nAfter absorbing enough basic rules and patterns through this supervised stage, the model moved to an environment where it played thousands of Go matches against versions of itself, getting better and better through reinforcement.\n\nThis combined approach worked wonders. By the end, AlphaGo had learned enough to successfully defeat the greatest Go player in the world.","91fa6f55-58ea-42e6-8d95-0cc5c1bc40a7",[516,529],{"id":517,"data":518,"type":56,"version":26,"maxContentLevel":19},"523cf760-ad5f-4e96-8126-cc7fe4639d74",{"type":56,"reviewType":519,"spacingBehaviour":26,"orderAxisType":316,"orderQuestion":520,"orderItems":522},8,[521],"Put these stages of AlphaGo's development in order.",[523,525,527],{"label":524,"sortOrder":4},"Supervised learning (observing Go matches)",{"label":526,"sortOrder":26},"Reinforcement learning (playing Go against itself)",{"label":528,"sortOrder":20},"Real world play (defeating Lee Sedol)",{"id":530,"data":531,"type":56,"version":26,"maxContentLevel":19},"c95c116f-0078-4973-aea2-b58f3efc7e83",{"type":56,"reviewType":316,"spacingBehaviour":26,"matchPairsQuestion":532,"matchPairsPairs":534,"matchPairsShowExamples":6},[533],"In the context of reinforcement learning, match the pairs below:",[535,538,541,544],{"left":536,"right":537,"direction":19},"Positive value for a self-driving car","Arrive at correct destination",{"left":539,"right":540,"direction":19},"Negative value for a self-driving car","Arrive at wrong destination",{"left":542,"right":543,"direction":19},"Positive value for a voice assistant","Correctly obey command",{"left":545,"right":546,"direction":19},"Negative value for a voice assistant","Wrongly obey command",{"id":548,"data":549,"type":20,"version":26,"maxContentLevel":19,"summaryPage":551,"introPage":558,"pages":564},"f78bf9aa-b641-4651-83aa-6c37ef9d0fc0",{"type":20,"title":550},"Effectiveness of data",{"id":552,"data":553,"type":19,"maxContentLevel":19,"version":26},"d5a67cad-1f6a-48e8-ad73-9a758aa5dedf",{"type":19,"summary":554},[555,556,557],"Availability of data is crucial for machine learning models to be effective","Ideally, you'll need at least ten pieces of data for every parameter in your model","Lack of data can hinder AI development, especially when you're trying to use machine learning",{"id":559,"data":560,"type":37,"maxContentLevel":19,"version":26},"7e0c4ba5-2426-42fe-a779-b8731649bc9b",{"type":37,"intro":561},[562,563],"Why is data such an important part of machine learning?","Is it possible to develop models when you don't have data?",[565,582,595,617],{"id":566,"data":567,"type":26,"maxContentLevel":19,"version":26,"reviews":570},"a1975951-7b87-4dca-870d-8713d7e8b652",{"type":26,"contentRole":20,"markdownContent":568,"audioMediaId":569},"A lot of people are (rightly) excited about modern machine learning.\n\nAs we talked about earlier, it lets us build complex, powerful models without needing to programme all the rules of those models by hand. It also lets us build flexible, adaptive models, which can change and evolve, and handle unexpected problems.\n\nBut there's one important challenge that we need to be aware of: **availability of data**.\n\n![Graph](image://e3dcb218-1053-4f01-9fd3-7f22c128942a \"Data. By Chiffre01 (CC BY-SA 4.0) \u003Chttps://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons\")\n\nWhile reinforcement learning doesn't need a dataset, supervised learning and unsupervised learning do. And this can't just be any dataset. It needs to be relevant to the task that you want your AI to be learning.","8a507a33-67aa-4f9f-8f9e-11ac4c59d70f",[571],{"id":572,"data":573,"type":56,"version":26,"maxContentLevel":19},"8c8f2778-035b-4c98-8bd9-b8046c752e2c",{"type":56,"reviewType":19,"spacingBehaviour":26,"multiChoiceQuestion":574,"multiChoiceCorrect":576,"multiChoiceIncorrect":578,"multiChoiceMultiSelect":6,"multiChoiceRevealAnswerOption":6},[575],"Which of these is a major challenge for supervised and unsupervised learning?",[577],"All of these",[579,580,581],"Finding large enough datasets","Finding quality datasets","Finding relevant datasets",{"id":583,"data":584,"type":26,"maxContentLevel":19,"version":26,"reviews":587},"5b12189e-01b1-49bf-8f17-e2f544651161",{"type":26,"contentRole":20,"markdownContent":585,"audioMediaId":586},"As well as a relevant dataset, you'll also need a *large* dataset. AI learns best when it's able to repeat things thousands and thousands of times.\n\nIt's hard to put an exact number of the amount of data that you'll need. One general rule is that you'll need ten pieces of data (for example, ten photos) for each parameter in your model. If you had a model with 200 parameters? You'd need 2000 pieces of data.\n\nThat rule is pretty arbitrary. It all depends on the model, and what you want it to learn. But these approximate numbers still help to highlight how much data is often needed.","81d55c47-f16a-4bf0-96fc-2491b8450c0e",[588],{"id":589,"data":590,"type":56,"version":26,"maxContentLevel":19},"d2282dba-3d94-4621-b5bf-165d1feecd58",{"type":56,"reviewType":211,"spacingBehaviour":26,"clozeQuestion":591,"clozeWords":593},[592],"As a general (but arbitrary) rule, you need ten pieces of data for every one parameter in your model.",[594],"ten",{"id":596,"data":597,"type":26,"maxContentLevel":19,"version":26,"reviews":600},"7d1272a1-4597-4127-958c-1dd874972511",{"type":26,"contentRole":20,"markdownContent":598,"audioMediaId":599},"If you do manage to find enough usable data, you'll want to divide it into parts. The first part will be used as **training data**, which your AI model can learn from.\n\nThe second part will be some **test data**, which you can give to the AI after the learning stage, and check how well it performs.\n\nAs another general rule: of your total data, 80% should be used for training, and 20% for testing. So if you had 2000 photos in total, you'd use 1600 to train your AI, and 400 to test it later on.","4f1a8efc-64cf-4d60-9225-9649337c40cb",[601,610],{"id":602,"data":603,"type":56,"version":26,"maxContentLevel":19},"bf7a8bdd-2dc7-4680-af4b-f3185d338f1f",{"type":56,"reviewType":20,"spacingBehaviour":26,"binaryQuestion":604,"binaryCorrect":606,"binaryIncorrect":608},[605],"When training an AI, what percentage of your data should you use as training data?",[607],"80%",[609],"20%",{"id":611,"data":612,"type":56,"version":26,"maxContentLevel":19},"e17540b4-6039-4c7a-ad24-cb16a1a4eb7e",{"type":56,"reviewType":20,"spacingBehaviour":26,"binaryQuestion":613,"binaryCorrect":615,"binaryIncorrect":616},[614],"When training an AI, what percentage of your data should you use as testing data?",[609],[607],{"id":618,"data":619,"type":26,"maxContentLevel":19,"version":26,"reviews":622},"a01130c0-2aaf-4609-956b-77b51bb0e216",{"type":26,"contentRole":20,"markdownContent":620,"audioMediaId":621},"If you can find all that data, that's amazing news. AI scientists often talk about the **unreasonable effectiveness of data**. They're basically saying, when you *do* have enough data, it will probably be more effective than you'd ever expect.