transcripts/IHZwWFHWa-w.json

{"message": {"transcript": [{"chunks": [{"end": 0.52, "start": 0.0, "text": "Last"}, {"end": 1.64, "start": 0.52, "text": "video"}, {"end": 1.92, "start": 1.64, "text": "I"}, {"end": 2.6, "start": 1.92, "text": "laid"}, {"end": 3.04, "start": 2.6, "text": "out"}, {"end": 3.44, "start": 3.04, "text": "the"}, {"end": 4.4, "start": 3.44, "text": "structure"}, {"end": 4.72, "start": 4.4, "text": "of"}, {"end": 5.16, "start": 4.72, "text": "a"}, {"end": 5.92, "start": 5.16, "text": "neural"}, {"end": 6.68, "start": 5.92, "text": "network."}, {"end": 7.24, "start": 6.68, "text": "I'll"}, {"end": 7.48, "start": 7.24, "text": "give"}, {"end": 7.56, "start": 7.48, "text": "a"}, {"end": 7.6, "start": 7.56, "text": "quick"}, {"end": 8.24, "start": 7.6, "text": "recap"}, {"end": 8.4, "start": 8.24, "text": "here,"}, {"end": 8.48, "start": 8.4, "text": "just"}, {"end": 8.76, "start": 8.48, "text": "so"}, {"end": 8.8, "start": 8.76, "text": "that"}, {"end": 9.32, "start": 8.8, "text": "it's"}, {"end": 9.68, "start": 9.32, "text": "fresh"}, {"end": 9.76, "start": 9.68, "text": "in"}, {"end": 9.84, "start": 9.76, "text": "our"}, {"end": 10.28, "start": 9.84, "text": "minds,"}, {"end": 10.32, "start": 10.28, "text": "and"}, {"end": 10.32, "start": 10.32, "text": "then"}, {"end": 10.32, "start": 10.32, "text": "I"}, {"end": 10.32, "start": 10.32, "text": "have"}, {"end": 10.44, "start": 10.32, "text": "two"}, {"end": 11.0, "start": 10.44, "text": "main"}, {"end": 11.44, "start": 11.0, "text": "goals"}, {"end": 12.0, "start": 11.44, "text": "for"}, {"end": 12.16, "start": 12.0, "text": "this"}, {"end": 12.8, "start": 12.16, "text": "video."}, {"end": 13.12, "start": 12.8, "text": "The"}, {"end": 13.48, "start": 13.12, "text": "first"}, {"end": 13.64, "start": 13.48, "text": "is"}, {"end": 13.88, "start": 13.64, "text": "to"}, {"end": 14.24, "start": 13.88, "text": "introduce"}, {"end": 14.44, "start": 14.24, "text": "the"}, {"end": 14.84, "start": 14.44, "text": "idea"}, {"end": 14.84, "start": 14.84, "text": "of"}, {"end": 15.36, "start": 14.84, "text": "gradient"}, {"end": 16.04, "start": 15.36, "text": "descent,"}, {"end": 16.36, "start": 16.04, "text": "which"}, {"end": 16.88, "start": 16.36, "text": "underlies"}, {"end": 17.16, "start": 16.88, "text": "not"}, {"end": 17.24, "start": 17.16, "text": "only"}, {"end": 17.4, "start": 17.24, "text": "how"}, {"end": 17.68, "start": 17.4, "text": "neural"}, {"end": 18.08, "start": 17.68, "text": "networks"}, {"end": 18.28, "start": 18.08, "text": "learn,"}, {"end": 18.36, "start": 18.28, "text": "but"}, {"end": 18.6, "start": 18.36, "text": "how"}, {"end": 18.64, "start": 18.6, "text": "a"}, {"end": 18.8, "start": 18.64, "text": "lot"}, {"end": 19.12, "start": 18.8, "text": "of"}, {"end": 19.28, "start": 19.12, "text": "other"}, {"end": 19.64, "start": 19.28, "text": "machine"}, {"end": 19.88, "start": 19.64, "text": "learning"}, {"end": 20.28, "start": 19.88, "text": "works"}, {"end": 20.56, "start": 20.28, "text": "as"}, {"end": 20.64, "start": 20.56, "text": "well."}, {"end": 21.08, "start": 20.64, "text": "Then"}, {"end": 21.36, "start": 21.08, "text": "after"}, {"end": 21.6, "start": 21.36, "text": "that"}, {"end": 21.72, "start": 21.6, "text": "we're"}, {"end": 21.92, "start": 21.72, "text": "going"}, {"end": 21.92, "start": 21.92, "text": "to"}, {"end": 21.92, "start": 21.92, "text": "dig"}, {"end": 21.92, "start": 21.92, "text": "in"}, {"end": 21.96, "start": 21.92, "text": "a"}, {"end": 22.48, "start": 21.96, "text": "little"}, {"end": 22.68, "start": 22.48, "text": "more"}, {"end": 22.68, "start": 22.68, "text": "to"}, {"end": 22.96, "start": 22.68, "text": "how"}, {"end": 23.16, "start": 22.96, "text": "this"}, {"end": 23.8, "start": 23.16, "text": "particular"}, {"end": 24.08, "start": 23.8, "text": "network"}, {"end": 24.76, "start": 24.08, "text": "performs,"}, {"end": 24.84, "start": 24.76, "text": "and"}, {"end": 24.92, "start": 24.84, "text": "what"}, {"end": 25.44, "start": 24.92, "text": "those"}, {"end": 25.64, "start": 25.44, "text": "hidden"}, {"end": 26.2, "start": 25.64, "text": "layers"}, {"end": 26.52, "start": 26.2, "text": "of"}, {"end": 27.32, "start": 26.52, "text": "neurons"}, {"end": 27.64, "start": 27.32, "text": "end"}, {"end": 28.0, "start": 27.64, "text": "up"}, {"end": 28.76, "start": 28.0, "text": "actually"}, {"end": 28.84, "start": 28.76, "text": "looking"}, {"end": 29.08, "start": 28.84, "text": "for."}, {"end": 29.2, "start": 29.08, "text": "As"}, {"end": 29.4, "start": 29.2, "text": "a"}, {"end": 29.96, "start": 29.4, "text": "reminder,"}], "text": " Last video I laid out the structure of a neural network. I'll give a quick recap here, just so that it's fresh in our minds, and then I have two main goals for this video. The first is to introduce the idea of gradient descent, which underlies not only how neural networks learn, but how a lot of other machine learning works as well. Then after that we're going to dig in a little more to how this particular network performs, and what those hidden layers of neurons end up actually looking for. As a reminder,"}, {"chunks": [{"end": 30.24, "start": 30.0, "text": "Our"}, {"end": 30.64, "start": 30.24, "text": "goal"}, {"end": 31.08, "start": 30.64, "text": "here"}, {"end": 31.16, "start": 31.08, "text": "is"}, {"end": 31.16, "start": 31.16, "text": "the"}, {"end": 31.68, "start": 31.16, "text": "classic"}, {"end": 32.28, "start": 31.68, "text": "example"}, {"end": 32.56, "start": 32.28, "text": "of"}, {"end": 33.16, "start": 32.56, "text": "handwritten"}, {"end": 33.48, "start": 33.16, "text": "digit"}, {"end": 34.08, "start": 33.48, "text": "recognition,"}, {"end": 34.68, "start": 34.08, "text": "the"}, {"end": 35.480000000000004, "start": 34.68, "text": "hello"}, {"end": 35.8, "start": 35.480000000000004, "text": "world"}, {"end": 35.96, "start": 35.8, "text": "of"}, {"end": 36.04, "start": 35.96, "text": "neural"}, {"end": 36.4, "start": 36.04, "text": "networks."}, {"end": 36.92, "start": 36.4, "text": "These"}, {"end": 37.72, "start": 36.92, "text": "digits"}, {"end": 38.04, "start": 37.72, "text": "are"}, {"end": 38.36, "start": 38.04, "text": "rendered"}, {"end": 38.72, "start": 38.36, "text": "on"}, {"end": 38.88, "start": 38.72, "text": "a"}, {"end": 39.519999999999996, "start": 38.88, "text": "28"}, {"end": 39.72, "start": 39.519999999999996, "text": "by"}, {"end": 39.88, "start": 39.72, "text": "28"}, {"end": 40.12, "start": 39.88, "text": "pixel"}, {"end": 40.2, "start": 40.12, "text": "grid,"}, {"end": 40.36, "start": 40.2, "text": "each"}, {"end": 40.76, "start": 40.36, "text": "pixel"}, {"end": 40.88, "start": 40.76, "text": "with"}, {"end": 41.24, "start": 40.88, "text": "some"}, {"end": 41.8, "start": 41.24, "text": "grayscale"}, {"end": 42.16, "start": 41.8, "text": "value"}, {"end": 42.56, "start": 42.16, "text": "between"}, {"end": 43.12, "start": 42.56, "text": "0"}, {"end": 43.56, "start": 43.12, "text": "and"}, {"end": 43.88, "start": 43.56, "text": "1."}, {"end": 44.480000000000004, "start": 43.88, "text": "Those"}, {"end": 44.56, "start": 44.480000000000004, "text": "are"}, {"end": 44.56, "start": 44.56, "text": "what"}, {"end": 44.68, "start": 44.56, "text": "determine"}, {"end": 45.24, "start": 44.68, "text": "the"}, {"end": 46.0, "start": 45.24, "text": "activations"}, {"end": 46.2, "start": 46.0, "text": "of"}, {"end": 47.44, "start": 46.2, "text": "784"}, {"end": 48.120000000000005, "start": 47.44, "text": "neurons"}, {"end": 48.36, "start": 48.120000000000005, "text": "in"}, {"end": 48.519999999999996, "start": 48.36, "text": "the"}, {"end": 49.0, "start": 48.519999999999996, "text": "input"}, {"end": 49.4, "start": 49.0, "text": "layer"}, {"end": 49.4, "start": 49.4, "text": "of"}, {"end": 49.6, "start": 49.4, "text": "the"}, {"end": 50.32, "start": 49.6, "text": "network."}, {"end": 51.08, "start": 50.32, "text": "And"}, {"end": 51.56, "start": 51.08, "text": "then"}, {"end": 51.72, "start": 51.56, "text": "the"}, {"end": 52.32, "start": 51.72, "text": "activation"}, {"end": 52.4, "start": 52.32, "text": "for"}, {"end": 52.44, "start": 52.4, "text": "each"}, {"end": 52.760000000000005, "start": 52.44, "text": "neuron"}, {"end": 52.8, "start": 52.760000000000005, "text": "in"}, {"end": 53.08, "start": 52.8, "text": "the"}, {"end": 53.8, "start": 53.08, "text": "following"}, {"end": 54.36, "start": 53.8, "text": "layers"}, {"end": 54.519999999999996, "start": 54.36, "text": "is"}, {"end": 54.760000000000005, "start": 54.519999999999996, "text": "based"}, {"end": 54.8, "start": 54.760000000000005, "text": "on"}, {"end": 55.08, "start": 54.8, "text": "a"}, {"end": 55.4, "start": 55.08, "text": "weighted"}, {"end": 56.16, "start": 55.4, "text": "sum"}, {"end": 56.36, "start": 56.16, "text": "of"}, {"end": 56.760000000000005, "start": 56.36, "text": "all"}, {"end": 57.04, "start": 56.760000000000005, "text": "the"}, {"end": 57.8, "start": 57.04, "text": "activations"}, {"end": 58.120000000000005, "start": 57.8, "text": "in"}, {"end": 58.879999999999995, "start": 58.120000000000005, "text": "the"}, {"end": 59.56, "start": 58.879999999999995, "text": "previous"}, {"end": 59.96, "start": 59.56, "text": "layer,"}], "text": " Our goal here is the classic example of handwritten digit recognition, the hello world of neural networks. These digits are rendered on a 28 by 28 pixel grid, each pixel with some grayscale value between 0 and 1. Those are what determine the activations of 784 neurons in the input layer of the network. And then the activation for each neuron in the following layers is based on a weighted sum of all the activations in the previous layer,"}, {"chunks": [{"end": 60.4, "start": 60.0, "text": "Then"}, {"end": 60.4, "start": 60.4, "text": "you"}, {"end": 60.96, "start": 60.4, "text": "compose"}, {"end": 60.96, "start": 60.96, "text": "that"}, {"end": 61.6, "start": 60.96, "text": "sum"}, {"end": 62.04, "start": 61.6, "text": "with"}, {"end": 62.52, "start": 62.04, "text": "some"}, {"end": 62.96, "start": 62.52, "text": "other"}, {"end": 63.88, "start": 62.96, "text": "function,"}, {"end": 64.4, "start": 63.88, "text": "like"}, {"end": 64.88, "start": 64.4, "text": "the"}, {"end": 65.52, "start": 64.88, "text": "sigmoid"}, {"end": 66.4, "start": 65.52, "text": "squishification"}, {"end": 66.68, "start": 66.4, "text": "or"}, {"end": 66.72, "start": 66.68, "text": "a"}, {"end": 67.16, "start": 66.72, "text": "ReLU,"}, {"end": 67.44, "start": 67.16, "text": "the"}, {"end": 67.88, "start": 67.44, "text": "way"}, {"end": 68.24, "start": 67.88, "text": "I"}, {"end": 68.48, "start": 68.24, "text": "walked"}, {"end": 68.48, "start": 68.48, "text": "through"}, {"end": 68.52, "start": 68.48, "text": "last"}, {"end": 69.24, "start": 68.52, "text": "video."}, {"end": 69.48, "start": 69.24, "text": "In"}, {"end": 70.0, "start": 69.48, "text": "total,"}, {"end": 70.12, "start": 70.0, "text": "given"}, {"end": 70.52, "start": 70.12, "text": "the"}, {"end": 70.92, "start": 70.52, "text": "somewhat"}, {"end": 71.36, "start": 70.92, "text": "arbitrary"}, {"end": 71.88, "start": 71.36, "text": "choice"}, {"end": 71.88, "start": 71.88, "text": "of"}, {"end": 72.16, "start": 71.88, "text": "two"}, {"end": 72.48, "start": 72.16, "text": "hidden"}, {"end": 72.84, "start": 72.48, "text": "layers"}, {"end": 72.88, "start": 72.84, "text": "here"}, {"end": 73.24, "start": 72.88, "text": "with"}, {"end": 73.48, "start": 73.24, "text": "16"}, {"end": 74.28, "start": 73.48, "text": "neurons"}, {"end": 74.84, "start": 74.28, "text": "each,"}, {"end": 75.24, "start": 74.84, "text": "the"}, {"end": 75.88, "start": 75.24, "text": "network"}, {"end": 76.28, "start": 75.88, "text": "has"}, {"end": 76.48, "start": 76.28, "text": "about"}, {"end": 77.08, "start": 76.48, "text": "13,000"}, {"end": 77.4, "start": 77.08, "text": "weights"}, {"end": 77.44, "start": 77.4, "text": "and"}, {"end": 78.03999999999999, "start": 77.44, "text": "biases"}, {"end": 78.32, "start": 78.03999999999999, "text": "that"}, {"end": 78.4, "start": 78.32, "text": "we"}, {"end": 78.48, "start": 78.4, "text": "can"}, {"end": 78.96000000000001, "start": 78.48, "text": "adjust,"}, {"end": 79.44, "start": 78.96000000000001, "text": "and"}, {"end": 79.84, "start": 79.44, "text": "it's"}, {"end": 80.24, "start": 79.84, "text": "these"}, {"end": 80.92, "start": 80.24, "text": "values"}, {"end": 81.08, "start": 80.92, "text": "that"}, {"end": 81.56, "start": 81.08, "text": "determine"}, {"end": 81.84, "start": 81.56, "text": "what"}, {"end": 82.56, "start": 81.84, "text": "exactly"}, {"end": 83.08, "start": 82.56, "text": "the"}, {"end": 83.56, "start": 83.08, "text": "network"}, {"end": 83.96000000000001, "start": 83.56, "text": "actually"}, {"end": 84.48, "start": 83.96000000000001, "text": "does."