\n\nBut the opposite is true as well. When you *don't* have that data, you're seriously going to struggle. It's an important drawback to modern AI. It's still pretty hard to develop models for problems that don't have much data.\n\nImagine, for example, that you wanted to train an AI model to spot the symptoms of a rare disease. But you can only find about twenty case studies of people who have had that disease. That's nowhere near enough data – machine learning wouldn't be possible.\n\n![Graph](image://d968dfc2-1a18-4a2e-9179-6cd741e841c4 \"Medical data. Image via Pexels\")","40033b4b-6457-448a-987b-4c5e37b131bf",[623,630],{"id":624,"data":625,"type":56,"version":26,"maxContentLevel":19},"3b379141-10e9-4cfc-926e-bfd129282f57",{"type":56,"reviewType":211,"spacingBehaviour":26,"clozeQuestion":626,"clozeWords":628},[627],"AI scientists often talk about the unreasonable effectiveness of data.",[629],"effectiveness",{"id":631,"data":632,"type":56,"version":26,"maxContentLevel":19},"882f552d-4443-4de8-a2ff-cb27e54af335",{"type":56,"reviewType":26,"spacingBehaviour":26,"activeRecallQuestion":633,"activeRecallAnswers":635},[634],"Why might it be hard to train an AI to identify rare diseases?",[636],"There isn't much data on rare diseases",[638,747,839,1009,1151,1312,1435],{"id":23,"data":24,"type":20,"version":26,"maxContentLevel":19,"summaryPage":27,"introPage":34,"pages":639},[640,680,712],{"id":43,"data":44,"type":26,"maxContentLevel":19,"version":26,"parsed":641},{"data":642,"body":645,"toc":678},{"title":643,"description":644},"","Last time, we saw how symbolic AI has a major limitation. It relies on manual programming – and manual programming can only take us so far. In order to build more complex AIs, we need to use a different approach.",{"type":646,"children":647},"root",[648,655,660,665],{"type":649,"tag":650,"props":651,"children":652},"element","p",{},[653],{"type":654,"value":644},"text",{"type":649,"tag":650,"props":656,"children":657},{},[658],{"type":654,"value":659},"But what's the alternative to manual programming?",{"type":649,"tag":650,"props":661,"children":662},{},[663],{"type":654,"value":664},"To answer that question, we can look at a human brain. Your own brain (obviously) wasn't manually programmed. Instead, you filled it with facts and ideas via a process of gradual learning. Taste an apple? That taste was added to your brain. Ride a bike? That skill was automatically stored.",{"type":649,"tag":650,"props":666,"children":667},{},[668,670,676],{"type":654,"value":669},"Over the last few decades, many scientists have taken the same approach to AI. They call this approach ",{"type":649,"tag":671,"props":672,"children":673},"strong",{},[674],{"type":654,"value":675},"machine learning",{"type":654,"value":677},", and it's literally changed the world.",{"title":643,"searchDepth":20,"depth":20,"links":679},[],{"id":48,"data":49,"type":26,"maxContentLevel":19,"version":26,"reviews":52,"parsed":681},{"data":682,"body":684,"toc":710},{"title":643,"description":683},"We'll get into the complexities of machine learning later, but here's a simple definition. An AI is said to 'learn' when you give it an input – like a big list of numbers, or some photos, or videos – and it changes its behavior afterwards. As we said, this is how humans learn as well. We experience an input, then we change.",{"type":646,"children":685},[686,690,695,705],{"type":649,"tag":650,"props":687,"children":688},{},[689],{"type":654,"value":683},{"type":649,"tag":650,"props":691,"children":692},{},[693],{"type":654,"value":694},"Let's run with that photo example. Imagine you've built an Artificial Intelligence that struggles to tell the difference between cats and dogs. You decide to show it a pile of animal photos and tell it over and over: \"this one’s a cat, this one’s a cat, this one’s a dog, this one’s a dog...\"",{"type":649,"tag":650,"props":696,"children":697},{},[698],{"type":649,"tag":699,"props":700,"children":704},"img",{"alt":701,"src":702,"title":703},"Graph","image://45419589-20de-4180-8724-3e3f8ab53ff2","Cats and dogs.",[],{"type":649,"tag":650,"props":706,"children":707},{},[708],{"type":654,"value":709},"Your AI starts to notice patterns in the photos, like the shape of the ears or the whiskers. And by the end, it's pretty good at recognizing cats and dogs on its own.",{"title":643,"searchDepth":20,"depth":20,"links":711},[],{"id":62,"data":63,"type":26,"maxContentLevel":19,"version":26,"reviews":66,"parsed":713},{"data":714,"body":716,"toc":745},{"title":643,"description":715},"Learning wouldn't be the only way to get an AI to recognize cats. Alternatively, a scientist could write symbolic rules (\"fluffy tail\", \"pointy ears\", \"long whiskers\", etc.), then manually build these rules into the AI's design.",{"type":646,"children":717},[718,722,727,732,737],{"type":649,"tag":650,"props":719,"children":720},{},[721],{"type":654,"value":715},{"type":649,"tag":650,"props":723,"children":724},{},[725],{"type":654,"value":726},"But it's often a lot faster, and a lot more effective, to take a learning approach instead.",{"type":649,"tag":650,"props":728,"children":729},{},[730],{"type":654,"value":731},"Let's think about another example: an AI that's designed to translate text. Instead of manually programming all the words from every language, you can train the AI on a pile of texts which have already been translated. As it works through those texts, and compares the languages, it can learn all the different words.",{"type":649,"tag":650,"props":733,"children":734},{},[735],{"type":654,"value":736},"This is exactly what Google did in 2016. They built an AI, then trained it on millions of translated texts. Whenever you load up Google Translate, you're interacting with an impressive, self-taught AI.",{"type":649,"tag":650,"props":738,"children":739},{},[740],{"type":649,"tag":699,"props":741,"children":744},{"alt":701,"src":742,"title":743},"image://cd3c914a-7b0d-41a2-82eb-1a294c3cfeb8","Google Translate. Image (CC BY-SA 