}, {"end": 84.92, "start": 84.48, "text": "Then"}, {"end": 85.12, "start": 84.92, "text": "what"}, {"end": 85.2, "start": 85.12, "text": "we"}, {"end": 85.24, "start": 85.2, "text": "mean"}, {"end": 85.8, "start": 85.24, "text": "when"}, {"end": 86.03999999999999, "start": 85.8, "text": "we"}, {"end": 86.32, "start": 86.03999999999999, "text": "say"}, {"end": 86.36, "start": 86.32, "text": "that"}, {"end": 86.52, "start": 86.36, "text": "this"}, {"end": 86.96000000000001, "start": 86.52, "text": "network"}, {"end": 87.8, "start": 86.96000000000001, "text": "classifies"}, {"end": 88.03999999999999, "start": 87.8, "text": "a"}, {"end": 88.4, "start": 88.03999999999999, "text": "given"}, {"end": 88.68, "start": 88.4, "text": "digit"}, {"end": 88.92, "start": 88.68, "text": "is"}, {"end": 88.96000000000001, "start": 88.92, "text": "that"}, {"end": 89.2, "start": 88.96000000000001, "text": "the"}, {"end": 89.72, "start": 89.2, "text": "brightest"}, {"end": 89.96000000000001, "start": 89.72, "text": "of"}], "text": " Then you compose that sum with some other function, like the sigmoid squishification or a ReLU, the way I walked through last video. In total, given the somewhat arbitrary choice of two hidden layers here with 16 neurons each, the network has about 13,000 weights and biases that we can adjust, and it's these values that determine what exactly the network actually does. Then what we mean when we say that this network classifies a given digit is that the brightest of"}, {"chunks": [{"end": 90.16, "start": 90.0, "text": "Those"}, {"end": 90.48, "start": 90.16, "text": "10"}, {"end": 91.04, "start": 90.48, "text": "neurons"}, {"end": 91.04, "start": 91.04, "text": "in"}, {"end": 91.16, "start": 91.04, "text": "the"}, {"end": 91.52, "start": 91.16, "text": "final"}, {"end": 91.88, "start": 91.52, "text": "layer"}, {"end": 92.0, "start": 91.88, "text": "correspond"}, {"end": 92.0, "start": 92.0, "text": "to"}, {"end": 92.0, "start": 92.0, "text": "that"}, {"end": 92.04, "start": 92.0, "text": "digit."}, {"end": 92.44, "start": 92.04, "text": "And"}, {"end": 92.6, "start": 92.44, "text": "remember,"}, {"end": 92.92, "start": 92.6, "text": "the"}, {"end": 93.36, "start": 92.92, "text": "motivation"}, {"end": 93.52, "start": 93.36, "text": "that"}, {"end": 93.92, "start": 93.52, "text": "we"}, {"end": 94.64, "start": 93.92, "text": "had"}, {"end": 94.76, "start": 94.64, "text": "in"}, {"end": 95.4, "start": 94.76, "text": "mind"}, {"end": 95.68, "start": 95.4, "text": "here"}, {"end": 95.92, "start": 95.68, "text": "for"}, {"end": 96.48, "start": 95.92, "text": "the"}, {"end": 96.8, "start": 96.48, "text": "layered"}, {"end": 97.72, "start": 96.8, "text": "structure"}, {"end": 97.88, "start": 97.72, "text": "was"}, {"end": 98.32, "start": 97.88, "text": "that"}, {"end": 99.0, "start": 98.32, "text": "maybe"}, {"end": 99.28, "start": 99.0, "text": "the"}, {"end": 99.68, "start": 99.28, "text": "second"}, {"end": 100.03999999999999, "start": 99.68, "text": "layer"}, {"end": 100.03999999999999, "start": 100.03999999999999, "text": "could"}, {"end": 100.03999999999999, "start": 100.03999999999999, "text": "pick"}, {"end": 100.12, "start": 100.03999999999999, "text": "up"}, {"end": 100.12, "start": 100.12, "text": "on"}, {"end": 100.16, "start": 100.12, "text": "the"}, {"end": 100.68, "start": 100.16, "text": "edges,"}, {"end": 101.8, "start": 100.68, "text": "and"}, {"end": 102.08, "start": 101.8, "text": "the"}, {"end": 102.32, "start": 102.08, "text": "third"}, {"end": 102.64, "start": 102.32, "text": "layer"}, {"end": 102.8, "start": 102.64, "text": "might"}, {"end": 102.8, "start": 102.8, "text": "pick"}, {"end": 102.8, "start": 102.8, "text": "up"}, {"end": 102.92, "start": 102.8, "text": "on"}, {"end": 104.0, "start": 102.92, "text": "patterns"}, {"end": 104.16, "start": 104.0, "text": "like"}, {"end": 104.48, "start": 104.16, "text": "loops"}, {"end": 104.72, "start": 104.48, "text": "and"}, {"end": 105.12, "start": 104.72, "text": "lines,"}, {"end": 105.32, "start": 105.12, "text": "and"}, {"end": 105.44, "start": 105.32, "text": "the"}, {"end": 105.8, "start": 105.44, "text": "last"}, {"end": 106.0, "start": 105.8, "text": "one"}, {"end": 106.12, "start": 106.0, "text": "could"}, {"end": 106.36, "start": 106.12, "text": "just"}, {"end": 106.56, "start": 106.36, "text": "piece"}, {"end": 106.92, "start": 106.56, "text": "together"}, {"end": 107.12, "start": 106.92, "text": "those"}, {"end": 107.64, "start": 107.12, "text": "patterns"}, {"end": 107.88, "start": 107.64, "text": "to"}, {"end": 108.36, "start": 107.88, "text": "recognize"}, {"end": 109.72, "start": 108.36, "text": "digits."}, {"end": 109.88, "start": 109.72, "text": "So"}, {"end": 110.4, "start": 109.88, "text": "here,"}, {"end": 110.68, "start": 110.4, "text": "we"}, {"end": 111.0, "start": 110.68, "text": "learn"}, {"end": 111.2, "start": 111.0, "text": "how"}, {"end": 111.48, "start": 111.2, "text": "the"}, {"end": 112.16, "start": 111.48, "text": "network"}, {"end": 112.6, "start": 112.16, "text": "learns."}, {"end": 112.64, "start": 112.6, "text": "What"}, {"end": 112.72, "start": 112.64, "text": "we"}, {"end": 113.0, "start": 112.72, "text": "want"}, {"end": 113.03999999999999, "start": 113.0, "text": "is"}, {"end": 113.68, "start": 113.03999999999999, "text": "an"}, {"end": 114.08, "start": 113.68, "text": "algorithm"}, {"end": 114.4, "start": 114.08, "text": "where"}, {"end": 114.4, "start": 114.4, "text": "you"}, {"end": 114.96000000000001, "start": 114.4, "text": "can"}, {"end": 115.2, "start": 114.96000000000001, "text": "show"}, {"end": 115.4, "start": 115.2, "text": "this"}, {"end": 115.76, "start": 115.4, "text": "network"}, {"end": 115.88, "start": 115.76, "text": "a"}, {"end": 116.32, "start": 115.88, "text": "whole"}, {"end": 116.68, "start": 116.32, "text": "bunch"}, {"end": 116.92, "start": 116.68, "text": "of"}, {"end": 117.28, "start": 116.92, "text": "training"}, {"end": 117.6, "start": 117.28, "text": "data,"}, {"end": 117.76, "start": 117.6, "text": "which"}, {"end": 117.76, "start": 117.76, "text": "comes"}, {"end": 117.76, "start": 117.76, "text": "in"}, {"end": 118.12, "start": 117.76, "text": "the"}, {"end": 118.64, "start": 118.12, "text": "form"}, {"end": 118.76, "start": 118.64, "text": "of"}, {"end": 118.76, "start": 118.76, "text": "a"}, {"end": 119.08, "start": 118.76, "text": "bunch"}, {"end": 119.08, "start": 119.08, "text": "of"}, {"end": 119.44, "start": 119.08, "text": "different"}, {"end": 119.96000000000001, "start": 119.44, "text": "images"}], "text": " Those 10 neurons in the final layer correspond to that digit. And remember, the motivation that we had in mind here for the layered structure was that maybe the second layer could pick up on the edges, and the third layer might pick up on patterns like loops and lines, and the last one could just piece together those patterns to recognize digits. So here, we learn how the network learns. What we want is an algorithm where you can show this network a whole bunch of training data, which comes in the form of a bunch of different images"}, {"chunks": [{"end": 120.6, "start": 120.0, "text": "handwritten"}, {"end": 121.04, "start": 120.6, "text": "digits,"}, {"end": 121.28, "start": 121.04, "text": "along"}, {"end": 121.28, "start": 121.28, "text": "with"}, {"end": 121.28, "start": 121.28, "text": "labels"}, {"end": 121.8, "start": 121.28, "text": "for"}, {"end": 122.32, "start": 121.8, "text": "what"}, {"end": 122.72, "start": 122.32, "text": "they're"}, {"end": 123.0, "start": 122.72, "text": "supposed"}, {"end": 123.2, "start": 123.0, "text": "to"}, {"end": 123.52, "start": 123.2, "text": "be,"}, {"end": 124.2, "start": 123.52, "text": "and"}, {"end": 124.52, "start": 124.2, "text": "it'll"}, {"end": 125.0, "start": 124.52, "text": "adjust"}, {"end": 125.28, "start": 125.0, "text": "those"}, {"end": 126.16, "start": 125.28, "text": "13,000"}, {"end": 126.48, "start": 126.16, "text": "weights"}, {"end": 126.52, "start": 126.48, "text": "and"}, {"end": 127.2, "start": 126.52, "text": "biases"}, {"end": 127.24, "start": 127.2, "text": "so"}, {"end": 127.52, "start": 127.24, "text": "as"}, {"end": 127.88, "start": 127.52, "text": "to"}, {"end": 128.16, "start": 127.88, "text": "improve"}, {"end": 128.4, "start": 128.16, "text": "its"}, {"end": 129.04, "start": 128.4, "text": "performance"}, {"end": 129.28, "start": 129.04, "text": "on"}, {"end": 129.36, "start": 129.28, "text": "the"}, {"end": 129.76, "start": 129.36, "text": "training"}, {"end": 130.76, "start": 129.76, "text": "data."}, {"end": 131.64, "start": 130.76, "text": "Hopefully,"}, {"end": 131.88, "start": 131.64, "text": "this"}, {"end": 132.12, "start": 131.88, "text": "layered"}, {"end": 132.72, "start": 132.12, "text": "structure"}, {"end": 132.88, "start": 132.72, "text": "will"}, {"end": 133.04, "start": 132.88, "text": "mean"}, {"end": 133.36, "start": 133.04, "text": "that"}, {"end": 133.8, "start": 133.36, "text": "what"}, {"end": 133.84, "start": 133.8, "text": "it"}, {"end": 133.96, "start": 133.84, "text": "learns"}, {"end": 134.8, "start": 133.96, "text": "generalizes"}, {"end": 135.2, "start": 134.8, "text": "to"}, {"end": 135.56, "start": 135.2, "text": "images"}, {"end": 135.92, "start": 135.56, "text": "beyond"}, {"end": 136.16, "start": 135.92, "text": "that"}, {"end": 136.52, "start": 136.16, "text": "training"}, {"end": 136.92000000000002, "start": 136.52, "text": "data."}, {"end": 137.04, "start": 136.92000000000002, "text": "And"}, {"end": 137.76, "start": 137.04, "text": "the"}, {"end": 138.07999999999998, "start": 137.76, "text": "way"}, {"end": 138.07999999999998, "start": 138.07999999999998, "text": "we"}, {"end": 138.36, "start": 138.07999999999998, "text": "test"}, {"end": 138.64, "start": 138.36, "text": "that"}, {"end": 138.64, "start": 138.64, "text": "is"}, {"end": 138.68, "start": 138.64, "text": "that"}, {"end": 138.72, "start": 138.68, "text": "after"}, {"end": 138.76, "start": 138.72, "text": "you"}, {"end": 139.2, "start": 138.76, "text": "train"}, {"end": 139.24, "start": 139.2, "text": "the"}, {"end": 139.88, "start": 139.24, "text": "network,"}, {"end": 140.72, "start": 139.88, "text": "you"}, {"end": 141.12, "start": 140.72, "text": "show"}, {"end": 141.2, "start": 141.12, "text": "it"}, {"end": 141.6, "start": 141.2, "text": "more"}, {"end": 141.76, "start": 141.6, "text": "labeled"}, {"end": 142.2, "start": 141.76, "text": "data"}, {"end": 142.24, "start": 142.2, "text": "that"}, {"end": 142.52, "start": 142.24, "text": "it's"}, {"end": 142.68, "start": 142.52, "text": "never"}, {"end": 142.96, "start": 142.68, "text": "seen"}, {"end": 143.16, "start": 142.96, "text": "before,"}, {"end": 143.8, "start": 143.16, "text": "and"}, {"end": 143.92000000000002, "start": 143.8, "text": "you"}, {"end": 144.24, "start": 143.92000000000002, "text": "see"}, {"end": 144.48, "start": 144.24, "text": "how"}, {"end": 145.12, "start": 144.48, "text": "accurately"}, {"end": 145.12, "start": 145.12, "text": "it"}, {"end": 145.76, "start": 145.12, "text": "classifies"}, {"end": 146.0, "start": 145.76, "text": "those"}, {"end": 146.32, "start": 146.0, "text": "new"}, {"end": 149.96, "start": 146.32, "text": "images."}], "text": " handwritten digits, along with labels for what they're supposed to be, and it'll adjust those 13,000 weights and biases so as to improve its performance on the training data. Hopefully, this layered structure will mean that what it learns generalizes to images beyond that training data. And the way we test that is that after you train the network, you show it more labeled data that it's never seen before, and you see how accurately it classifies those new images."}, {"chunks": [{"end": 151.04, "start": 150.0, "text": "Fortunately"}, {"end": 151.6, "start": 151.04, "text": "for"}, {"end": 152.08, "start": 151.6, "text": "us,"}, {"end": 152.16, "start": 152.08, "text": "and"}, {"end": 152.2, "start": 152.16, "text": "what"}, {"end": 152.4, "start": 152.2, "text": "makes"}, {"end": 152.6, "start": 152.4, "text": "this"}, {"end": 152.76, "start": 152.6, "text": "such"}, {"end": 152.8, "start": 152.76, "text": "a"}, {"end": 153.48, "start": 152.8, "text": "common"}, {"end": 154.2, "start": 153.48, "text": "example"}, {"end": 154.4, "start": 154.2, "text": "to"}, {"end": 154.4, "start": 154.4, "text": "start"}, {"end": 154.44, "start": 154.4, "text": "with,"}, {"end": 154.56, "start": 154.44, "text": "is"}, {"end": 154.72, "start": 154.56, "text": "that"}, {"end": 154.96, "start": 154.72, "text": "the"}, {"end": 155.0, "start": 154.96, "text": "good"}, {"end": 155.24, "start": 155.0, "text": "people"}, {"end": 155.6, "start": 155.24, "text": "behind"}, {"end": 155.76, "start": 155.6, "text": "the"}, {"end": 156.2, "start": 155.76, "text": "MNIST"}, {"end": 156.84, "start": 156.2, "text": "database"}, {"end": 157.0, "start": 156.84, "text": "have"}, {"end": 157.16, "start": 157.0, "text": "put"}, {"end": 157.48, "start": 157.16, "text": "together"}, {"end": 157.72, "start": 157.48, "text": "a"}, {"end": 158.2, "start": 157.72, "text": "collection"}, {"end": 158.32, "start": 158.2, "text": "of"}, {"end": 158.88, "start": 158.32, "text": "tens"}, {"end": 159.0, "start": 158.88, "text": "of"}, {"end": 159.8, "start": 159.0, "text": "thousands"}, {"end": 159.96, "start": 159.8, "text": "of"}, {"end": 160.32, "start": 159.96, "text": "handwritten"}, {"end": 160.64, "start": 160.32, "text": "digit"}, {"end": 161.44, "start": 160.64, "text": "images,"}, {"end": 161.64, "start": 161.44, "text": "each"}, {"end": 161.84, "start": 161.64, "text": "one"}, {"end": 162.36, "start": 161.84, "text": "labeled"}, {"end": 162.4, "start": 162.36, "text": "with"}, {"end": 162.6, "start": 162.4, "text": "the"}, {"end": 163.04, "start": 162.6, "text": "numbers"}, {"end": 163.12, "start": 163.04, "text": "that"}, {"end": 163.28, "start": 163.12, "text": "they're"}, {"end": 163.76, "start": 163.28, "text": "supposed"}, {"end": 163.84, "start": 163.76, "text": "to"}, {"end": 164.44, "start": 163.84, "text": "be."