4.0) \u003Chttps://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons",[],{"title":643,"searchDepth":20,"depth":20,"links":746},[],{"id":85,"data":86,"type":20,"version":26,"maxContentLevel":19,"summaryPage":88,"introPage":95,"pages":748},[749,779,801],{"id":103,"data":104,"type":26,"maxContentLevel":19,"version":26,"parsed":750},{"data":751,"body":753,"toc":777},{"title":643,"description":752},"So, we know that machine learning will often be faster and easier than manual programming. But there's also another important benefit. If it's done correctly, this approach can actually allow a model to 'think outside the box'.",{"type":646,"children":754},[755,759,764,769],{"type":649,"tag":650,"props":756,"children":757},{},[758],{"type":654,"value":752},{"type":649,"tag":650,"props":760,"children":761},{},[762],{"type":654,"value":763},"Imagine that you manually programmed an AI to close your window when it’s rainy, and to open your window when it’s sunny. Then one day it snows, and the AI doesn’t know what to do.",{"type":649,"tag":650,"props":765,"children":766},{},[767],{"type":654,"value":768},"It's another big drawback with symbolic AI. However much time you spend on the programming, you might struggle to think of every possible scenario. And if you forget a scenario – if you don't programme a rule – your AI model won't know how to respond.",{"type":649,"tag":650,"props":770,"children":771},{},[772],{"type":649,"tag":699,"props":773,"children":776},{"alt":701,"src":774,"title":775},"image://13600504-58cb-43d8-8df4-a8cd12a12a2c","Snow. Image via Pexels",[],{"title":643,"searchDepth":20,"depth":20,"links":778},[],{"id":108,"data":109,"type":26,"maxContentLevel":19,"version":26,"reviews":112,"parsed":780},{"data":781,"body":783,"toc":799},{"title":643,"description":782},"Compare that to machine learning, where an AI model has watched thousands of videos of people opening and closing their windows.",{"type":646,"children":784},[785,789,794],{"type":649,"tag":650,"props":786,"children":787},{},[788],{"type":654,"value":782},{"type":649,"tag":650,"props":790,"children":791},{},[792],{"type":654,"value":793},"None of these videos have snow in them – but the AI has noticed some patterns. People usually close their windows when something falls from the sky. Especially when that something is cold and wet.",{"type":649,"tag":650,"props":795,"children":796},{},[797],{"type":654,"value":798},"Based on these patterns, the AI is able to extrapolate: the snow appears to be cold and wet, so it must be time to close the window.",{"title":643,"searchDepth":20,"depth":20,"links":800},[],{"id":121,"data":122,"type":26,"maxContentLevel":19,"version":26,"reviews":125,"parsed":802},{"data":803,"body":805,"toc":837},{"title":643,"description":804},"That snow scenario is an example of generalization. In the context of AI, this term refers to a model's ability to handle data it hasn't seen before, as opposed to just handling the data it originally learned from.",{"type":646,"children":806},[807,819,824,832],{"type":649,"tag":650,"props":808,"children":809},{},[810,812,817],{"type":654,"value":811},"That snow scenario is an example of ",{"type":649,"tag":671,"props":813,"children":814},{},[815],{"type":654,"value":816},"generalization",{"type":654,"value":818},". In the context of AI, this term refers to a model's ability to handle data it hasn't seen before, as opposed to just handling the data it originally learned from.",{"type":649,"tag":650,"props":820,"children":821},{},[822],{"type":654,"value":823},"Humans are pretty good at this. Imagine showing a child some pictures of dogs: big dogs, small dogs, furry dogs, short-haired dogs. Later, in the park, they see a breed of dog that wasn't in the photos. But they still know it's a dog, because they have a 'generalized' understanding of the data.",{"type":649,"tag":650,"props":825,"children":826},{},[827],{"type":649,"tag":699,"props":828,"children":831},{"alt":701,"src":829,"title":830},"image://00510760-a931-406e-a9e4-b5a4b162fb17","Image via Pexels",[],{"type":649,"tag":650,"props":833,"children":834},{},[835],{"type":654,"value":836},"Machine learning allows an AI model to gain a generalized understanding too. This is extremely useful – as we said, it's hard to predict (and manually programme) every possible scenario that an AI model might face.",{"title":643,"searchDepth":20,"depth":20,"links":838},[],{"id":155,"data":156,"type":20,"version":26,"maxContentLevel":19,"summaryPage":158,"introPage":165,"pages":840},[841,890,933,971],{"id":173,"data":174,"type":26,"maxContentLevel":19,"version":26,"reviews":177,"parsed":842},{"data":843,"body":845,"toc":888},{"title":643,"description":844},"Like everything in AI, machine learning is essentially a computer programme which mimics a human-like process. But how does this programme work?",{"type":646,"children":846},[847,851,862,870,875,880],{"type":649,"tag":650,"props":848,"children":849},{},[850],{"type":654,"value":844},{"type":649,"tag":650,"props":852,"children":853},{},[854,856,860],{"type":654,"value":855},"You can think of it as a template with some numbers attached. AI scientists refer to these numbers as ",{"type":649,"tag":671,"props":857,"children":858},{},[859],{"type":654,"value":215},{"type":654,"value":861},".",{"type":649,"tag":650,"props":863,"children":864},{},[865],{"type":649,"tag":699,"props":866,"children":869},{"alt":701,"src":867,"title":868},"image://61864a07-3264-42a2-b193-74302e343166","A (very) simple template with parameters.",[],{"type":649,"tag":650,"props":871,"children":872},{},[873],{"type":654,"value":874},"When data is fed into this template, those parameters automatically adjust. The numbers will either get higher or lower, depending on the nature of the data. Just imagine someone sitting there, adjusting dials up and down, as the different bits of data come through.",{"type":649,"tag":650,"props":876,"children":877},{},[878],{"type":654,"value":879},"We'll look at an example in a second. But in simple terms: when we say that a machine is 'learning', we really just mean that these parameters are changing up and down.",{"type":649,"tag":650,"props":881,"children":882},{},[883],{"type":649,"tag":699,"props":884,"children":887},{"alt":701,"src":885,"title":886},"image://a198586f-bd2c-491c-8382-823794434830","Parameters changing up and down.",[],{"title":643,"searchDepth":20,"depth":20,"links":889},[],{"id":198,"data":199,"type":26,"maxContentLevel":19,"version":26,"parsed":891},{"data":892,"body":894,"toc":931},{"title":643,"description":893},"Here's that example we promised. Let's imagine you're building