}, {"end": 164.64, "start": 164.44, "text": "And"}, {"end": 164.84, "start": 164.64, "text": "as"}, {"end": 165.32, "start": 164.84, "text": "provocative"}, {"end": 165.6, "start": 165.32, "text": "as"}, {"end": 165.92, "start": 165.6, "text": "it"}, {"end": 166.12, "start": 165.92, "text": "is"}, {"end": 166.28, "start": 166.12, "text": "to"}, {"end": 166.72, "start": 166.28, "text": "describe"}, {"end": 166.72, "start": 166.72, "text": "a"}, {"end": 167.24, "start": 166.72, "text": "machine"}, {"end": 167.4, "start": 167.24, "text": "as"}, {"end": 167.72, "start": 167.4, "text": "learning,"}, {"end": 167.92000000000002, "start": 167.72, "text": "once"}, {"end": 168.28, "start": 167.92000000000002, "text": "you"}, {"end": 168.8, "start": 168.28, "text": "actually"}, {"end": 169.04, "start": 168.8, "text": "see"}, {"end": 169.28, "start": 169.04, "text": "how"}, {"end": 169.32, "start": 169.28, "text": "it"}, {"end": 169.76, "start": 169.32, "text": "works,"}, {"end": 169.76, "start": 169.76, "text": "it"}, {"end": 169.76, "start": 169.76, "text": "feels"}, {"end": 169.76, "start": 169.76, "text": "a"}, {"end": 169.8, "start": 169.76, "text": "lot"}, {"end": 170.28, "start": 169.8, "text": "less"}, {"end": 170.88, "start": 170.28, "text": "like"}, {"end": 171.16, "start": 170.88, "text": "some"}, {"end": 171.52, "start": 171.16, "text": "crazy"}, {"end": 172.0, "start": 171.52, "text": "sci-fi"}, {"end": 172.56, "start": 172.0, "text": "premise,"}, {"end": 172.6, "start": 172.56, "text": "and"}, {"end": 172.6, "start": 172.6, "text": "a"}, {"end": 172.6, "start": 172.6, "text": "lot"}, {"end": 173.24, "start": 172.6, "text": "more"}, {"end": 174.07999999999998, "start": 173.24, "text": "like,"}, {"end": 174.32, "start": 174.07999999999998, "text": "well,"}, {"end": 174.32, "start": 174.32, "text": "a"}, {"end": 174.96, "start": 174.32, "text": "calculus"}, {"end": 176.0, "start": 174.96, "text": "exercise."}, {"end": 176.07999999999998, "start": 176.0, "text": "I"}, {"end": 176.36, "start": 176.07999999999998, "text": "mean,"}, {"end": 176.88, "start": 176.36, "text": "basically"}, {"end": 176.88, "start": 176.88, "text": "it"}, {"end": 177.2, "start": 176.88, "text": "comes"}, {"end": 177.4, "start": 177.2, "text": "down"}, {"end": 177.48, "start": 177.4, "text": "to"}, {"end": 178.32, "start": 177.48, "text": "finding"}, {"end": 178.44, "start": 178.32, "text": "the"}, {"end": 178.8, "start": 178.44, "text": "minimum"}, {"end": 178.8, "start": 178.8, "text": "of"}, {"end": 178.84, "start": 178.8, "text": "a"}, {"end": 179.4, "start": 178.84, "text": "certain"}, {"end": 179.96, "start": 179.4, "text": "function."}], "text": " Fortunately for us, and what makes this such a common example to start with, is that the good people behind the MNIST database have put together a collection of tens of thousands of handwritten digit images, each one labeled with the numbers that they're supposed to be. And as provocative as it is to describe a machine as learning, once you actually see how it works, it feels a lot less like some crazy sci-fi premise, and a lot more like, well, a calculus exercise. I mean, basically it comes down to finding the minimum of a certain function."}, {"chunks": [{"end": 180.16, "start": 180.0, "text": "Remember,"}, {"end": 181.84, "start": 180.16, "text": "conceptually"}, {"end": 182.48, "start": 181.84, "text": "we're"}, {"end": 183.04, "start": 182.48, "text": "thinking"}, {"end": 183.36, "start": 183.04, "text": "of"}, {"end": 183.72, "start": 183.36, "text": "each"}, {"end": 184.32, "start": 183.72, "text": "neuron"}, {"end": 184.56, "start": 184.32, "text": "as"}, {"end": 184.8, "start": 184.56, "text": "being"}, {"end": 185.28, "start": 184.8, "text": "connected"}, {"end": 185.76, "start": 185.28, "text": "to"}, {"end": 185.96, "start": 185.76, "text": "all"}, {"end": 186.4, "start": 185.96, "text": "of"}, {"end": 186.44, "start": 186.4, "text": "the"}, {"end": 187.04, "start": 186.44, "text": "neurons"}, {"end": 187.04, "start": 187.04, "text": "in"}, {"end": 187.08, "start": 187.04, "text": "the"}, {"end": 187.56, "start": 187.08, "text": "previous"}, {"end": 187.8, "start": 187.56, "text": "layer,"}, {"end": 188.2, "start": 187.8, "text": "and"}, {"end": 188.44, "start": 188.2, "text": "the"}, {"end": 189.24, "start": 188.44, "text": "weights"}, {"end": 189.48, "start": 189.24, "text": "in"}, {"end": 189.48, "start": 189.48, "text": "the"}, {"end": 189.52, "start": 189.48, "text": "weighted"}, {"end": 189.8, "start": 189.52, "text": "sum"}, {"end": 190.28, "start": 189.8, "text": "defining"}, {"end": 190.56, "start": 190.28, "text": "its"}, {"end": 191.44, "start": 190.56, "text": "activation"}, {"end": 191.56, "start": 191.44, "text": "are"}, {"end": 192.0, "start": 191.56, "text": "kind"}, {"end": 192.16, "start": 192.0, "text": "of"}, {"end": 192.48, "start": 192.16, "text": "like"}, {"end": 192.52, "start": 192.48, "text": "the"}, {"end": 193.08, "start": 192.52, "text": "strengths"}, {"end": 193.16, "start": 193.08, "text": "of"}, {"end": 193.44, "start": 193.16, "text": "those"}, {"end": 194.52, "start": 193.44, "text": "connections."}, {"end": 194.72, "start": 194.52, "text": "The"}, {"end": 195.24, "start": 194.72, "text": "bias"}, {"end": 195.32, "start": 195.24, "text": "is"}, {"end": 195.68, "start": 195.32, "text": "some"}, {"end": 196.48, "start": 195.68, "text": "indication"}, {"end": 196.68, "start": 196.48, "text": "of"}, {"end": 197.04, "start": 196.68, "text": "whether"}, {"end": 197.04, "start": 197.04, "text": "that"}, {"end": 197.04, "start": 197.04, "text": "neuron"}, {"end": 197.4, "start": 197.04, "text": "tends"}, {"end": 197.52, "start": 197.4, "text": "to"}, {"end": 197.68, "start": 197.52, "text": "be"}, {"end": 197.88, "start": 197.68, "text": "active"}, {"end": 197.92000000000002, "start": 197.88, "text": "or"}, {"end": 198.24, "start": 197.92000000000002, "text": "inactive."}, {"end": 198.76, "start": 198.24, "text": "And"}, {"end": 199.12, "start": 198.76, "text": "to"}, {"end": 199.12, "start": 199.12, "text": "start"}, {"end": 199.56, "start": 199.12, "text": "things"}, {"end": 199.68, "start": 199.56, "text": "off,"}, {"end": 200.2, "start": 199.68, "text": "we're"}, {"end": 200.64, "start": 200.2, "text": "just"}, {"end": 200.8, "start": 200.64, "text": "going"}, {"end": 201.04, "start": 200.8, "text": "to"}, {"end": 201.84, "start": 201.04, "text": "initialize"}, {"end": 202.04, "start": 201.84, "text": "all"}, {"end": 202.04, "start": 202.04, "text": "of"}, {"end": 202.28, "start": 202.04, "text": "those"}, {"end": 202.64, "start": 202.28, "text": "weights"}, {"end": 202.68, "start": 202.64, "text": "and"}, {"end": 203.24, "start": 202.68, "text": "biases"}, {"end": 203.84, "start": 203.24, "text": "totally"}, {"end": 204.88, "start": 203.84, "text": "randomly."}, {"end": 205.24, "start": 204.88, "text": "Needless"}, {"end": 205.64, "start": 205.24, "text": "to"}, {"end": 205.68, "start": 205.64, "text": "say,"}, {"end": 205.92000000000002, "start": 205.68, "text": "this"}, {"end": 206.2, "start": 205.92000000000002, "text": "network"}, {"end": 206.32, "start": 206.2, "text": "is"}, {"end": 206.52, "start": 206.32, "text": "going"}, {"end": 206.52, "start": 206.52, "text": "to"}, {"end": 206.96, "start": 206.52, "text": "perform"}, {"end": 206.96, "start": 206.96, "text": "pretty"}, {"end": 207.16, "start": 206.96, "text": "horribly"}, {"end": 207.6, "start": 207.16, "text": "on"}, {"end": 207.88, "start": 207.6, "text": "a"}, {"end": 208.04, "start": 207.88, "text": "given"}, {"end": 208.36, "start": 208.04, "text": "training"}, {"end": 208.96, "start": 208.36, "text": "example,"}, {"end": 209.36, "start": 208.96, "text": "since"}, {"end": 209.48, "start": 209.36, "text": "it's"}, {"end": 209.68, "start": 209.48, "text": "just"}, {"end": 209.96, "start": 209.68, "text": "doing"}], "text": " Remember, conceptually we're thinking of each neuron as being connected to all of the neurons in the previous layer, and the weights in the weighted sum defining its activation are kind of like the strengths of those connections. The bias is some indication of whether that neuron tends to be active or inactive. And to start things off, we're just going to initialize all of those weights and biases totally randomly. Needless to say, this network is going to perform pretty horribly on a given training example, since it's just doing"}, {"chunks": [{"end": 210.32, "start": 210.0, "text": "something"}, {"end": 210.68, "start": 210.32, "text": "random."}, {"end": 211.16, "start": 210.68, "text": "For"}, {"end": 211.64, "start": 211.16, "text": "example,"}, {"end": 211.88, "start": 211.64, "text": "you"}, {"end": 212.0, "start": 211.88, "text": "feed"}, {"end": 212.32, "start": 212.0, "text": "in"}, {"end": 212.92, "start": 212.32, "text": "this"}, {"end": 213.16, "start": 212.92, "text": "image"}, {"end": 213.2, "start": 213.16, "text": "of"}, {"end": 213.36, "start": 213.2, "text": "a"}, {"end": 213.44, "start": 213.36, "text": "3,"}, {"end": 213.44, "start": 213.44, "text": "and"}, {"end": 213.56, "start": 213.44, "text": "the"}, {"end": 214.08, "start": 213.56, "text": "output"}, {"end": 215.0, "start": 214.08, "text": "layer"}, {"end": 215.6, "start": 215.0, "text": "just"}, {"end": 216.32, "start": 215.6, "text": "looks"}, {"end": 216.48, "start": 216.32, "text": "like"}, {"end": 216.52, "start": 216.48, "text": "a"}, {"end": 216.72, "start": 216.52, "text": "mess."}, {"end": 217.08, "start": 216.72, "text": "So"}, {"end": 217.12, "start": 217.08, "text": "what"}, {"end": 217.44, "start": 217.12, "text": "you"}, {"end": 217.48, "start": 217.44, "text": "do"}, {"end": 217.64, "start": 217.48, "text": "is"}, {"end": 217.76, "start": 217.64, "text": "you"}, {"end": 218.28, "start": 217.76, "text": "define"}, {"end": 218.28, "start": 218.28, "text": "a"}, {"end": 218.92, "start": 218.28, "text": "cost"}, {"end": 219.08, "start": 218.92, "text": "function,"}, {"end": 219.16, "start": 219.08, "text": "a"}, {"end": 219.2, "start": 219.16, "text": "way"}, {"end": 219.24, "start": 219.2, "text": "of"}, {"end": 219.96, "start": 219.24, "text": "telling"}, {"end": 220.36, "start": 219.96, "text": "the"}, {"end": 221.2, "start": 220.36, "text": "computer,"}, {"end": 221.88, "start": 221.2, "text": "no,"}, {"end": 222.24, "start": 221.88, "text": "bad"}, {"end": 223.08, "start": 222.24, "text": "computer,"}, {"end": 223.44, "start": 223.08, "text": "that"}, {"end": 223.88, "start": 223.44, "text": "output"}, {"end": 224.32, "start": 223.88, "text": "should"}, {"end": 224.72, "start": 224.32, "text": "have"}, {"end": 225.32, "start": 224.72, "text": "activations"}, {"end": 225.84, "start": 225.32, "text": "which"}, {"end": 226.12, "start": 225.84, "text": "are"}, {"end": 226.12, "start": 226.12, "text": "0"}, {"end": 226.28, "start": 226.12, "text": "for"}, {"end": 226.52, "start": 226.28, "text": "most"}, {"end": 227.04, "start": 226.52, "text": "neurons,"}, {"end": 227.24, "start": 227.04, "text": "but"}, {"end": 227.72, "start": 227.24, "text": "1"}, {"end": 227.88, "start": 227.72, "text": "for"}, {"end": 228.07999999999998, "start": 227.88, "text": "this"}, {"end": 228.8, "start": 228.07999999999998, "text": "neuron."}, {"end": 229.0, "start": 228.8, "text": "What"}, {"end": 229.24, "start": 229.0, "text": "you"}, {"end": 229.52, "start": 229.24, "text": "gave"}, {"end": 229.84, "start": 229.52, "text": "me"}, {"end": 230.04, "start": 229.84, "text": "is"}, {"end": 230.24, "start": 230.04, "text": "utter"}, {"end": 230.84, "start": 230.24, "text": "trash."}, {"end": 231.64, "start": 230.84, "text": "To"}, {"end": 231.92000000000002, "start": 231.64, "text": "say"}, {"end": 232.16, "start": 231.92000000000002, "text": "that"}, {"end": 232.16, "start": 232.16, "text": "a"}, {"end": 232.24, "start": 232.16, "text": "little"}, {"end": 232.84, "start": 232.24, "text": "more"}, {"end": 233.64, "start": 232.84, "text": "mathematically,"}, {"end": 233.84, "start": 233.64, "text": "what"}, {"end": 233.92000000000002, "start": 233.84, "text": "you"}, {"end": 233.96, "start": 233.92000000000002, "text": "do"}, {"end": 234.32, "start": 233.96, "text": "is"}, {"end": 234.32, "start": 234.32, "text": "add"}, {"end": 234.32, "start": 234.32, "text": "up"}, {"end": 234.88, "start": 234.32, "text": "the"}, {"end": 235.68, "start": 234.88, "text": "squares"}, {"end": 235.92000000000002, "start": 235.68, "text": "of"}, {"end": 235.96, "start": 235.92000000000002, "text": "the"}, {"end": 236.76, "start": 235.96, "text": "differences"}, {"end": 237.07999999999998, "start": 236.76, "text": "between"}, {"end": 237.36, "start": 237.07999999999998, "text": "each"}, {"end": 237.36, "start": 237.36, "text": "of"}, {"end": 237.64, "start": 237.36, "text": "those"}, {"end": 238.12, "start": 237.64, "text": "trash"}, {"end": 238.6, "start": 238.12, "text": "output"}, {"end": 239.28, "start": 238.6, "text": "activations"}, {"end": 239.48, "start": 239.28, "text": "and"}, {"end": 239.68, "start": 239.48, "text": "the"}, {"end": 239.96, "start": 239.68, "text": "value"}], "text": " something random. For example, you feed in this image of a 3, and the output layer just looks like a mess. So what you do is you define a cost function, a way of telling the computer, no, bad computer, that output should have activations which are 0 for most neurons, but 1 for this neuron. What you gave me is utter trash. To say that a little more mathematically, what you do is add up the squares of the differences between each of those trash output activations and the value"}, {"chunks": [{"end": 240.2, "start": 240.0, "text": "that"}, {"end": 240.4, "start": 240.2, "text": "you"}, {"end": 240.72, "start": 240.4, "text": "want"}, {"end": 241.16, "start": 240.72, "text": "them"}, {"end": 241.28, "start": 241.16, "text": "to"}, {"end": 241.28, "start": 241.28, "text": "have"}, {"end": 241.28, "start": 241.28, "text": "and"}, {"end": 241.28, "start": 241.28, "text": "this"}, {"end": 241.72, "start": 241.28, "text": "is"}, {"end": 241.8, "start": 241.72, "text": "what"}, {"end": 242.44, "start": 241.8, "text": "we'll"}, {"end": 242.8, "start": 242.44, "text": "call"}, {"end": 242.84, "start": 242.8, "text": "the"}, {"end": 243.32, "start": 242.84, "text": "cost"}, {"end": 243.36, "start": 243.32, "text": "of"}, {"end": 243.64, "start": 243.36, "text": "a"}, {"end": 244.6, "start": 243.64, "text": "single"}, {"end": 244.84, "start": 244.6, "text": "training"}, {"end": 245.92, "start": 244.84, "text": "example."