an AI model which can predict the price of a pizza based on its size.",{"type":646,"children":895},[896,900,905,913,918,923],{"type":649,"tag":650,"props":897,"children":898},{},[899],{"type":654,"value":893},{"type":649,"tag":650,"props":901,"children":902},{},[903],{"type":654,"value":904},"Your template has parameters for all the different sizes: 10-inch pizza, 11-inch pizza, 12-inch pizza, and so on. These parameters all have prices attached, which you've estimated yourself.",{"type":649,"tag":650,"props":906,"children":907},{},[908],{"type":649,"tag":699,"props":909,"children":912},{"alt":701,"src":910,"title":911},"image://9300acc9-f0fc-4285-aa7c-32325ceb510f","Estimated pizza prices.",[],{"type":649,"tag":650,"props":914,"children":915},{},[916],{"type":654,"value":917},"You then give the AI some menus from local pizza restaurants. As each menu comes through, it tweaks its parameters, and the prices of each pizza gradually go up or down.",{"type":649,"tag":650,"props":919,"children":920},{},[921],{"type":654,"value":922},"For example, if one menu showed a 10-inch pizza for $8.50, the AI would tweak that original parameter down a couple of notches. If the next menu showed a 10-inch pizza for $9.50, it tweaks the number up again. Every time it does this, it gets closer and closer to an average price for a 10-inch pizza in your area.",{"type":649,"tag":650,"props":924,"children":925},{},[926],{"type":649,"tag":699,"props":927,"children":930},{"alt":701,"src":928,"title":929},"image://9e5222e3-c4b8-416a-b8b9-24657cb27570","Adjusting prices up and down.",[],{"title":643,"searchDepth":20,"depth":20,"links":932},[],{"id":203,"data":204,"type":26,"maxContentLevel":19,"version":26,"reviews":207,"parsed":934},{"data":935,"body":937,"toc":969},{"title":643,"description":936},"By the end, the AI will have a new set of numbers: it dropped from $10.00 to $9.00 for a 10-inch, from $11.00 to $10.00 for an 11-inch, and from $12.00 to $11.00 for a 12-inch.",{"type":646,"children":938},[939,943,951,956],{"type":649,"tag":650,"props":940,"children":941},{},[942],{"type":654,"value":936},{"type":649,"tag":650,"props":944,"children":945},{},[946],{"type":649,"tag":699,"props":947,"children":950},{"alt":701,"src":948,"title":949},"image://e5bfef35-995a-49ee-a5cb-f098c43a3d8d","Pizza prices.",[],{"type":649,"tag":650,"props":952,"children":953},{},[954],{"type":654,"value":955},"In other words, this AI has 'learned' some average prices for pizzas being sold in your area. And if you asked it to estimate the cost of, say, a 14-inch pizza, it would hopefully give you a decent answer. In this case, maybe $13.00.",{"type":649,"tag":650,"props":957,"children":958},{},[959,961,967],{"type":654,"value":960},"You can check if that number is accurate by looking in another pizza menu. You glance through the options, and find that this particular pizza restaurant is selling 14-inch pizzas for $13.25. That's not ",{"type":649,"tag":962,"props":963,"children":964},"em",{},[965],{"type":654,"value":966},"exactly",{"type":654,"value":968}," the same as your AI's estimate, but all things considered, it's pretty close.",{"title":643,"searchDepth":20,"depth":20,"links":970},[],{"id":217,"data":218,"type":26,"maxContentLevel":19,"version":26,"reviews":221,"parsed":972},{"data":973,"body":975,"toc":1007},{"title":643,"description":974},"There's actually a name for this 'gap' between the AI's prediction ($13.00) and the real-life data ($13.25). Scientists call it the loss function, and it's an important part of machine learning.",{"type":646,"children":976},[977,989,994,999],{"type":649,"tag":650,"props":978,"children":979},{},[980,982,987],{"type":654,"value":981},"There's actually a name for this 'gap' between the AI's prediction ($13.00) and the real-life data ($13.25). Scientists call it the ",{"type":649,"tag":671,"props":983,"children":984},{},[985],{"type":654,"value":986},"loss function",{"type":654,"value":988},", and it's an important part of machine learning.",{"type":649,"tag":650,"props":990,"children":991},{},[992],{"type":654,"value":993},"The smaller the loss function, the better your machine has learned. In an ideal scenario, the loss function would be non-existent – your AI would have predicted $13.25 in the first place.",{"type":649,"tag":650,"props":995,"children":996},{},[997],{"type":654,"value":998},"To reduce the loss function for this particular AI, you could give it more menus to learn from. In theory, with every piece of data it encounters, it will get closer and closer to the 'truth'. It's not always that simple, but as a general rule, larger datasets usually lead to more effective AI models.",{"type":649,"tag":650,"props":1000,"children":1001},{},[1002],{"type":649,"tag":699,"props":1003,"children":1006},{"alt":701,"src":1004,"title":1005},"image://9a205d1e-fff1-4b1c-93e4-71d7062907ab","Pizza by Valerio Capello, English Wikipedia (CC BY-SA 3.0) \u003Chttp://creativecommons.org/licenses/by-sa/3.0/>, via Wikimedia Commons",[],{"title":643,"searchDepth":20,"depth":20,"links":1008},[],{"id":254,"data":255,"type":20,"version":26,"maxContentLevel":19,"summaryPage":257,"introPage":264,"pages":1010},[1011,1045,1087,1117],{"id":272,"data":273,"type":26,"maxContentLevel":19,"version":26,"reviews":276,"parsed":1012},{"data":1013,"body":1015,"toc":1043},{"title":643,"description":1014},"When you want an AI to learn from a dataset, like in that pizza example we talked about last time, there are a few different approaches you can take.",{"type":646,"children":1016},[1017,1021,1033,1038],{"type":649,"tag":650,"props":1018,"children":1019},{},[1020],{"type":654,"value":1014},{"type":649,"tag":650,"props":1022,"children":1023},{},[1024,1026,1031],{"type":654,"value":1025},"One of these approaches is ",{"type":649,"tag":671,"props":1027,"children":1028},{},[1029],{"type":654,"value":1030},"supervised learning",{"type":654,"value":1032},". This is when you train your AI using a labeled set of data. When we say 'labeled', we mean that the data has all been carefully arranged into something called input-output pairs.",{"type":649,"tag":650,"props":1034,"children":1035},{},[1036],{"type":654,"value":1037},"That's what we did with those pizzas last time. The input side of the pair was the size of each pizza (e.g. 10 inch) and the output side of the pair was the price of each pizza (e.g. $9.00).",{"type":649,"tag":650,"props":1039,"children":1040},{},[1041],{"type":654,"value":1042},"In this example, our labeled data is essentially just a list of these size-price (input-output) pairs.",{"title":643,"searchDepth":20,"depth":20,"links":1044},[],{"id":303,"data":304,"type":26,"maxContentLevel":19,"version":26,"parsed":1046},{"data":1047,"body":1049,"toc":1085},{"title":643,"description":1048},"An input-output pair will hopefully have some kind of relationship or rule. In our pizza example, the rule went like this: the output number (price) was always one less than the input number (size).",{"type":646,"children":1050},[1051,1063,1070,1075,1080],{"type":649,"tag":650,"props":1052,"children":1053},{},[1054,1056,1061],{"type":654,"value":1055},"An input-output pair will hopefully have some kind of relationship or rule. In our pizza example, the rule went like this: the output number (price) was always ",{"type":649,"tag":962,"props":1057,"children":1058},{},[1059],{"type":654,"value":1060},"one less",{"type":654,"value":1062}," than the input number (size).",{"type":649,"tag":650,"props":1064,"children":1065},{},[1066],{"type":649,"tag":699,"props":1067,"children":1069},{"alt":701,"src":1068,"title":949},"image://974fb9ca-d2dd-4247-aaee-ca0d4c526f2c",[],{"type":649,"tag":650,"props":1071,"children":1072},{},[1073],{"type":654,"value":1074},"In a lot of cases, we won't actually know that relationship or rule ourselves. Instead, we want the AI to look at the input-output pairs, and search for patterns that help it to establish the relationship or rule by itself.",{"type":649,"tag":650,"props":1076,"children":1077},{},[1078],{"type":654,"value":1079},"Once it's established the rule, it can start using it. For example, we could give an input number to our pizza AI: \"5-inch\". Using the rule above, the AI could predict the output number: \"$4.00\".",{"type":649,"tag":650,"props":1081,"children":1082},{},[1083],{"type":654,"value":1084},"In plainer language, we could ask the AI, \"How much do you think a 5-inch pizza would cost?\" and it could reply to you, \"It will probably cost $4.00\".",{"title":643,"searchDepth":20,"depth":20,"links":1086},[],{"id":308,"data":309,"type":26,"maxContentLevel":19,"version":26,"reviews":312,"parsed":1088},{"data":1089,"body":1091,"toc":1115},{"title":643,"description":1090},"Here's another example. We could give our AI a bunch of emails to look at. Half of these emails are labeled 'spam', and the other half are labeled 'not spam'. This time, the input is the email itself, and the output is either 'spam' or 'not spam'.",{"type":646,"children":1092},[1093,1097,1102,1107],{"type":649,"tag":650,"props":1094,"children":1095},{},[1096],{"type":654,"value":1090},{"type":649,"tag":650,"props":1098,"children":1099},{},[1100],{"type":654,"value":1101},"Using this labeled data, the AI can now learn some rules or patterns that distinguish spam from non-spam. For example, it might notice that an input email with more spelling mistakes, or strange punctuation, is typically linked to an output label saying 'spam'.",{"type":649,"tag":650,"props":1103,"children":1104},{},[1105],{"type":654,"value":1106},"Later, whenever we give this AI a random email, it can use this rule to sort it. Lots of spelling mistakes? Or unusual punctuation? The email is sent to spam.",{"type":649,"tag":650,"props":1108,"children":1109},{},[1110],{"type":649,"tag":699,"props":1111,"children":1114},{"alt":701,"src":1112,"title":1113},"image://4cabceb9-902b-487f-adfd-c2aac2e32121","Spam or not spam? Image: gmcgcc, CC BY-SA 4.0 \u003Chttps://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons",[],{"title":643,"searchDepth":20,"depth":20,"links":1116},[],{"id":333,"data":334,"type":26,"maxContentLevel":19,"version":26,"reviews":337,"parsed":1118},{"data":1119,"body":1121,"toc":1149},{"title":643,"description":1120},"Supervised learning is great in scenarios like that email example – when there's a clear pair of inputs and outputs.",{"type":646,"children":1122},[1123,1127,1132,1137,1144],{"type":649,"tag":650,"props":1124,"children":1125},{},[1126],{"type":654,"value":1120},{"type":649,"tag":650,"props":1128,"children":1129},{},[1130],{"type":654,"value":1131},"Here are some other examples: when you give your AI some text in one language (input) then get a translated version (output). When you play a piece of music to your AI (input) then get the title of that song (output).",{"type":649,"tag":650,"props":1133,"children":1134},{},[1135],{"type":654,"value":1136},"Also, that cat and dog example we mentioned earlier on. You can give it a photo of an animal (input) then get the name of that animal (output).",{"type":649,"tag":650,"props":1138,"children":1139},{},[1140],{"type":649,"tag":699,"props":1141,"children":1143},{"alt":701,"src":1142,"title":703},"image://af27fc29-bc2d-4547-96eb-053ed754f436",[],{"type":649,"tag":650,"props":1145,"children":1146},{},[1147],{"type":654,"value":1148},"This isn't the only approach to machine learning. We'll look at some other types later. But it's a very common approach right now, with thousands of different uses.",{"title":643,"searchDepth":20,"depth":20,"links":1150},[],{"id":357,"data":358,"type":20,"version":20,"maxContentLevel":19,"summaryPage":360,"introPage":367,"pages":1152},[1153,1187,1230,1260,1290],{"id":374,"data":375,"type":26,"maxContentLevel":19,"version":26,"reviews":378,"parsed":1154},{"data":1155,"body":1157,"toc":1185},{"title":643,"description":1156},"The main alternative to supervised learning is (can you guess it?) unsupervised learning.",{"type":646,"children":1158},[1159,1170,1175,1180],{"type":649,"tag":650,"props":1160,"children":1161},{},[1162,1164,1169],{"type":654,"value":1163},"The main alternative to supervised learning is (can you guess it?) ",{"type":649,"tag":671,"props":1165,"children":1166},{},[1167],{"type":654,"value":1168},"unsupervised learning",{"type":654,"value":861},{"type":649,"tag":650,"props":1171,"children":1172},{},[1173],{"type":654,"value":1174},"This time, you won't be giving your AI a set of nicely labeled pairs. Instead, you'll be giving it an unstructured pile of raw, unlabelled data.",{"type":649,"tag":650,"props":1176,"children":1177},{},[1178],{"type":654,"value":1179},"This data might actually follow some interesting rules and patterns. But as a human, you don't know what they are. The data is too messy, too large, too confusing. There's certainly nothing as intuitive here as a simple input-output pair.",{"type":649,"tag":650,"props":1181,"children":1182},{},[1183],{"type":654,"value":1184},"So you ask your AI \"are there any patterns in here?\" and see if it can learn anything useful.",{"title":643,"searchDepth":20,"depth":20,"links":1186},[],{"id":398,"data":399,"type":26,"maxContentLevel":19,"version":26,"parsed":1188},{"data":1189,"body":1191,"toc":1228},{"title":643,"description":1190},"Typically, an unsupervised learning model will use a technique called clustering. This technique involves sorting data into groups based on apparent similarities and differences.",{"type":646,"children":1192},[1193,1205,1210,1215,1220],{"type":649,"tag":650,"props":1194,"children":1195},{},[1196,1198,1203],{"type":654,"value":1197},"Typically, an unsupervised learning model will use a technique called ",{"type":649,"tag":671,"props":1199,"children":1200},{},[1201],{"type":654,"value":1202},"clustering",{"type":654,"value":1204},". This technique involves sorting data into groups based on apparent similarities and differences.",{"type":649,"tag":650,"props":1206,"children":1207},{},[1208],{"type":654,"value":1209},"It's like giving a child a handful of marbles, then asking them to sort them. They may start to sort them by size, or color, or opacity, or weight, or whatever other patterns they come up with. An AI can do the same with data, seeking patterns, then learning from them.",{"type":649,"tag":650,"props":1211,"children":1212},{},[1213],{"type":654,"value":1214},"Here's a real world example. Imagine that you run a gym. You ask an AI to look for patterns in a database of all your members.",{"type":649,"tag":650,"props":1216,"children":1217},{},[1218],{"type":654,"value":1219},"It finds some interesting clusters: people who live in the east of the city seem more interested in taking yoga classes, whereas people who live in the west of the city are more interested in taking spin classes. Why? You have no idea. But later, when you open a new gym in the east, you make sure it specializes in yoga.",{"type":649,"tag":650,"props":1221,"children":1222},{},[1223],{"type":649,"tag":699,"props":1224,"children":1227},{"alt":701,"src":1225,"title":1226},"image://450f2747-75d6-477c-9fbb-473fd3bd4f66","Clustered data. Image: BogdanShevchenko (CC BY-SA 4.0) \u003Chttps://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons",[],{"title":643,"searchDepth":20,"depth":20,"links":1229},[],{"id":403,"data":404,"type":26,"maxContentLevel":19,"version":26,"reviews":407,"parsed":1231},{"data":1232,"body":1234,"toc":1258},{"title":643,"description":1233},"Another technique that a model might use is association. This involves looking for connections between pieces of data. Often, these connections are causal: if X happens, then Y will happen next.",{"type":646,"children":1235},[1236,1248,1253],{"type":649,"tag":650,"props":1237,"children":1238},{},[1239,1241,1246],{"type":654,"value":1240},"Another technique that a model might use is ",{"type":649,"tag":671,"props":1242,"children":1243},{},[1244],{"type":654,"value":1245},"association",{"type":654,"value":1247},". This involves looking for connections between pieces of data. Often, these connections are causal: if X happens, then Y will happen next.",{"type":649,"tag":650,"props":1249,"children":1250},{},[1251],{"type":654,"value":1252},"As another example, imagine that you run a streaming service like Netflix. You ask an AI to look for patterns in a database of all your members. It finds an association: after watching a scary horror movie, it's common for people to calm down with their favorite rom com.",{"type":649,"tag":650,"props":1254,"children":1255},{},[1256],{"type":654,"value":1257},"This knowledge helps you adjust your service – you make sure that rom coms are typically suggested after horrors.",{"title":643,"searchDepth":20,"depth":20,"links":1259},[],{"id":425,"data":426,"type":26,"maxContentLevel":19,"version":20,"parsed":1261},{"data":1262,"body":1264,"toc":1288},{"title":643,"description":1263},"Unsupervised learning is incredibly powerful for exploratory analysis. It can identify all kinds of rules and patterns that you never would have thought of on your own. But remember: these rules will sometimes be bizarre and useless.",{"type":646,"children":1265},[1266,1270,1275,1283],{"type":649,"tag":650,"props":1267,"children":1268},{},[1269],{"type":654,"value":1263},{"type":649,"tag":650,"props":1271,"children":1272},{},[1273],{"type":654,"value":1274},"Your AI might notice that people born in November like movies about dogs, while people born in September prefer movies which feature cats – but only if those cats are black. For whatever reason, this pattern might genuinely be present in your data.",{"type":649,"tag":650,"props":1276,"children":1277},{},[1278],{"type":649,"tag":699,"props":1279,"children":1282},{"alt":701,"src":1280,"title":1281},"image://c370f4f2-a295-4f26-87e7-aece84df2053","Black cat by Dmitry Makeev (CC BY-SA 4.0) \u003Chttps://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons",[],{"type":649,"tag":650,"props":1284,"children":1285},{},[1286],{"type":654,"value":1287},"But is it useful? That's for you to decide.",{"title":643,"searchDepth":20,"depth":20,"links":1289},[],{"id":430,"data":431,"type":26,"maxContentLevel":19,"version":26,"reviews":434,"parsed":1291},{"data":1292,"body":1294,"toc":1310},{"title":643,"description":1293},"It's worth pointing out that unsupervised learning only works on massive datasets. And it's more computationally complex than supervised learning – it takes up a lot more power.",{"type":646,"children":1295},[1296,1300,1305],{"type":649,"tag":650,"props":1297,"children":1298},{},[1299],{"type":654,"value":1293},{"type":649,"tag":650,"props":1301,"children":1302},{},[1303],{"type":654,"value":1304},"There's a trade off here. Unsupervised learning is more work for a computer, but it's much less work for a human. Supervised learning, on the other hand, is easier for the computer, but it takes time and effort on the human side to sort and label the data.",{"type":649,"tag":650,"props":1306,"children":1307},{},[1308],{"type":654,"value":1309},"A scientist will generally choose between them based on the ultimate goal of their project. If they're looking to find some unexpected patterns in a giant dataset, they'll use unsupervised. If they just want to find some simple relationships between inputs and outputs, they'll use supervised learning instead.",{"title":643,"searchDepth":20,"depth":20,"links":1311},[],{"id":449,"data":450,"type":20,"version":26,"maxContentLevel":19,"summaryPage":452,"introPage":459,"pages":1313},[1314,1348,1383,1413],{"id":466,"data":467,"type":26,"maxContentLevel":19,"version":26,"reviews":470,"parsed":1315},{"data":1316,"body":1318,"toc":1346},{"title":643,"description":1317},"Along with supervised learning and