}, {"end": 246.32, "start": 245.92, "text": "Notice"}, {"end": 246.48, "start": 246.32, "text": "this"}, {"end": 246.96, "start": 246.48, "text": "sum"}, {"end": 247.24, "start": 246.96, "text": "is"}, {"end": 247.96, "start": 247.24, "text": "small"}, {"end": 248.48, "start": 247.96, "text": "when"}, {"end": 248.48, "start": 248.48, "text": "the"}, {"end": 248.52, "start": 248.48, "text": "network"}, {"end": 249.16, "start": 248.52, "text": "confidently"}, {"end": 249.84, "start": 249.16, "text": "classifies"}, {"end": 250.12, "start": 249.84, "text": "the"}, {"end": 250.24, "start": 250.12, "text": "image"}, {"end": 252.12, "start": 250.24, "text": "correctly"}, {"end": 252.64, "start": 252.12, "text": "but"}, {"end": 252.88, "start": 252.64, "text": "it's"}, {"end": 253.44, "start": 252.88, "text": "large"}, {"end": 253.44, "start": 253.44, "text": "when"}, {"end": 253.52, "start": 253.44, "text": "the"}, {"end": 254.04, "start": 253.52, "text": "network"}, {"end": 254.36, "start": 254.04, "text": "seems"}, {"end": 254.52, "start": 254.36, "text": "like"}, {"end": 254.64, "start": 254.52, "text": "it"}, {"end": 254.92, "start": 254.64, "text": "doesn't"}, {"end": 255.12, "start": 254.92, "text": "really"}, {"end": 255.44, "start": 255.12, "text": "know"}, {"end": 255.44, "start": 255.44, "text": "what"}, {"end": 255.56, "start": 255.44, "text": "it's"}, {"end": 256.44, "start": 255.56, "text": "doing."}, {"end": 257.12, "start": 256.44, "text": "So"}, {"end": 258.28, "start": 257.12, "text": "then"}, {"end": 259.12, "start": 258.28, "text": "what"}, {"end": 259.44, "start": 259.12, "text": "you"}, {"end": 259.52, "start": 259.44, "text": "do"}, {"end": 259.64, "start": 259.52, "text": "is"}, {"end": 260.28, "start": 259.64, "text": "consider"}, {"end": 260.6, "start": 260.28, "text": "the"}, {"end": 260.76, "start": 260.6, "text": "average"}, {"end": 261.28, "start": 260.76, "text": "cost"}, {"end": 261.48, "start": 261.28, "text": "over"}, {"end": 261.88, "start": 261.48, "text": "all"}, {"end": 262.44, "start": 261.88, "text": "of"}, {"end": 262.6, "start": 262.44, "text": "the"}, {"end": 263.2, "start": 262.6, "text": "tens"}, {"end": 263.24, "start": 263.2, "text": "of"}, {"end": 263.84, "start": 263.24, "text": "thousands"}, {"end": 263.88, "start": 263.84, "text": "of"}, {"end": 264.28, "start": 263.88, "text": "training"}, {"end": 264.56, "start": 264.28, "text": "examples"}, {"end": 264.6, "start": 264.56, "text": "at"}, {"end": 264.92, "start": 264.6, "text": "your"}, {"end": 267.4, "start": 264.92, "text": "disposal."}, {"end": 267.84, "start": 267.4, "text": "This"}, {"end": 267.96, "start": 267.84, "text": "average"}, {"end": 268.36, "start": 267.96, "text": "cost"}, {"end": 268.56, "start": 268.36, "text": "is"}, {"end": 268.84, "start": 268.56, "text": "our"}, {"end": 269.32, "start": 268.84, "text": "measure"}, {"end": 269.52, "start": 269.32, "text": "for"}, {"end": 269.6, "start": 269.52, "text": "how"}, {"end": 269.96, "start": 269.6, "text": "loud"}], "text": " that you want them to have and this is what we'll call the cost of a single training example. Notice this sum is small when the network confidently classifies the image correctly but it's large when the network seems like it doesn't really know what it's doing. So then what you do is consider the average cost over all of the tens of thousands of training examples at your disposal. This average cost is our measure for how loud"}, {"chunks": [{"end": 270.24, "start": 270.0, "text": "the"}, {"end": 270.6, "start": 270.24, "text": "network"}, {"end": 270.92, "start": 270.6, "text": "is,"}, {"end": 271.08, "start": 270.92, "text": "and"}, {"end": 271.32, "start": 271.08, "text": "how"}, {"end": 271.6, "start": 271.32, "text": "bad"}, {"end": 271.76, "start": 271.6, "text": "the"}, {"end": 272.32, "start": 271.76, "text": "computer"}, {"end": 272.8, "start": 272.32, "text": "should"}, {"end": 273.08, "start": 272.8, "text": "feel."}, {"end": 273.4, "start": 273.08, "text": "And"}, {"end": 273.64, "start": 273.4, "text": "that's"}, {"end": 273.64, "start": 273.64, "text": "a"}, {"end": 274.24, "start": 273.64, "text": "complicated"}, {"end": 275.0, "start": 274.24, "text": "thing."}, {"end": 275.24, "start": 275.0, "text": "Remember"}, {"end": 275.52, "start": 275.24, "text": "how"}, {"end": 275.72, "start": 275.52, "text": "the"}, {"end": 276.08, "start": 275.72, "text": "network"}, {"end": 276.48, "start": 276.08, "text": "itself"}, {"end": 276.68, "start": 276.48, "text": "was"}, {"end": 277.04, "start": 276.68, "text": "basically"}, {"end": 277.28, "start": 277.04, "text": "a"}, {"end": 277.72, "start": 277.28, "text": "function,"}, {"end": 278.04, "start": 277.72, "text": "one"}, {"end": 278.52, "start": 278.04, "text": "that"}, {"end": 278.96, "start": 278.52, "text": "takes"}, {"end": 279.08, "start": 278.96, "text": "in"}, {"end": 280.04, "start": 279.08, "text": "784"}, {"end": 280.64, "start": 280.04, "text": "numbers"}, {"end": 280.96, "start": 280.64, "text": "as"}, {"end": 281.36, "start": 280.96, "text": "inputs,"}, {"end": 281.48, "start": 281.36, "text": "the"}, {"end": 281.76, "start": 281.48, "text": "pixel"}, {"end": 282.24, "start": 281.76, "text": "values,"}, {"end": 282.28, "start": 282.24, "text": "and"}, {"end": 282.32, "start": 282.28, "text": "spits"}, {"end": 282.4, "start": 282.32, "text": "out"}, {"end": 282.6, "start": 282.4, "text": "10"}, {"end": 283.12, "start": 282.6, "text": "numbers"}, {"end": 283.36, "start": 283.12, "text": "as"}, {"end": 283.68, "start": 283.36, "text": "its"}, {"end": 284.32, "start": 283.68, "text": "output,"}, {"end": 284.96, "start": 284.32, "text": "and"}, {"end": 285.2, "start": 284.96, "text": "in"}, {"end": 285.48, "start": 285.2, "text": "a"}, {"end": 286.0, "start": 285.48, "text": "sense"}, {"end": 286.16, "start": 286.0, "text": "it's"}, {"end": 287.12, "start": 286.16, "text": "parameterized"}, {"end": 287.4, "start": 287.12, "text": "by"}, {"end": 287.72, "start": 287.4, "text": "all"}, {"end": 287.92, "start": 287.72, "text": "these"}, {"end": 288.28, "start": 287.92, "text": "weights"}, {"end": 288.28, "start": 288.28, "text": "and"}, {"end": 289.24, "start": 288.28, "text": "biases?"}, {"end": 289.28, "start": 289.24, "text": "Well"}, {"end": 289.72, "start": 289.28, "text": "the"}, {"end": 290.08, "start": 289.72, "text": "cost"}, {"end": 290.56, "start": 290.08, "text": "function"}, {"end": 291.0, "start": 290.56, "text": "is"}, {"end": 291.2, "start": 291.0, "text": "a"}, {"end": 291.2, "start": 291.2, "text": "layer"}, {"end": 291.2, "start": 291.2, "text": "of"}, {"end": 292.12, "start": 291.2, "text": "complexity"}, {"end": 292.48, "start": 292.12, "text": "on"}, {"end": 292.8, "start": 292.48, "text": "top"}, {"end": 292.8, "start": 292.8, "text": "of"}, {"end": 292.8, "start": 292.8, "text": "that."}, {"end": 292.8, "start": 292.8, "text": "It"}, {"end": 293.68, "start": 292.8, "text": "takes"}, {"end": 293.84, "start": 293.68, "text": "as"}, {"end": 294.04, "start": 293.84, "text": "its"}, {"end": 294.8, "start": 294.04, "text": "input"}, {"end": 295.04, "start": 294.8, "text": "those"}, {"end": 295.44, "start": 295.04, "text": "13,000"}, {"end": 295.88, "start": 295.44, "text": "or"}, {"end": 296.12, "start": 295.88, "text": "so"}, {"end": 296.44, "start": 296.12, "text": "weights"}, {"end": 296.52, "start": 296.44, "text": "and"}, {"end": 297.16, "start": 296.52, "text": "biases,"}, {"end": 297.36, "start": 297.16, "text": "and"}, {"end": 297.48, "start": 297.36, "text": "it"}, {"end": 297.88, "start": 297.48, "text": "spits"}, {"end": 298.28, "start": 297.88, "text": "out"}, {"end": 298.4, "start": 298.28, "text": "a"}, {"end": 298.88, "start": 298.4, "text": "single"}, {"end": 299.32, "start": 298.88, "text": "number"}, {"end": 299.84, "start": 299.32, "text": "describing"}, {"end": 299.96, "start": 299.84, "text": "how"}], "text": " the network is, and how bad the computer should feel. And that's a complicated thing. Remember how the network itself was basically a function, one that takes in 784 numbers as inputs, the pixel values, and spits out 10 numbers as its output, and in a sense it's parameterized by all these weights and biases? Well the cost function is a layer of complexity on top of that. It takes as its input those 13,000 or so weights and biases, and it spits out a single number describing how"}, {"chunks": [{"end": 300.12, "start": 300.0, "text": "bad"}, {"end": 300.56, "start": 300.12, "text": "those"}, {"end": 300.76, "start": 300.56, "text": "weights"}, {"end": 301.08, "start": 300.76, "text": "and"}, {"end": 301.6, "start": 301.08, "text": "biases"}, {"end": 301.72, "start": 301.6, "text": "are,"}, {"end": 302.12, "start": 301.72, "text": "and"}, {"end": 302.64, "start": 302.12, "text": "the"}, {"end": 303.12, "start": 302.64, "text": "way"}, {"end": 303.28, "start": 303.12, "text": "it's"}, {"end": 303.88, "start": 303.28, "text": "defined"}, {"end": 303.88, "start": 303.88, "text": "depends"}, {"end": 303.88, "start": 303.88, "text": "on"}, {"end": 304.04, "start": 303.88, "text": "the"}, {"end": 305.12, "start": 304.04, "text": "network's"}, {"end": 305.72, "start": 305.12, "text": "behavior"}, {"end": 306.12, "start": 305.72, "text": "over"}, {"end": 306.48, "start": 306.12, "text": "all"}, {"end": 306.56, "start": 306.48, "text": "the"}, {"end": 307.08, "start": 306.56, "text": "tens"}, {"end": 307.08, "start": 307.08, "text": "of"}, {"end": 307.72, "start": 307.08, "text": "thousands"}, {"end": 307.76, "start": 307.72, "text": "of"}, {"end": 308.16, "start": 307.76, "text": "pieces"}, {"end": 308.16, "start": 308.16, "text": "of"}, {"end": 308.56, "start": 308.16, "text": "training"}, {"end": 308.88, "start": 308.56, "text": "data."}, {"end": 308.88, "start": 308.88, "text": "That's"}, {"end": 308.88, "start": 308.88, "text": "a"}, {"end": 308.88, "start": 308.88, "text": "lot"}, {"end": 309.04, "start": 308.88, "text": "to"}, {"end": 309.52, "start": 309.04, "text": "think"}, {"end": 310.36, "start": 309.52, "text": "about."}, {"end": 310.56, "start": 310.36, "text": "But"}, {"end": 310.64, "start": 310.56, "text": "just"}, {"end": 311.48, "start": 310.64, "text": "telling"}, {"end": 311.56, "start": 311.48, "text": "the"}, {"end": 312.88, "start": 311.56, "text": "computer"}, {"end": 313.04, "start": 312.88, "text": "what"}, {"end": 313.12, "start": 313.04, "text": "a"}, {"end": 313.96, "start": 313.12, "text": "crappy"}, {"end": 314.2, "start": 313.96, "text": "job"}, {"end": 314.4, "start": 314.2, "text": "it's"}, {"end": 314.88, "start": 314.4, "text": "doing"}, {"end": 315.16, "start": 314.88, "text": "isn't"}, {"end": 315.64, "start": 315.16, "text": "very"}, {"end": 315.8, "start": 315.64, "text": "helpful."}, {"end": 315.8, "start": 315.8, "text": "You"}, {"end": 315.84, "start": 315.8, "text": "want"}, {"end": 315.84, "start": 315.84, "text": "to"}, {"end": 316.04, "start": 315.84, "text": "tell"}, {"end": 316.24, "start": 316.04, "text": "it"}, {"end": 316.6, "start": 316.24, "text": "how"}, {"end": 317.12, "start": 316.6, "text": "to"}, {"end": 317.72, "start": 317.12, "text": "change"}, {"end": 318.0, "start": 317.72, "text": "those"}, {"end": 318.32, "start": 318.0, "text": "weights"}, {"end": 318.32, "start": 318.32, "text": "and"}, {"end": 319.04, "start": 318.32, "text": "biases"}, {"end": 319.24, "start": 319.04, "text": "so"}, {"end": 319.44, "start": 319.24, "text": "that"}, {"end": 319.64, "start": 319.44, "text": "it"}, {"end": 319.64, "start": 319.64, "text": "gets"}, {"end": 320.04, "start": 319.64, "text": "better."}, {"end": 320.8, "start": 320.04, "text": "To"}, {"end": 321.24, "start": 320.8, "text": "make"}, {"end": 321.52, "start": 321.24, "text": "it"}, {"end": 321.92, "start": 321.52, "text": "easier,"}, {"end": 321.96, "start": 321.92, "text": "rather"}, {"end": 322.44, "start": 321.96, "text": "than"}, {"end": 322.84, "start": 322.44, "text": "struggling"}, {"end": 323.04, "start": 322.84, "text": "to"}, {"end": 323.4, "start": 323.04, "text": "imagine"}, {"end": 323.56, "start": 323.4, "text": "a"}, {"end": 324.08, "start": 323.56, "text": "function"}, {"end": 324.08, "start": 324.08, "text": "with"}, {"end": 324.88, "start": 324.08, "text": "13,000"}, {"end": 325.36, "start": 324.88, "text": "inputs,"}, {"end": 325.84, "start": 325.36, "text": "just"}, {"end": 326.24, "start": 325.84, "text": "imagine"}, {"end": 326.32, "start": 326.24, "text": "a"}, {"end": 326.64, "start": 326.32, "text": "simple"}, {"end": 327.16, "start": 326.64, "text": "function"}, {"end": 327.24, "start": 327.16, "text": "that"}, {"end": 327.64, "start": 327.24, "text": "has"}, {"end": 327.76, "start": 327.64, "text": "one"}, {"end": 328.16, "start": 327.76, "text": "number"}, {"end": 328.52, "start": 328.16, "text": "as"}, {"end": 328.64, "start": 328.52, "text": "an"}, {"end": 328.68, "start": 328.64, "text": "input"}, {"end": 328.72, "start": 328.68, "text": "and"}, {"end": 328.96, "start": 328.72, "text": "one"}, {"end": 329.12, "start": 328.96, "text": "number"}, {"end": 329.16, "start": 329.12, "text": "as"}, {"end": 329.96, "start": 329.16, "text": "an"}], "text": " bad those weights and biases are, and the way it's defined depends on the network's behavior over all the tens of thousands of pieces of training data. That's a lot to think about. But just telling the computer what a crappy job it's doing isn't very helpful. You want to tell it how to change those weights and biases so that it gets better. To make it easier, rather than struggling to imagine a function with 13,000 inputs, just imagine a simple function that has one number as an input and one number as an"}, {"chunks": [{"end": 330.24, "start": 330.0, "text": "output."