unsupervised learning, another popular approach to machine learning is something called reinforcement learning.",{"type":646,"children":1319},[1320,1331,1336,1341],{"type":649,"tag":650,"props":1321,"children":1322},{},[1323,1325,1330],{"type":654,"value":1324},"Along with supervised learning and unsupervised learning, another popular approach to machine learning is something called ",{"type":649,"tag":671,"props":1326,"children":1327},{},[1328],{"type":654,"value":1329},"reinforcement learning",{"type":654,"value":861},{"type":649,"tag":650,"props":1332,"children":1333},{},[1334],{"type":654,"value":1335},"This one is pretty different from the others, as it doesn't require a dataset. Instead, you're going to be putting your AI in a closed environment, where it can learn via rewards and punishments.",{"type":649,"tag":650,"props":1337,"children":1338},{},[1339],{"type":654,"value":1340},"These rewards and punishments are just positive values and negative values which are assigned to different actions. Whenever the AI performs an action, the corresponding value will tell the AI whether to perform that action again.",{"type":649,"tag":650,"props":1342,"children":1343},{},[1344],{"type":654,"value":1345},"Fundamentally, it's just like operant conditioning. But instead of using rewards and punishments to train rats and monkeys, it's being used to train an AI.",{"title":643,"searchDepth":20,"depth":20,"links":1347},[],{"id":483,"data":484,"type":26,"maxContentLevel":19,"version":26,"reviews":487,"parsed":1349},{"data":1350,"body":1352,"toc":1381},{"title":643,"description":1351},"Reinforcement learning would be a great approach if you wanted to teach an AI model how to play a game of chess.",{"type":646,"children":1353},[1354,1358,1366,1371,1376],{"type":649,"tag":650,"props":1355,"children":1356},{},[1357],{"type":654,"value":1351},{"type":649,"tag":650,"props":1359,"children":1360},{},[1361],{"type":649,"tag":699,"props":1362,"children":1365},{"alt":701,"src":1363,"title":1364},"image://37632363-4caa-4564-9930-115482039eab","Chess by User:Cburnett (CC BY-SA 3.0) \u003Chttp://creativecommons.org/licenses/by-sa/3.0/>, via Wikimedia Commons",[],{"type":649,"tag":650,"props":1367,"children":1368},{},[1369],{"type":654,"value":1370},"You'd need to build an environment where the AI can play chess matches over and over and over. Each win is associated with a positive value; each loss is associated with a negative value.",{"type":649,"tag":650,"props":1372,"children":1373},{},[1374],{"type":654,"value":1375},"By the time it's played thousands of matches, this AI might have learned which sets of moves are most likely to lead to a positive, winning outcome. It might also have learned which moves to avoid, or which tactics are too risky to pull off.",{"type":649,"tag":650,"props":1377,"children":1378},{},[1379],{"type":654,"value":1380},"Reinforcement learning can also be used in other scenarios, like teaching a robot how to walk. Assign a positive value to a successful step, and a negative value if the robot falls to the floor.",{"title":643,"searchDepth":20,"depth":20,"links":1382},[],{"id":496,"data":497,"type":26,"maxContentLevel":19,"version":26,"reviews":500,"parsed":1384},{"data":1385,"body":1387,"toc":1411},{"title":643,"description":1386},"In general, you'd use reinforcement learning if you wanted your AI to learn how to deal with dynamic, real world situations. The kinds of situations that can't easily be summed up in a database.",{"type":646,"children":1388},[1389,1393,1398,1406],{"type":649,"tag":650,"props":1390,"children":1391},{},[1392],{"type":654,"value":1386},{"type":649,"tag":650,"props":1394,"children":1395},{},[1396],{"type":654,"value":1397},"Another example would be training a self-driving car. Instead of getting that AI to look for patterns in data about collisions and traffic codes, it might work better to build an artificial environment where the AI can simulate driving around, and learn through trial and error.",{"type":649,"tag":650,"props":1399,"children":1400},{},[1401],{"type":649,"tag":699,"props":1402,"children":1405},{"alt":701,"src":1403,"title":1404},"image://27294a6d-78b9-4ed8-aae7-9cee59c6c4b3","Self-driving car by Grendelkhan (CC BY-SA 4.0) \u003Chttps://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons",[],{"type":649,"tag":650,"props":1407,"children":1408},{},[1409],{"type":654,"value":1410},"Having said that, there might also be value in getting an AI model to learn some basic principles using supervised learning, then to graduate it to reinforcement learning later. This approach was actually how AlphaGo was trained to play the game of Go.",{"title":643,"searchDepth":20,"depth":20,"links":1412},[],{"id":511,"data":512,"type":26,"maxContentLevel":19,"version":26,"reviews":515,"parsed":1414},{"data":1415,"body":1417,"toc":1433},{"title":643,"description":1416},"AlphaGo initially learned from a dataset: it observed thousands of recordings of real-life moves by expert human Go players. This was a labeled dataset – the model could see when a set of moves (input) resulted in a victory (output).",{"type":646,"children":1418},[1419,1423,1428],{"type":649,"tag":650,"props":1420,"children":1421},{},[1422],{"type":654,"value":1416},{"type":649,"tag":650,"props":1424,"children":1425},{},[1426],{"type":654,"value":1427},"After absorbing enough basic rules and patterns through this supervised stage, the model moved to an environment where it played thousands of Go matches against versions of itself, getting better and better through reinforcement.",{"type":649,"tag":650,"props":1429,"children":1430},{},[1431],{"type":654,"value":1432},"This combined approach worked wonders. By the end, AlphaGo had learned enough to successfully defeat the greatest Go player in the world.",{"title":643,"searchDepth":20,"depth":20,"links":1434},[],{"id":548,"data":549,"type":20,"version":26,"maxContentLevel":19,"summaryPage":551,"introPage":558,"pages":1436},[1437,1478,1508,1545],{"id":566,"data":567,"type":26,"maxContentLevel":19,"version":26,"reviews":570,"parsed":1438},{"data":1439,"body":1441,"toc":1476},{"title":643,"description":1440},"A lot of people are (rightly) excited about modern machine learning.",{"type":646,"children":1442},[1443,1447,1452,1463,1471],{"type":649,"tag":650,"props":1444,"children":1445},{},[1446],{"type":654,"value":1440},{"type":649,"tag":650,"props":1448,"children":1449},{},[1450],{"type":654,"value":1451},"As we talked about earlier, it lets us build complex, powerful models without needing to programme all the rules of those models by hand. It also lets us build flexible, adaptive models, which can change and evolve, and handle unexpected problems.",{"type":649,"tag":650,"props":1453,"children":1454},{},[1455,1457,1462],{"type":654,"value":1456},"But there's one important challenge that we need to be aware of: ",{"type":649,"tag":671,"props":1458,"children":1459},{},[1460],{"type":654,"value":1461},"availability of data",{"type":654,"value":861},{"type":649,"tag":650,"props":1464,"children":1465},{},[1466],{"type":649,"tag":699,"props":1467,"children":1470},{"alt":701,"src":1468,"title":1469},"image://e3dcb218-1053-4f01-9fd3-7f22c128942a","Data. By Chiffre01 (CC BY-SA 4.0) \u003Chttps://creativecommons.org/licenses/by-sa/4.0>, via Wikimedia Commons",[],{"type":649,"tag":650,"props":1472,"children":1473},{},[1474],{"type":654,"value":1475},"While reinforcement learning doesn't need a dataset, supervised learning and unsupervised learning do. And this can't just be any dataset. It needs to be relevant to the task that you want your AI to be learning.",{"title":643,"searchDepth":20,"depth":20,"links":1477},[],{"id":583,"data":584,"type":26,"maxContentLevel":19,"version":26,"reviews":587,"parsed":1479},{"data":1480,"body":1482,"toc":1506},{"title":643,"description":1481},"As well as a relevant dataset, you'll also need a large dataset. AI learns best when it's able to repeat things thousands and thousands of times.",{"type":646,"children":1483},[1484,1496,1501],{"type":649,"tag":650,"props":1485,"children":1486},{},[1487,1489,1494],{"type":654,"value":1488},"As well as a relevant dataset, you'll also need a ",{"type":649,"tag":962,"props":1490,"children":1491},{},[1492],{"type":654,"value":1493},"large",{"type":654,"value":1495}," dataset. AI learns best when it's able to repeat things thousands and thousands of times.",{"type":649,"tag":650,"props":1497,"children":1498},{},[1499],{"type":654,"value":1500},"It's hard to put an exact number of the amount of data that you'll need. One general rule is that you'll need ten pieces of data (for example, ten photos) for each parameter in your model. If you had a model with 200 parameters? You'd need 2000 pieces of data.",{"type":649,"tag":650,"props":1502,"children":1503},{},[1504],{"type":654,"value":1505},"That rule is pretty arbitrary. It all depends on the model, and what you want it to learn. But these approximate numbers still help to highlight how much data is often needed.",{"title":643,"searchDepth":20,"depth":20,"links":1507},[],{"id":596,"data":597,"type":26,"maxContentLevel":19,"version":26,"reviews":600,"parsed":1509},{"data":1510,"body":1512,"toc":1543},{"title":643,"description":1511},"If you do manage to find enough usable data, you'll want to divide it into parts. The first part will be used as training data, which your AI model can learn from.",{"type":646,"children":1513},[1514,1526,1538],{"type":649,"tag":650,"props":1515,"children":1516},{},[1517,1519,1524],{"type":654,"value":1518},"If you do manage to find enough usable data, you'll want to divide it into parts. The first part will be used as ",{"type":649,"tag":671,"props":1520,"children":1521},{},[1522],{"type":654,"value":1523},"training data",{"type":654,"value":1525},", which your AI model can learn from.",{"type":649,"tag":650,"props":1527,"children":1528},{},[1529,1531,1536],{"type":654,"value":1530},"The second part will be some ",{"type":649,"tag":671,"props":1532,"children":1533},{},[1534],{"type":654,"value":1535},"test data",{"type":654,"value":1537},", which you can give to the AI after the learning stage, and check how well it performs.",{"type":649,"tag":650,"props":1539,"children":1540},{},[1541],{"type":654,"value":1542},"As another general rule: of your total data, 80% should be used for training, and 20% for testing. So if you had 2000 photos in total, you'd use 1600 to train your AI, and 400 to test it later on.",{"title":643,"searchDepth":20,"depth":20,"links":1544},[],{"id":618,"data":619,"type":26,"maxContentLevel":19,"version":26,"reviews":622,"parsed":1546},{"data":1547,"body":1549,"toc":1595},{"title":643,"description":1548},"If you can find all that data, that's amazing news. AI scientists often talk about the unreasonable effectiveness of data. They're basically saying, when you do have enough data, it will probably be more effective than you'd ever expect.",{"type":646,"children":1550},[1551,1570,1582,1587],{"type":649,"tag":650,"props":1552,"children":1553},{},[1554,1556,1561,1563,1568],{"type":654,"value":1555},"If you can find all that data, that's amazing news. AI scientists often talk about the ",{"type":649,"tag":671,"props":1557,"children":1558},{},[1559],{"type":654,"value":1560},"unreasonable effectiveness of data",{"type":654,"value":1562},". They're basically saying, when you ",{"type":649,"tag":962,"props":1564,"children":1565},{},[1566],{"type":654,"value":1567},"do",{"type":654,"value":1569}," have enough data, it will probably be more effective than you'd ever expect.",{"type":649,"tag":650,"props":1571,"children":1572},{},[1573,1575,1580],{"type":654,"value":1574},"But the opposite is true as well. When you ",{"type":649,"tag":962,"props":1576,"children":1577},{},[1578],{"type":654,"value":1579},"don't",{"type":654,"value":1581}," have that data, you're seriously going to struggle. It's an important drawback to modern AI. It's still pretty hard to develop models for problems that don't have much data.",{"type":649,"tag":650,"props":1583,"children":1584},{},[1585],{"type":654,"value":1586},"Imagine, for example, that you wanted to train an AI model to spot the symptoms of a rare disease. But you can only find about twenty case studies of people who have had that disease. That's nowhere near enough data – machine learning wouldn't be possible.",{"type":649,"tag":650,"props":1588,"children":1589},{},[1590],{"type":649,"tag":699,"props":1591,"children":1594},{"alt":701,"src":1592,"title":1593},"image://d968dfc2-1a18-4a2e-9179-6cd741e841c4","Medical data. Image via Pexels",[],{"title":643,"searchDepth":20,"depth":20,"links":1596},[],{"left":4,"top":4,"width":1598,"height":1598,"rotate":4,"vFlip":6,"hFlip":6,"body":1599},24,"\u003Cpath fill=\"none\" stroke=\"currentColor\" stroke-linecap=\"round\" stroke-linejoin=\"round\" stroke-width=\"2\" d=\"m9 18l6-6l-6-6\"/>",{"left":4,"top":4,"width":1598,"height":1598,"rotate":4,"vFlip":6,"hFlip":6,"body":1601},"\u003Cpath fill=\"none\" stroke=\"currentColor\" stroke-linecap=\"round\" stroke-linejoin=\"round\" stroke-width=\"2\" d=\"M4 5h16M4 12h16M4 19h16\"/>",1778179449351]