}, {"end": 330.44, "start": 330.24, "text": "How"}, {"end": 330.44, "start": 330.44, "text": "do"}, {"end": 331.28, "start": 330.44, "text": "you"}, {"end": 331.84, "start": 331.28, "text": "find"}, {"end": 332.28, "start": 331.84, "text": "an"}, {"end": 332.44, "start": 332.28, "text": "input"}, {"end": 333.32, "start": 332.44, "text": "that"}, {"end": 334.0, "start": 333.32, "text": "minimizes"}, {"end": 334.12, "start": 334.0, "text": "the"}, {"end": 334.64, "start": 334.12, "text": "value"}, {"end": 334.64, "start": 334.64, "text": "of"}, {"end": 334.84, "start": 334.64, "text": "this"}, {"end": 336.36, "start": 334.84, "text": "function?"}, {"end": 336.76, "start": 336.36, "text": "Calculus"}, {"end": 337.16, "start": 336.76, "text": "students"}, {"end": 337.44, "start": 337.16, "text": "will"}, {"end": 337.64, "start": 337.44, "text": "know"}, {"end": 337.68, "start": 337.64, "text": "that"}, {"end": 337.76, "start": 337.68, "text": "you"}, {"end": 337.88, "start": 337.76, "text": "can"}, {"end": 338.4, "start": 337.88, "text": "sometimes"}, {"end": 338.76, "start": 338.4, "text": "figure"}, {"end": 339.12, "start": 338.76, "text": "out"}, {"end": 339.28, "start": 339.12, "text": "that"}, {"end": 339.4, "start": 339.28, "text": "minimum"}, {"end": 340.64, "start": 339.4, "text": "explicitly,"}, {"end": 340.92, "start": 340.64, "text": "but"}, {"end": 341.04, "start": 340.92, "text": "that's"}, {"end": 341.48, "start": 341.04, "text": "not"}, {"end": 341.8, "start": 341.48, "text": "always"}, {"end": 342.36, "start": 341.8, "text": "feasible"}, {"end": 342.64, "start": 342.36, "text": "for"}, {"end": 343.0, "start": 342.64, "text": "really"}, {"end": 343.44, "start": 343.0, "text": "complicated"}, {"end": 344.36, "start": 343.44, "text": "functions,"}, {"end": 344.8, "start": 344.36, "text": "certainly"}, {"end": 345.04, "start": 344.8, "text": "not"}, {"end": 345.4, "start": 345.04, "text": "in"}, {"end": 345.44, "start": 345.4, "text": "the"}, {"end": 346.52, "start": 345.44, "text": "13,000"}, {"end": 346.68, "start": 346.52, "text": "input"}, {"end": 347.12, "start": 346.68, "text": "version"}, {"end": 347.12, "start": 347.12, "text": "of"}, {"end": 347.44, "start": 347.12, "text": "this"}, {"end": 347.96, "start": 347.44, "text": "situation"}, {"end": 348.2, "start": 347.96, "text": "for"}, {"end": 348.68, "start": 348.2, "text": "our"}, {"end": 349.12, "start": 348.68, "text": "crazy"}, {"end": 349.72, "start": 349.12, "text": "complicated"}, {"end": 349.76, "start": 349.72, "text": "neural"}, {"end": 350.32, "start": 349.76, "text": "network"}, {"end": 350.64, "start": 350.32, "text": "cost"}, {"end": 351.0, "start": 350.64, "text": "function."}, {"end": 351.64, "start": 351.0, "text": "A"}, {"end": 351.96, "start": 351.64, "text": "more"}, {"end": 352.4, "start": 351.96, "text": "flexible"}, {"end": 353.0, "start": 352.4, "text": "tactic"}, {"end": 353.16, "start": 353.0, "text": "is"}, {"end": 353.32, "start": 353.16, "text": "to"}, {"end": 353.72, "start": 353.32, "text": "start"}, {"end": 353.96, "start": 353.72, "text": "at"}, {"end": 354.56, "start": 353.96, "text": "any"}, {"end": 355.12, "start": 354.56, "text": "old"}, {"end": 355.2, "start": 355.12, "text": "input"}, {"end": 355.4, "start": 355.2, "text": "and"}, {"end": 355.64, "start": 355.4, "text": "figure"}, {"end": 355.92, "start": 355.64, "text": "out"}, {"end": 356.28, "start": 355.92, "text": "which"}, {"end": 356.84, "start": 356.28, "text": "direction"}, {"end": 357.04, "start": 356.84, "text": "you"}, {"end": 357.24, "start": 357.04, "text": "should"}, {"end": 357.52, "start": 357.24, "text": "step"}, {"end": 357.84, "start": 357.52, "text": "to"}, {"end": 358.16, "start": 357.84, "text": "make"}, {"end": 358.2, "start": 358.16, "text": "that"}, {"end": 358.8, "start": 358.2, "text": "output"}, {"end": 359.96, "start": 358.8, "text": "lower."}], "text": " output. How do you find an input that minimizes the value of this function? Calculus students will know that you can sometimes figure out that minimum explicitly, but that's not always feasible for really complicated functions, certainly not in the 13,000 input version of this situation for our crazy complicated neural network cost function. A more flexible tactic is to start at any old input and figure out which direction you should step to make that output lower."}, {"chunks": [{"end": 360.88, "start": 360.0, "text": "Specifically,"}, {"end": 361.08, "start": 360.88, "text": "if"}, {"end": 361.12, "start": 361.08, "text": "you"}, {"end": 361.28, "start": 361.12, "text": "can"}, {"end": 361.48, "start": 361.28, "text": "figure"}, {"end": 361.68, "start": 361.48, "text": "out"}, {"end": 361.8, "start": 361.68, "text": "the"}, {"end": 362.32, "start": 361.8, "text": "slope"}, {"end": 362.48, "start": 362.32, "text": "of"}, {"end": 362.76, "start": 362.48, "text": "the"}, {"end": 363.52, "start": 362.76, "text": "function"}, {"end": 363.96, "start": 363.52, "text": "where"}, {"end": 364.28, "start": 363.96, "text": "you"}, {"end": 364.4, "start": 364.28, "text": "are,"}, {"end": 364.8, "start": 364.4, "text": "then"}, {"end": 365.0, "start": 364.8, "text": "shift"}, {"end": 365.0, "start": 365.0, "text": "to"}, {"end": 365.28, "start": 365.0, "text": "the"}, {"end": 365.56, "start": 365.28, "text": "left"}, {"end": 365.6, "start": 365.56, "text": "if"}, {"end": 365.92, "start": 365.6, "text": "that"}, {"end": 366.28, "start": 365.92, "text": "slope"}, {"end": 366.28, "start": 366.28, "text": "is"}, {"end": 367.2, "start": 366.28, "text": "positive,"}, {"end": 367.4, "start": 367.2, "text": "and"}, {"end": 367.6, "start": 367.4, "text": "shift"}, {"end": 367.6, "start": 367.6, "text": "the"}, {"end": 367.6, "start": 367.6, "text": "input"}, {"end": 367.88, "start": 367.6, "text": "to"}, {"end": 368.28, "start": 367.88, "text": "the"}, {"end": 368.72, "start": 368.28, "text": "right"}, {"end": 368.88, "start": 368.72, "text": "if"}, {"end": 369.4, "start": 368.88, "text": "that"}, {"end": 369.68, "start": 369.4, "text": "slope"}, {"end": 369.72, "start": 369.68, "text": "is"}, {"end": 369.88, "start": 369.72, "text": "negative."}, {"end": 370.8, "start": 369.88, "text": "If"}, {"end": 371.16, "start": 370.8, "text": "you"}, {"end": 372.6, "start": 371.16, "text": "do"}, {"end": 372.88, "start": 372.6, "text": "this"}, {"end": 373.28, "start": 372.88, "text": "repeatedly,"}, {"end": 373.52, "start": 373.28, "text": "at"}, {"end": 373.64, "start": 373.52, "text": "each"}, {"end": 373.92, "start": 373.64, "text": "point"}, {"end": 374.56, "start": 373.92, "text": "checking"}, {"end": 374.56, "start": 374.56, "text": "the"}, {"end": 375.16, "start": 374.56, "text": "new"}, {"end": 375.48, "start": 375.16, "text": "slope"}, {"end": 375.64, "start": 375.48, "text": "and"}, {"end": 375.96, "start": 375.64, "text": "taking"}, {"end": 376.12, "start": 375.96, "text": "the"}, {"end": 376.6, "start": 376.12, "text": "appropriate"}, {"end": 377.16, "start": 376.6, "text": "step,"}, {"end": 377.6, "start": 377.16, "text": "you're"}, {"end": 377.6, "start": 377.6, "text": "going"}, {"end": 377.6, "start": 377.6, "text": "to"}, {"end": 378.08, "start": 377.6, "text": "approach"}, {"end": 378.36, "start": 378.08, "text": "some"}, {"end": 378.72, "start": 378.36, "text": "local"}, {"end": 379.16, "start": 378.72, "text": "minimum"}, {"end": 379.28, "start": 379.16, "text": "of"}, {"end": 379.36, "start": 379.28, "text": "the"}, {"end": 379.84, "start": 379.36, "text": "function."}, {"end": 379.84, "start": 379.84, "text": "And"}, {"end": 380.6, "start": 379.84, "text": "the"}, {"end": 380.96, "start": 380.6, "text": "image"}, {"end": 381.04, "start": 380.96, "text": "you"}, {"end": 381.44, "start": 381.04, "text": "might"}, {"end": 381.68, "start": 381.44, "text": "have"}, {"end": 381.68, "start": 381.68, "text": "in"}, {"end": 381.72, "start": 381.68, "text": "mind"}, {"end": 382.04, "start": 381.72, "text": "here"}, {"end": 382.44, "start": 382.04, "text": "is"}, {"end": 382.44, "start": 382.44, "text": "a"}, {"end": 382.84, "start": 382.44, "text": "ball"}, {"end": 383.16, "start": 382.84, "text": "rolling"}, {"end": 383.48, "start": 383.16, "text": "down"}, {"end": 383.48, "start": 383.48, "text": "a"}, {"end": 383.8, "start": 383.48, "text": "hill."}, {"end": 384.52, "start": 383.8, "text": "And"}, {"end": 385.0, "start": 384.52, "text": "notice,"}, {"end": 385.48, "start": 385.0, "text": "even"}, {"end": 385.68, "start": 385.48, "text": "for"}, {"end": 385.88, "start": 385.68, "text": "this"}, {"end": 386.08, "start": 385.88, "text": "really"}, {"end": 386.68, "start": 386.08, "text": "simplified"}, {"end": 387.12, "start": 386.68, "text": "single"}, {"end": 387.24, "start": 387.12, "text": "input"}, {"end": 387.88, "start": 387.24, "text": "function,"}, {"end": 388.16, "start": 387.88, "text": "there"}, {"end": 388.56, "start": 388.16, "text": "are"}, {"end": 388.92, "start": 388.56, "text": "many"}, {"end": 389.48, "start": 388.92, "text": "possible"}, {"end": 389.96, "start": 389.48, "text": "valleys"}], "text": " Specifically, if you can figure out the slope of the function where you are, then shift to the left if that slope is positive, and shift the input to the right if that slope is negative. If you do this repeatedly, at each point checking the new slope and taking the appropriate step, you're going to approach some local minimum of the function. And the image you might have in mind here is a ball rolling down a hill. And notice, even for this really simplified single input function, there are many possible valleys"}, {"chunks": [{"end": 390.24, "start": 390.0, "text": "you"}, {"end": 390.44, "start": 390.24, "text": "might"}, {"end": 390.72, "start": 390.44, "text": "land"}, {"end": 390.88, "start": 390.72, "text": "in,"}, {"end": 391.84, "start": 390.88, "text": "depending"}, {"end": 391.92, "start": 391.84, "text": "on"}, {"end": 392.12, "start": 391.92, "text": "which"}, {"end": 392.44, "start": 392.12, "text": "random"}, {"end": 392.64, "start": 392.44, "text": "input"}, {"end": 392.96, "start": 392.64, "text": "you"}, {"end": 393.6, "start": 392.96, "text": "start"}, {"end": 393.76, "start": 393.6, "text": "at."}, {"end": 393.8, "start": 393.76, "text": "And"}, {"end": 394.12, "start": 393.8, "text": "there's"}, {"end": 394.36, "start": 394.12, "text": "no"}, {"end": 394.88, "start": 394.36, "text": "guarantee"}, {"end": 394.88, "start": 394.88, "text": "that"}, {"end": 395.16, "start": 394.88, "text": "the"}, {"end": 396.0, "start": 395.16, "text": "local"}, {"end": 396.44, "start": 396.0, "text": "minimum"}, {"end": 396.44, "start": 396.44, "text": "you"}, {"end": 396.44, "start": 396.44, "text": "land"}, {"end": 396.44, "start": 396.44, "text": "in"}, {"end": 396.44, "start": 396.44, "text": "is"}, {"end": 396.44, "start": 396.44, "text": "going"}, {"end": 396.44, "start": 396.44, "text": "to"}, {"end": 396.6, "start": 396.44, "text": "be"}, {"end": 397.12, "start": 396.6, "text": "the"}, {"end": 397.68, "start": 397.12, "text": "smallest"}, {"end": 398.16, "start": 397.68, "text": "possible"}, {"end": 398.56, "start": 398.16, "text": "value"}, {"end": 398.56, "start": 398.56, "text": "of"}, {"end": 398.68, "start": 398.56, "text": "the"}, {"end": 399.0, "start": 398.68, "text": "cost"}, {"end": 400.0, "start": 399.0, "text": "function."}, {"end": 400.4, "start": 400.0, "text": "That's"}, {"end": 400.88, "start": 400.4, "text": "going"}, {"end": 400.88, "start": 400.88, "text": "to"}, {"end": 400.88, "start": 400.88, "text": "carry"}, {"end": 400.92, "start": 400.88, "text": "over"}, {"end": 400.96, "start": 400.92, "text": "to"}, {"end": 401.16, "start": 400.96, "text": "our"}, {"end": 401.4, "start": 401.16, "text": "neural"}, {"end": 401.76, "start": 401.4, "text": "network"}, {"end": 402.04, "start": 401.76, "text": "case"}, {"end": 402.2, "start": 402.04, "text": "as"}, {"end": 402.44, "start": 402.2, "text": "well."}, {"end": 402.44, "start": 402.44, "text": "And"}, {"end": 403.08, "start": 402.44, "text": "I"}, {"end": 403.44, "start": 403.08, "text": "also"}, {"end": 403.64, "start": 403.44, "text": "want"}, {"end": 403.64, "start": 403.64, "text": "you"}, {"end": 403.68, "start": 403.64, "text": "to"}, {"end": 404.16, "start": 403.68, "text": "notice"}, {"end": 404.44, "start": 404.16, "text": "how"}, {"end": 404.52, "start": 404.44, "text": "if"}, {"end": 404.52, "start": 404.52, "text": "you"}, {"end": 404.72, "start": 404.52, "text": "make"}, {"end": 405.16, "start": 404.72, "text": "your"}, {"end": 405.44, "start": 405.16, "text": "step"}, {"end": 406.0, "start": 405.44, "text": "sizes"}, {"end": 406.8, "start": 406.0, "text": "proportional"}, {"end": 406.92, "start": 406.8, "text": "to"}, {"end": 406.96, "start": 406.92, "text": "the"}, {"end": 407.4, "start": 406.96, "text": "slope,"}, {"end": 407.44, "start": 407.4, "text": "then"}, {"end": 407.48, "start": 407.44, "text": "when"}, {"end": 408.4, "start": 407.48, "text": "the"}, {"end": 408.68, "start": 408.4, "text": "slope"}, {"end": 408.88, "start": 408.68, "text": "is"}, {"end": 409.36, "start": 408.88, "text": "flattening"}, {"end": 409.56, "start": 409.36, "text": "out"}, {"end": 409.92, "start": 409.56, "text": "towards"}, {"end": 409.96, "start": 409.92, "text": "the"}, {"end": 410.24, "start": 409.96, "text": "minimum,"}, {"end": 410.64, "start": 410.24, "text": "your"}, {"end": 411.36, "start": 410.64, "text": "steps"}, {"end": 411.6, "start": 411.36, "text": "get"}, {"end": 411.92, "start": 411.6, "text": "smaller"}, {"end": 412.04, "start": 411.92, "text": "and"}, {"end": 412.28, "start": 412.04, "text": "smaller,"}, {"end": 412.4, "start": 412.28, "text": "and"}, {"end": 412.6, "start": 412.4, "text": "that"}, {"end": 412.64, "start": 412.6, "text": "kind"}, {"end": 413.0, "start": 412.64, "text": "of"}, {"end": 413.52, "start": 413.0, "text": "helps"}, {"end": 413.56, "start": 413.52, "text": "you"}, {"end": 413.84, "start": 413.56, "text": "from"}, {"end": 415.88, "start": 413.84, "text": "overshooting."}, {"end": 416.32, "start": 415.88, "text": "Bumping"}, {"end": 416.32, "start": 416.32, "text": "up"}, {"end": 416.32, "start": 416.32, "text": "the"}, {"end": 417.0, "start": 416.32, "text": "complexity"}, {"end": 417.12, "start": 417.0, "text": "a"}, {"end": 417.6, "start": 417.12, "text": "bit,"}, {"end": 417.72, "start": 417.6, "text": "imagine"}, {"end": 417.76, "start": 417.72, "text": "instead"}, {"end": 418.16, "start": 417.76, "text": "a"}, {"end": 418.8, "start": 418.16, "text": "function"}, {"end": 419.08, "start": 418.8, "text": "with"}, {"end": 419.56, "start": 419.08, "text": "two"}, {"end": 419.96, "start": 419.56, "text": "inputs,"}], "text": " you might land in, depending on which random input you start at. And there's no guarantee that the local minimum you land in is going to be the smallest possible value of the cost function. That's going to carry over to our neural network case as well. And I also want you to notice how if you make your step sizes proportional to the slope, then when the slope is flattening out towards the minimum, your steps get smaller and smaller, and that kind of helps you from overshooting. Bumping up the complexity a bit, imagine instead a function with two inputs,"}, {"chunks": [{"end": 420.24, "start": 420.0, "text": "and"}, {"end": 420.48, "start": 420.24, "text": "one"}, {"end": 420.68, "start": 420.48, "text": "output."}, {"end": 421.08, "start": 420.68, "text": "You"}, {"end": 421.2, "start": 421.08, "text": "might"}, {"end": 421.36, "start": 421.2, "text": "think"}, {"end": 421.76, "start": 421.36, "text": "of"}, {"end": 422.08, "start": 421.76, "text": "the"}, {"end": 422.48, "start": 422.08, "text": "input"}, {"end": 423.04, "start": 422.48, "text": "space"}, {"end": 423.04, "start": 423.04, "text": "as"}, {"end": 423.08, "start": 423.04, "text": "the"}, {"end": 423.52, "start": 423.08, "text": "xy"}, {"end": 423.92, "start": 423.52, "text": "plane,"}, {"end": 424.04, "start": 423.92, "text": "and"}, {"end": 424.48, "start": 424.04, "text": "the"}, {"end": 424.8, "start": 424.48, "text": "cost"}, {"end": 425.28, "start": 424.8, "text": "function"}, {"end": 425.48, "start": 425.28, "text": "as"}, {"end": 425.8, "start": 425.48, "text": "being"}, {"end": 426.28, "start": 425.8, "text": "graphed"}, {"end": 426.36, "start": 426.28, "text": "as"}, {"end": 426.68, "start": 426.36, "text": "a"}, {"end": 427.36, "start": 426.68, "text": "surface"}, {"end": 427.84, "start": 427.36, "text": "above"}, {"end": 428.28, "start": 427.84, "text": "it."}, {"end": 428.48, "start": 428.28, "text": "Now"}, {"end": 428.52, "start": 428.48, "text": "instead"}, {"end": 428.68, "start": 428.52, "text": "of"}, {"end": 429.04, "start": 428.68, "text": "asking"}, {"end": 429.36, "start": 429.04, "text": "about"}, {"end": 429.8, "start": 429.36, "text": "the"}, {"end": 430.2, "start": 429.8, "text": "slope"}, {"end": 430.32, "start": 430.2, "text": "of"}, {"end": 430.44, "start": 430.32, "text": "the"}, {"end": 430.88, "start": 430.44, "text": "function,"}, {"end": 430.92, "start": 430.88, "text": "you"}, {"end": 431.32, "start": 430.92, "text": "have"}, {"end": 431.76, "start": 431.32, "text": "to"}, {"end": 432.28, "start": 431.76, "text": "ask"}, {"end": 432.6, "start": 432.28, "text": "which"}, {"end": 433.24, "start": 432.6, "text": "direction"}, {"end": 433.44, "start": 433.24, "text": "should"}, {"end": 433.52, "start": 433.44, "text": "you"}, {"end": 433.92, "start": 433.52, "text": "step"}, {"end": 434.08, "start": 433.92, "text": "in"}, {"end": 434.64, "start": 434.08, "text": "this"}, {"end": 434.8, "start": 434.64, "text": "input"}, {"end": 435.68, "start": 434.8, "text": "space"}, {"end": 435.8, "start": 435.68, "text": "so"}, {"end": 435.92, "start": 435.8, "text": "as"}, {"end": 436.2, "start": 435.92, "text": "to"}, {"end": 436.72, "start": 436.2, "text": "decrease"}, {"end": 436.72, "start": 436.72, "text": "the"}, {"end": 436.96, "start": 436.72, "text": "output"}, {"end": 437.12, "start": 436.96, "text": "of"}, {"end": 437.48, "start": 437.12, "text": "the"}, {"end": 438.2, "start": 437.48, "text": "function"}, {"end": 438.44, "start": 438.2, "text": "most"}, {"end": 438.88, "start": 438.44, "text": "quickly."}, {"end": 439.08, "start": 438.88, "text": "In"}, {"end": 439.72, "start": 439.08, "text": "other"}, {"end": 440.32, "start": 439.72, "text": "words,"}, {"end": 440.6, "start": 440.32, "text": "what's"}, {"end": 440.76, "start": 440.6, "text": "the"}, {"end": 441.2, "start": 440.76, "text": "downhill"}, {"end": 441.88, "start": 441.2, "text": "direction?"}, {"end": 442.04, "start": 441.88, "text": "And"}, {"end": 442.76, "start": 442.04, "text": "again,"}, {"end": 443.12, "start": 442.76, "text": "it's"}, {"end": 443.76, "start": 443.12, "text": "helpful"}, {"end": 443.76, "start": 443.76, "text": "to"}, {"end": 443.76, "start": 443.76, "text": "think"}, {"end": 443.76, "start": 443.76, "text": "of"}, {"end": 443.76, "start": 443.76, "text": "a"}, {"end": 444.32, "start": 443.76, "text": "ball"}, {"end": 444.72, "start": 444.32, "text": "rolling"}, {"end": 444.96, "start": 444.72, "text": "down"}, {"end": 445.16, "start": 444.96, "text": "that"}, {"end": 445.52, "start": 445.16, "text": "hill."}, {"end": 446.28, "start": 445.52, "text": "Those"}, {"end": 446.44, "start": 446.28, "text": "of"}, {"end": 446.48, "start": 446.44, "text": "you"}, {"end": 447.08, "start": 446.48, "text": "familiar"}, {"end": 447.72, "start": 447.08, "text": "with"}, {"end": 448.48, "start": 447.72, "text": "multivariable"}, {"end": 449.24, "start": 448.48, "text": "calculus"}, {"end": 449.88, "start": 449.24, "text": "will"}, {"end": 449.92, "start": 449.88, "text": "know"}, {"end": 449.96, "start": 449.92, "text": "that"}], "text": " and one output. You might think of the input space as the xy plane, and the cost function as being graphed as a surface above it. Now instead of asking about the slope of the function, you have to ask which direction should you step in this input space so as to decrease the output of the function most quickly. In other words, what's the downhill direction? And again, it's helpful to think of a ball rolling down that hill. Those of you familiar with multivariable calculus will know that"}, {"chunks": [{"end": 450.2, "start": 450.0, "text": "The"}, {"end": 451.0, "start": 450.2, "text": "gradient"}, {"end": 451.16, "start": 451.0, "text": "of"}, {"end": 451.16, "start": 451.16, "text": "a"}, {"end": 452.04, "start": 451.16, "text": "function"}, {"end": 452.52, "start": 452.04, "text": "gives"}, {"end": 452.64, "start": 452.52, "text": "you"}, {"end": 452.88, "start": 452.64, "text": "the"}, {"end": 453.32, "start": 452.88, "text": "direction"}, {"end": 453.56, "start": 453.32, "text": "of"}, {"end": 454.0, "start": 453.56, "text": "steepest"}, {"end": 454.88, "start": 454.0, "text": "ascent."}, {"end": 455.36, "start": 454.88, "text": "Basically,"}, {"end": 455.68, "start": 455.36, "text": "which"}, {"end": 456.0, "start": 455.68, "text": "direction"}, {"end": 456.16, "start": 456.0, "text": "should"}, {"end": 456.24, "start": 456.16, "text": "you"}, {"end": 456.76, "start": 456.24, "text": "step"}, {"end": 457.0, "start": 456.76, "text": "to"}, {"end": 457.52, "start": 457.0, "text": "increase"}, {"end": 457.6, "start": 457.52, "text": "the"}, {"end": 458.08, "start": 457.6, "text": "function"}, {"end": 458.24, "start": 458.08, "text": "most"}, {"end": 459.48, "start": 458.24, "text": "quickly."}, {"end": 459.84, "start": 459.48, "text": "Naturally"}, {"end": 460.12, "start": 459.84, "text": "enough,"}, {"end": 460.84, "start": 460.12, "text": "taking"}, {"end": 461.04, "start": 460.84, "text": "the"}, {"end": 461.48, "start": 461.04, "text": "negative"}, {"end": 461.8, "start": 461.48, "text": "of"}, {"end": 461.84, "start": 461.8, "text": "that"}, {"end": 462.36, "start": 461.84, "text": "gradient"}, {"end": 462.4, "start": 462.36, "text": "gives"}, {"end": 462.48, "start": 462.4, "text": "you"}, {"end": 462.92, "start": 462.48, "text": "the"}, {"end": 463.24, "start": 462.92, "text": "direction"}, {"end": 463.32, "start": 463.24, "text": "to"}, {"end": 463.72, "start": 463.32, "text": "step"}, {"end": 464.0, "start": 463.72, "text": "that"}, {"end": 464.64, "start": 464.0, "text": "decreases"}, {"end": 464.72, "start": 464.64, "text": "the"}, {"end": 465.32, "start": 464.72, "text": "function"}, {"end": 465.52, "start": 465.32, "text": "most"}, {"end": 465.56, "start": 465.52, "text": "quickly."}, {"end": 467.2, "start": 465.56, "text": "And"}, {"end": 467.4, "start": 467.2, "text": "even"}, {"end": 467.68, "start": 467.4, "text": "more"}, {"end": 467.96, "start": 467.68, "text": "than"}, {"end": 467.96, "start": 467.96, "text": "that,"}, {"end": 468.12, "start": 467.96, "text": "the"}, {"end": 468.64, "start": 468.12, "text": "length"}, {"end": 468.8, "start": 468.64, "text": "of"}, {"end": 469.04, "start": 468.8, "text": "this"}, {"end": 469.4, "start": 469.04, "text": "gradient"}, {"end": 469.84, "start": 469.4, "text": "vector"}, {"end": 469.84, "start": 469.84, "text": "is"}, {"end": 470.52, "start": 469.84, "text": "actually"}, {"end": 470.84, "start": 470.52, "text": "an"}, {"end": 471.16, "start": 470.84, "text": "indication"}, {"end": 471.4, "start": 471.16, "text": "for"}, {"end": 471.6, "start": 471.4, "text": "just"}, {"end": 471.92, "start": 471.6, "text": "how"}, {"end": 472.28, "start": 471.92, "text": "steep"}, {"end": 472.28, "start": 472.28, "text": "that"}, {"end": 472.8, "start": 472.28, "text": "steepest"}, {"end": 473.2, "start": 472.8, "text": "slope"}, {"end": 473.76, "start": 473.2, "text": "is."}, {"end": 474.0, "start": 473.76, "text": "Now"}, {"end": 474.48, "start": 474.0, "text": "if"}, {"end": 474.84, "start": 474.48, "text": "you're"}, {"end": 475.52, "start": 474.84, "text": "unfamiliar"}, {"end": 475.52, "start": 475.52, "text": "with"}, {"end": 476.08, "start": 475.52, "text": "multivariable"}, {"end": 476.84, "start": 476.08, "text": "calculus"}, {"end": 477.52, "start": 476.84, "text": "and"}, {"end": 477.52, "start": 477.52, "text": "you"}, {"end": 477.52, "start": 477.52, "text": "want"}, {"end": 477.52, "start": 477.52, "text": "to"}, {"end": 477.52, "start": 477.52, "text": "learn"}, {"end": 477.52, "start": 477.52, "text": "more,"}, {"end": 477.68, "start": 477.52, "text": "check"}, {"end": 477.68, "start": 477.68, "text": "out"}, {"end": 477.68, "start": 477.68, "text": "some"}, {"end": 477.68, "start": 477.68, "text": "of"}, {"end": 478.32, "start": 477.68, "text": "the"}, {"end": 478.6, "start": 478.32, "text": "work"}, {"end": 478.64, "start": 478.6, "text": "that"}, {"end": 478.8, "start": 478.64, "text": "I"}, {"end": 478.96, "start": 478.8, "text": "did"}, {"end": 479.08, "start": 478.96, "text": "for"}, {"end": 479.32, "start": 479.08, "text": "Khan"}, {"end": 479.68, "start": 479.32, "text": "Academy"}, {"end": 479.76, "start": 479.68, "text": "on"}, {"end": 479.84, "start": 479.76, "text": "the"}, {"end": 479.96, "start": 479.84, "text": "top."}], "text": " The gradient of a function gives you the direction of steepest ascent. Basically, which direction should you step to increase the function most quickly. Naturally enough, taking the negative of that gradient gives you the direction to step that decreases the function most quickly. And even more than that, the length of this gradient vector is actually an indication for just how steep that steepest slope is. Now if you're unfamiliar with multivariable calculus and you want to learn more, check out some of the work that I did for Khan Academy on the top."}, {"chunks": [{"end": 481.16, "start": 480.0, "text": "Honestly"}, {"end": 481.48, "start": 481.16, "text": "though,"}, {"end": 481.92, "start": 481.48, "text": "all"}, {"end": 481.92, "start": 481.92, "text": "that"}, {"end": 481.96, "start": 481.92, "text": "matters"}, {"end": 482.4, "start": 481.96, "text": "for"}, {"end": 482.52, "start": 482.4, "text": "you"}, {"end": 483.12, "start": 482.52, "text": "and"}, {"end": 483.4, "start": 483.12, "text": "me"}, {"end": 483.64, "start": 483.4, "text": "right"}, {"end": 484.32, "start": 483.64, "text": "now"}, {"end": 484.44, "start": 484.32, "text": "is"}, {"end": 484.76, "start": 484.44, "text": "that"}, {"end": 484.88, "start": 484.76, "text": "in"}, {"end": 485.56, "start": 484.88, "text": "principle"}, {"end": 485.8, "start": 485.56, "text": "there"}, {"end": 486.2, "start": 485.8, "text": "exists"}, {"end": 486.36, "start": 486.2, "text": "a"}, {"end": 486.68, "start": 486.36, "text": "way"}, {"end": 486.88, "start": 486.68, "text": "to"}, {"end": 487.2, "start": 486.88, "text": "compute"}, {"end": 487.48, "start": 487.2, "text": "this"}, {"end": 487.6, "start": 487.48, "text": "vector."}, {"end": 487.76, "start": 487.6, "text": "This"}, {"end": 488.12, "start": 487.76, "text": "vector"}, {"end": 488.16, "start": 488.12, "text": "that"}, {"end": 488.4, "start": 488.16, "text": "tells"}, {"end": 488.84, "start": 488.4, "text": "you"}, {"end": 489.2, "start": 488.84, "text": "what"}, {"end": 489.6, "start": 489.2, "text": "the"}, {"end": 490.04, "start": 489.6, "text": "downhill"}, {"end": 490.48, "start": 490.04, "text": "direction"}, {"end": 490.76, "start": 490.48, "text": "is"}, {"end": 490.92, "start": 490.76, "text": "and"}, {"end": 491.2, "start": 490.92, "text": "how"}, {"end": 491.48, "start": 491.2, "text": "steep"}, {"end": 491.56, "start": 491.48, "text": "it"}, {"end": 491.92, "start": 491.56, "text": "is."}, {"end": 492.32, "start": 491.92, "text": "You'll"}, {"end": 492.56, "start": 492.32, "text": "be"}, {"end": 493.08, "start": 492.56, "text": "okay"}, {"end": 493.28, "start": 493.08, "text": "if"}, {"end": 493.92, "start": 493.28, "text": "that's"}, {"end": 494.28, "start": 493.92, "text": "all"}, {"end": 494.4, "start": 494.28, "text": "you"}, {"end": 495.12, "start": 494.4, "text": "know"}, {"end": 495.68, "start": 495.12, "text": "and"}, {"end": 495.96, "start": 495.68, "text": "you're"}, {"end": 496.24, "start": 495.96, "text": "not"}, {"end": 496.44, "start": 496.24, "text": "rock"}, {"end": 496.92, "start": 496.44, "text": "solid"}, {"end": 497.08, "start": 496.92, "text": "on"}, {"end": 497.24, "start": 497.08, "text": "the"}, {"end": 497.8, "start": 497.24, "text": "details."}, {"end": 498.24, "start": 497.8, "text": "Because"}, {"end": 498.24, "start": 498.24, "text": "if"}, {"end": 498.28, "start": 498.24, "text": "you"}, {"end": 498.28, "start": 498.28, "text": "can"}, {"end": 498.28, "start": 498.28, "text": "get"}, {"end": 498.32, "start": 498.28, "text": "that,"}, {"end": 498.56, "start": 498.32, "text": "the"}, {"end": 499.04, "start": 498.56, "text": "algorithm"}, {"end": 499.32, "start": 499.04, "text": "for"}, {"end": 499.84, "start": 499.32, "text": "minimizing"}, {"end": 499.92, "start": 499.84, "text": "the"}, {"end": 500.84, "start": 499.92, "text": "function"}, {"end": 500.92, "start": 500.84, "text": "is"}, {"end": 501.2, "start": 500.92, "text": "to"}, {"end": 501.44, "start": 501.2, "text": "compute"}, {"end": 501.64, "start": 501.44, "text": "this"}, {"end": 502.0, "start": 501.64, "text": "gradient"}, {"end": 502.56, "start": 502.0, "text": "direction,"}, {"end": 502.72, "start": 502.56, "text": "then"}, {"end": 503.0, "start": 502.72, "text": "take"}, {"end": 503.36, "start": 503.0, "text": "a"}, {"end": 503.68, "start": 503.36, "text": "small"}, {"end": 503.92, "start": 503.68, "text": "step"}, {"end": 504.6, "start": 503.92, "text": "downhill"}, {"end": 505.0, "start": 504.6, "text": "and"}, {"end": 505.36, "start": 505.0, "text": "just"}, {"end": 505.68, "start": 505.36, "text": "repeat"}, {"end": 505.76, "start": 505.68, "text": "that"}, {"end": 506.52, "start": 505.76, "text": "over"}, {"end": 507.16, "start": 506.52, "text": "and"}, {"end": 508.16, "start": 507.16, "text": "over."}, {"end": 508.32, "start": 508.16, "text": "It's"}, {"end": 508.36, "start": 508.32, "text": "the"}, {"end": 508.6, "start": 508.36, "text": "same"}, {"end": 508.88, "start": 508.6, "text": "basic"}, {"end": 508.88, "start": 508.88, "text": "idea"}, {"end": 508.92, "start": 508.88, "text": "for"}, {"end": 509.08, "start": 508.92, "text": "a"}, {"end": 509.56, "start": 509.08, "text": "function"}, {"end": 509.96, "start": 509.56, "text": "that"}], "text": " Honestly though, all that matters for you and me right now is that in principle there exists a way to compute this vector. This vector that tells you what the downhill direction is and how steep it is. You'll be okay if that's all you know and you're not rock solid on the details. Because if you can get that, the algorithm for minimizing the function is to compute this gradient direction, then take a small step downhill and just repeat that over and over. It's the same basic idea for a function that"}, {"chunks": [{"end": 510.36, "start": 510.0, "text": "has"}, {"end": 511.32, "start": 510.36, "text": "13,000"}, {"end": 511.72, "start": 511.32, "text": "inputs"}, {"end": 511.92, "start": 511.72, "text": "instead"}, {"end": 512.08, "start": 511.92, "text": "of"}, {"end": 512.44, "start": 512.08, "text": "2"}, {"end": 513.36, "start": 512.44, "text": "inputs."}, {"end": 513.84, "start": 513.36, "text": "Imagine"}, {"end": 514.64, "start": 513.84, "text": "organizing"}, {"end": 515.04, "start": 514.64, "text": "all"}, {"end": 515.88, "start": 515.04, "text": "13,000"}, {"end": 516.28, "start": 515.88, "text": "weights"}, {"end": 516.32, "start": 516.28, "text": "and"}, {"end": 517.12, "start": 516.32, "text": "biases"}, {"end": 517.24, "start": 517.12, "text": "of"}, {"end": 517.24, "start": 517.24, "text": "our"}, {"end": 517.8, "start": 517.24, "text": "network"}, {"end": 518.08, "start": 517.8, "text": "into"}, {"end": 518.16, "start": 518.08, "text": "a"}, {"end": 518.64, "start": 518.16, "text": "giant"}, {"end": 519.04, "start": 518.64, "text": "column"}, {"end": 519.36, "start": 519.04, "text": "vector."}, {"end": 520.12, "start": 519.36, "text": "The"}, {"end": 520.84, "start": 520.12, "text": "negative"}, {"end": 521.56, "start": 520.84, "text": "gradient"}, {"end": 521.68, "start": 521.56, "text": "of"}, {"end": 521.8, "start": 521.68, "text": "the"}, {"end": 522.12, "start": 521.8, "text": "cost"}, {"end": 522.72, "start": 522.12, "text": "function"}, {"end": 523.08, "start": 522.72, "text": "is"}, {"end": 523.6, "start": 523.08, "text": "just"}, {"end": 523.76, "start": 523.6, "text": "a"}, {"end": 524.28, "start": 523.76, "text": "vector."}, {"end": 524.64, "start": 524.28, "text": "It's"}, {"end": 524.88, "start": 524.64, "text": "some"}, {"end": 525.84, "start": 524.88, "text": "direction"}, {"end": 526.36, "start": 525.84, "text": "inside"}, {"end": 527.0, "start": 526.36, "text": "this"}, {"end": 527.56, "start": 527.0, "text": "insanely"}, {"end": 527.8, "start": 527.56, "text": "huge"}, {"end": 527.8, "start": 527.8, "text": "input"}, {"end": 528.16, "start": 527.8, "text": "space"}, {"end": 528.32, "start": 528.16, "text": "that"}, {"end": 528.84, "start": 528.32, "text": "tells"}, {"end": 529.12, "start": 528.84, "text": "you"}, {"end": 529.52, "start": 529.12, "text": "which"}, {"end": 530.08, "start": 529.52, "text": "nudges"}, {"end": 530.08, "start": 530.08, "text": "to"}, {"end": 530.08, "start": 530.08, "text": "all"}, {"end": 530.12, "start": 530.08, "text": "of"}, {"end": 530.72, "start": 530.12, "text": "those"}, {"end": 531.48, "start": 530.72, "text": "numbers"}, {"end": 531.48, "start": 531.48, "text": "is"}, {"end": 531.52, "start": 531.48, "text": "going"}, {"end": 531.88, "start": 531.52, "text": "to"}, {"end": 532.2, "start": 531.88, "text": "cause"}, {"end": 532.24, "start": 532.2, "text": "the"}, {"end": 532.68, "start": 532.24, "text": "most"}, {"end": 533.24, "start": 532.68, "text": "rapid"}, {"end": 533.92, "start": 533.24, "text": "decrease"}, {"end": 533.96, "start": 533.92, "text": "to"}, {"end": 534.12, "start": 533.96, "text": "the"}, {"end": 534.44, "start": 534.12, "text": "cost"}, {"end": 534.76, "start": 534.44, "text": "function."}, {"end": 535.12, "start": 534.76, "text": "And"}, {"end": 535.72, "start": 535.12, "text": "of"}, {"end": 536.36, "start": 535.72, "text": "course,"}, {"end": 536.52, "start": 536.36, "text": "with"}, {"end": 536.68, "start": 536.52, "text": "our"}, {"end": 537.16, "start": 536.68, "text": "specially"}, {"end": 537.48, "start": 537.16, "text": "designed"}, {"end": 537.8, "start": 537.48, "text": "cost"}, {"end": 538.24, "start": 537.8, "text": "function,"}, {"end": 538.88, "start": 538.24, "text": "changing"}, {"end": 539.04, "start": 538.88, "text": "the"}, {"end": 539.52, "start": 539.04, "text": "weights"}, {"end": 539.52, "start": 539.52, "text": "and"}, {"end": 539.96, "start": 539.52, "text": "biases"}], "text": " has 13,000 inputs instead of 2 inputs. Imagine organizing all 13,000 weights and biases of our network into a giant column vector. The negative gradient of the cost function is just a vector. It's some direction inside this insanely huge input space that tells you which nudges to all of those numbers is going to cause the most rapid decrease to the cost function. And of course, with our specially designed cost function, changing the weights and biases"}, {"chunks": [{"end": 540.4, "start": 540.0, "text": "To"}, {"end": 541.08, "start": 540.4, "text": "decrease"}, {"end": 541.6, "start": 541.08, "text": "it"}, {"end": 541.92, "start": 541.6, "text": "means"}, {"end": 542.4, "start": 541.92, "text": "making"}, {"end": 542.48, "start": 542.4, "text": "the"}, {"end": 542.72, "start": 542.48, "text": "output"}, {"end": 542.72, "start": 542.72, "text": "of"}, {"end": 543.04, "start": 542.72, "text": "the"}, {"end": 543.64, "start": 543.04, "text": "network"}, {"end": 543.96, "start": 543.64, "text": "on"}, {"end": 544.04, "start": 543.96, "text": "each"}, {"end": 544.32, "start": 544.04, "text": "piece"}, {"end": 544.36, "start": 544.32, "text": "of"}, {"end": 544.72, "start": 544.36, "text": "training"}, {"end": 545.36, "start": 544.72, "text": "data"}, {"end": 545.48, "start": 545.36, "text": "look"}, {"end": 545.48, "start": 545.48, "text": "less"}, {"end": 545.48, "start": 545.48, "text": "like"}, {"end": 545.8, "start": 545.48, "text": "a"}, {"end": 546.68, "start": 545.8, "text": "random"}, {"end": 546.8, "start": 546.68, "text": "array"}, {"end": 547.16, "start": 546.8, "text": "of"}, {"end": 547.4, "start": 547.16, "text": "10"}, {"end": 547.92, "start": 547.4, "text": "values"}, {"end": 548.04, "start": 547.92, "text": "and"}, {"end": 548.16, "start": 548.04, "text": "more"}, {"end": 548.24, "start": 548.16, "text": "like"}, {"end": 548.84, "start": 548.24, "text": "an"}, {"end": 549.24, "start": 548.84, "text": "actual"}, {"end": 550.0, "start": 549.24, "text": "decision"}, {"end": 550.28, "start": 550.0, "text": "that"}, {"end": 550.36, "start": 550.28, "text": "we"}, {"end": 550.48, "start": 550.36, "text": "want"}, {"end": 550.64, "start": 550.48, "text": "it"}, {"end": 550.64, "start": 550.64, "text": "to"}, {"end": 551.4, "start": 550.64, "text": "make."}, {"end": 551.64, "start": 551.4, "text": "It's"}, {"end": 551.84, "start": 551.64, "text": "important"}, {"end": 552.16, "start": 551.84, "text": "to"}, {"end": 552.44, "start": 552.16, "text": "remember,"}, {"end": 552.64, "start": 552.44, "text": "this"}, {"end": 553.04, "start": 552.64, "text": "cost"}, {"end": 553.36, "start": 553.04, "text": "function"}, {"end": 553.96, "start": 553.36, "text": "involves"}, {"end": 554.24, "start": 553.96, "text": "an"}, {"end": 554.64, "start": 554.24, "text": "average"}, {"end": 555.04, "start": 554.64, "text": "over"}, {"end": 555.76, "start": 555.04, "text": "all"}, {"end": 555.96, "start": 555.76, "text": "of"}, {"end": 555.96, "start": 555.96, "text": "the"}, {"end": 556.12, "start": 555.96, "text": "training"}, {"end": 556.76, "start": 556.12, "text": "data,"}, {"end": 556.92, "start": 556.76, "text": "so"}, {"end": 557.0, "start": 556.92, "text": "if"}, {"end": 557.08, "start": 557.0, "text": "you"}, {"end": 557.72, "start": 557.08, "text": "minimize"}, {"end": 557.72, "start": 557.72, "text": "it,"}, {"end": 557.96, "start": 557.72, "text": "it"}, {"end": 558.36, "start": 557.96, "text": "means"}, {"end": 558.56, "start": 558.36, "text": "it's"}, {"end": 558.6, "start": 558.56, "text": "a"}, {"end": 558.8, "start": 558.6, "text": "better"}, {"end": 559.4, "start": 558.8, "text": "performance"}, {"end": 559.76, "start": 559.4, "text": "on"}, {"end": 560.08, "start": 559.76, "text": "all"}, {"end": 560.16, "start": 560.08, "text": "of"}, {"end": 560.64, "start": 560.16, "text": "those"}, {"end": 561.0, "start": 560.64, "text": "samples."}, {"end": 563.8, "start": 561.0, "text": "The"}, {"end": 564.2, "start": 563.8, "text": "algorithm"}, {"end": 564.84, "start": 564.2, "text": "for"}, {"end": 565.16, "start": 564.84, "text": "computing"}, {"end": 565.36, "start": 565.16, "text": "this"}, {"end": 566.08, "start": 565.36, "text": "gradient"}, {"end": 566.64, "start": 566.08, "text": "efficiently,"}, {"end": 566.92, "start": 566.64, "text": "which"}, {"end": 567.24, "start": 566.92, "text": "is"}, {"end": 567.64, "start": 567.24, "text": "effectively"}, {"end": 567.76, "start": 567.64, "text": "the"}, {"end": 567.76, "start": 567.76, "text": "heart"}, {"end": 568.12, "start": 567.76, "text": "of"}, {"end": 568.64, "start": 568.12, "text": "how"}, {"end": 568.92, "start": 568.64, "text": "a"}, {"end": 569.16, "start": 568.92, "text": "neural"}, {"end": 569.52, "start": 569.16, "text": "network"}, {"end": 569.96, "start": 569.52, "text": "learns,"}], "text": " To decrease it means making the output of the network on each piece of training data look less like a random array of 10 values and more like an actual decision that we want it to make. It's important to remember, this cost function involves an average over all of the training data, so if you minimize it, it means it's a better performance on all of those samples. The algorithm for computing this gradient efficiently, which is effectively the heart of how a neural network learns,"}, {"chunks": [{"end": 571.2, "start": 570.0, "text": "Backpropagation"}, {"end": 571.96, "start": 571.2, "text": "is"}, {"end": 572.24, "start": 571.96, "text": "what"}, {"end": 572.68, "start": 572.24, "text": "I'll"}, {"end": 572.92, "start": 572.68, "text": "talk"}, {"end": 573.24, "start": 572.92, "text": "about"}, {"end": 573.52, "start": 573.24, "text": "next"}, {"end": 574.48, "start": 573.52, "text": "video."}, {"end": 574.88, "start": 574.48, "text": "There,"}, {"end": 575.16, "start": 574.88, "text": "I"}, {"end": 575.56, "start": 575.16, "text": "want"}, {"end": 575.56, "start": 575.56, "text": "to"}, {"end": 575.56, "start": 575.56, "text": "take"}, {"end": 575.64, "start": 575.56, "text": "the"}, {"end": 576.0, "start": 575.64, "text": "time"}, {"end": 576.28, "start": 576.0, "text": "to"}, {"end": 576.6, "start": 576.28, "text": "walk"}, {"end": 577.04, "start": 576.6, "text": "through"}, {"end": 577.36, "start": 577.04, "text": "what"}, {"end": 577.92, "start": 577.36, "text": "exactly"}, {"end": 578.44, "start": 577.92, "text": "happens"}, {"end": 578.6, "start": 578.44, "text": "to"}, {"end": 579.12, "start": 578.6, "text": "each"}, {"end": 579.24, "start": 579.12, "text": "weight"}, {"end": 579.56, "start": 579.24, "text": "and"}, {"end": 580.16, "start": 579.56, "text": "bias"}, {"end": 580.4, "start": 580.16, "text": "for"}, {"end": 580.6, "start": 580.4, "text": "a"}, {"end": 580.6, "start": 580.6, "text": "given"}, {"end": 580.92, "start": 580.6, "text": "piece"}, {"end": 580.92, "start": 580.92, "text": "of"}, {"end": 581.28, "start": 580.92, "text": "training"}, {"end": 581.4, "start": 581.28, "text": "data,"}, {"end": 581.64, "start": 581.4, "text": "trying"}, {"end": 581.68, "start": 581.64, "text": "to"}, {"end": 582.12, "start": 581.68, "text": "give"}, {"end": 582.32, "start": 582.12, "text": "an"}, {"end": 582.92, "start": 582.32, "text": "intuitive"}, {"end": 583.32, "start": 582.92, "text": "feel"}, {"end": 583.6, "start": 583.32, "text": "for"}, {"end": 584.0, "start": 583.6, "text": "what's"}, {"end": 584.44, "start": 584.0, "text": "happening"}, {"end": 585.04, "start": 584.44, "text": "beyond"}, {"end": 585.04, "start": 585.04, "text": "the"}, {"end": 585.44, "start": 585.04, "text": "pile"}, {"end": 585.48, "start": 585.44, "text": "of"}, {"end": 585.84, "start": 585.48, "text": "relevant"}, {"end": 586.44, "start": 585.84, "text": "calculus"}, {"end": 586.44, "start": 586.44, "text": "and"}, {"end": 587.84, "start": 586.44, "text": "formulas."}, {"end": 588.0, "start": 587.84, "text": "Right"}, {"end": 588.36, "start": 588.0, "text": "here,"}, {"end": 588.52, "start": 588.36, "text": "right"}, {"end": 588.84, "start": 588.52, "text": "now,"}, {"end": 588.92, "start": 588.84, "text": "the"}, {"end": 589.4, "start": 588.92, "text": "main"}, {"end": 589.72, "start": 589.4, "text": "thing"}, {"end": 589.96, "start": 589.72, "text": "I"}, {"end": 590.36, "start": 589.96, "text": "want"}, {"end": 590.36, "start": 590.36, "text": "you"}, {"end": 590.48, "start": 590.36, "text": "to"}, {"end": 590.72, "start": 590.48, "text": "know,"}, {"end": 591.28, "start": 590.72, "text": "independent"}, {"end": 591.4, "start": 591.28, "text": "of"}, {"end": 591.8, "start": 591.4, "text": "implementation"}, {"end": 591.88, "start": 591.8, "text": "details,"}, {"end": 591.92, "start": 591.88, "text": "is"}, {"end": 592.0, "start": 591.92, "text": "that"}, {"end": 592.44, "start": 592.0, "text": "what"}, {"end": 592.6, "start": 592.44, "text": "we"}, {"end": 592.8, "start": 592.6, "text": "mean"}, {"end": 593.48, "start": 592.8, "text": "when"}, {"end": 594.16, "start": 593.48, "text": "we"}, {"end": 594.48, "start": 594.16, "text": "talk"}, {"end": 594.84, "start": 594.48, "text": "about"}, {"end": 594.92, "start": 594.84, "text": "a"}, {"end": 595.16, "start": 594.92, "text": "network"}, {"end": 595.68, "start": 595.16, "text": "learning"}, {"end": 595.68, "start": 595.68, "text": "is"}, {"end": 596.08, "start": 595.68, "text": "that"}, {"end": 596.52, "start": 596.08, "text": "it's"}, {"end": 596.92, "start": 596.52, "text": "just"}, {"end": 597.56, "start": 596.92, "text": "minimizing"}, {"end": 597.56, "start": 597.56, "text": "a"}, {"end": 598.04, "start": 597.56, "text": "cost"}, {"end": 598.72, "start": 598.04, "text": "function."}, {"end": 599.16, "start": 598.72, "text": "And"}, {"end": 599.6, "start": 599.16, "text": "notice,"}, {"end": 599.96, "start": 599.6, "text": "one"}], "text": " Backpropagation is what I'll talk about next video. There, I want to take the time to walk through what exactly happens to each weight and bias for a given piece of training data, trying to give an intuitive feel for what's happening beyond the pile of relevant calculus and formulas. Right here, right now, the main thing I want you to know, independent of implementation details, is that what we mean when we talk about a network learning is that it's just minimizing a cost function. And notice, one"}, {"chunks": [{"end": 600.0, "start": 600.0, "text": "The"}, {"end": 600.48, "start": 600.0, "text": "consequence"}, {"end": 600.68, "start": 600.48, "text": "of"}, {"end": 600.76, "start": 600.68, "text": "that"}, {"end": 601.12, "start": 600.76, "text": "is"}, {"end": 601.2, "start": 601.12, "text": "that"}, {"end": 601.36, "start": 601.2, "text": "it's"}, {"end": 601.76, "start": 601.36, "text": "important"}, {"end": 601.84, "start": 601.76, "text": "for"}, {"end": 602.04, "start": 601.84, "text": "this"}, {"end": 602.28, "start": 602.04, "text": "cost"}, {"end": 602.36, "start": 602.28, "text": "function"}, {"end": 602.36, "start": 602.36, "text": "to"}, {"end": 602.76, "start": 602.36, "text": "have"}, {"end": 603.12, "start": 602.76, "text": "a"}, {"end": 603.44, "start": 603.12, "text": "nice"}, {"end": 603.44, "start": 603.44, "text": "smooth"}, {"end": 603.48, "start": 603.44, "text": "output"}, {"end": 603.96, "start": 603.48, "text": "so"}, {"end": 604.28, "start": 603.96, "text": "that"}, {"end": 604.28, "start": 604.28, "text": "we"}, {"end": 604.76, "start": 604.28, "text": "can"}, {"end": 605.48, "start": 604.76, "text": "find"}, {"end": 605.64, "start": 605.48, "text": "a"}, {"end": 606.44, "start": 605.64, "text": "local"}, {"end": 606.44, "start": 606.44, "text": "minimum"}, {"end": 606.64, "start": 606.44, "text": "by"}, {"end": 606.76, "start": 606.64, "text": "taking"}, {"end": 607.2, "start": 606.76, "text": "little"}, {"end": 607.6, "start": 607.2, "text": "steps"}, {"end": 608.88, "start": 607.6, "text": "downhill."}, {"end": 609.24, "start": 608.88, "text": "This"}, {"end": 609.44, "start": 609.24, "text": "is"}, {"end": 609.88, "start": 609.44, "text": "why,"}, {"end": 610.36, "start": 609.88, "text": "by"}, {"end": 610.36, "start": 610.36, "text": "the"}, {"end": 610.4, "start": 610.36, "text": "way,"}, {"end": 611.0, "start": 610.4, "text": "artificial"}, {"end": 611.48, "start": 611.0, "text": "neurons"}, {"end": 611.68, "start": 611.48, "text": "have"}, {"end": 612.52, "start": 611.68, "text": "continuously"}, {"end": 613.08, "start": 612.52, "text": "ranging"}, {"end": 613.92, "start": 613.08, "text": "activations,"}, {"end": 614.28, "start": 613.92, "text": "rather"}, {"end": 614.48, "start": 614.28, "text": "than"}, {"end": 614.64, "start": 614.48, "text": "simply"}, {"end": 614.88, "start": 614.64, "text": "being"}, {"end": 614.96, "start": 614.88, "text": "active"}, {"end": 615.12, "start": 614.96, "text": "or"}, {"end": 615.16, "start": 615.12, "text": "inactive"}, {"end": 615.6, "start": 615.16, "text": "in"}, {"end": 615.76, "start": 615.6, "text": "a"}, {"end": 616.12, "start": 615.76, "text": "binary"}, {"end": 616.32, "start": 616.12, "text": "way,"}, {"end": 616.6, "start": 616.32, "text": "the"}, {"end": 616.96, "start": 616.6, "text": "way"}, {"end": 617.44, "start": 616.96, "text": "that"}, {"end": 618.0, "start": 617.44, "text": "biological"}, {"end": 618.44, "start": 618.0, "text": "neurons"}, {"end": 620.28, "start": 618.44, "text": "are."}, {"end": 620.44, "start": 620.28, "text": "This"}, {"end": 620.96, "start": 620.44, "text": "process"}, {"end": 621.16, "start": 620.96, "text": "of"}, {"end": 621.48, "start": 621.16, "text": "repeatedly"}, {"end": 621.84, "start": 621.48, "text": "nudging"}, {"end": 622.2, "start": 621.84, "text": "an"}, {"end": 622.36, "start": 622.2, "text": "input"}, {"end": 622.4, "start": 622.36, "text": "of"}, {"end": 622.68, "start": 622.4, "text": "a"}, {"end": 623.08, "start": 622.68, "text": "function"}, {"end": 623.24, "start": 623.08, "text": "by"}, {"end": 623.56, "start": 623.24, "text": "some"}, {"end": 623.8, "start": 623.56, "text": "multiple"}, {"end": 623.84, "start": 623.8, "text": "of"}, {"end": 624.16, "start": 623.84, "text": "the"}, {"end": 624.56, "start": 624.16, "text": "negative"}, {"end": 625.32, "start": 624.56, "text": "gradient"}, {"end": 625.48, "start": 625.32, "text": "is"}, {"end": 625.8, "start": 625.48, "text": "called"}, {"end": 626.32, "start": 625.8, "text": "gradient"}, {"end": 626.76, "start": 626.32, "text": "descent."}, {"end": 627.12, "start": 626.76, "text": "It's"}, {"end": 627.28, "start": 627.12, "text": "a"}, {"end": 627.76, "start": 627.28, "text": "way"}, {"end": 627.76, "start": 627.76, "text": "to"}, {"end": 628.36, "start": 627.76, "text": "converge"}, {"end": 628.56, "start": 628.36, "text": "toward"}, {"end": 628.84, "start": 628.56, "text": "some"}, {"end": 629.12, "start": 628.84, "text": "local"}, {"end": 629.64, "start": 629.12, "text": "minimum"}, {"end": 629.8, "start": 629.64, "text": "of"}, {"end": 629.8, "start": 629.8, "text": "a"}, {"end": 629.96, "start": 629.8, "text": "cost"}], "text": " The consequence of that is that it's important for this cost function to have a nice smooth output so that we can find a local minimum by taking little steps downhill. This is why, by the way, artificial neurons have continuously ranging activations, rather than simply being active or inactive in a binary way, the way that biological neurons are. This process of repeatedly nudging an input of a function by some multiple of the negative gradient is called gradient descent. It's a way to converge toward some local minimum of a cost"}, {"chunks": [{"end": 630.32, "start": 630.0, "text": "I'm"}, {"end": 630.84, "start": 630.32, "text": "still"}, {"end": 631.36, "start": 630.84, "text": "showing"}, {"end": 631.72, "start": 631.36, "text": "the"}, {"end": 632.16, "start": 631.72, "text": "picture"}, {"end": 632.2, "start": 632.16, "text": "of"}, {"end": 633.56, "start": 632.2, "text": "a"}, {"end": 634.16, "start": 633.56, "text": "function"}, {"end": 634.84, "start": 634.16, "text": "with"}, {"end": 635.16, "start": 634.84, "text": "two"}, {"end": 635.8, "start": 635.16, "text": "inputs,"}, {"end": 635.88, "start": 635.8, "text": "of"}, {"end": 636.24, "start": 635.88, "text": "course,"}, {"end": 636.72, "start": 636.24, "text": "because"}, {"end": 637.08, "start": 636.72, "text": "nudges"}, {"end": 637.4, "start": 637.08, "text": "in"}, {"end": 637.52, "start": 637.4, "text": "a"}, {"end": 638.08, "start": 637.52, "text": "13,000"}, {"end": 638.64, "start": 638.08, "text": "dimensional"}, {"end": 638.88, "start": 638.64, "text": "input"}, {"end": 639.24, "start": 638.88, "text": "space"}, {"end": 639.24, "start": 639.24, "text": "are"}, {"end": 639.24, "start": 639.24, "text": "a"}, {"end": 639.24, "start": 639.24, "text": "little"}, {"end": 639.24, "start": 639.24, "text": "hard"}, {"end": 639.28, "start": 639.24, "text": "to"}, {"end": 639.96, "start": 639.28, "text": "wrap"}, {"end": 640.36, "start": 639.96, "text": "your"}, {"end": 640.6, "start": 640.36, "text": "mind"}, {"end": 641.0, "start": 640.6, "text": "around,"}, {"end": 641.32, "start": 641.0, "text": "but"}, {"end": 641.52, "start": 641.32, "text": "there"}, {"end": 641.76, "start": 641.52, "text": "is"}, {"end": 642.12, "start": 641.76, "text": "actually"}, {"end": 642.36, "start": 642.12, "text": "a"}, {"end": 642.84, "start": 642.36, "text": "nice"}, {"end": 643.12, "start": 642.84, "text": "non-spatial"}, {"end": 643.28, "start": 643.12, "text": "way"}, {"end": 643.44, "start": 643.28, "text": "to"}, {"end": 643.68, "start": 643.44, "text": "think"}, {"end": 643.88, "start": 643.68, "text": "about"}, {"end": 645.04, "start": 643.88, "text": "this."}, {"end": 645.16, "start": 645.04, "text": "Each"}, {"end": 645.84, "start": 645.16, "text": "component"}, {"end": 645.88, "start": 645.84, "text": "of"}, {"end": 646.2, "start": 645.88, "text": "the"}, {"end": 646.72, "start": 646.2, "text": "negative"}, {"end": 647.24, "start": 646.72, "text": "gradient"}, {"end": 647.56, "start": 647.24, "text": "tells"}, {"end": 647.76, "start": 647.56, "text": "us"}, {"end": 648.0, "start": 647.76, "text": "two"}, {"end": 648.52, "start": 648.0, "text": "things."}, {"end": 648.6, "start": 648.52, "text": "The"}, {"end": 649.72, "start": 648.6, "text": "sign,"}, {"end": 650.12, "start": 649.72, "text": "of"}, {"end": 650.56, "start": 650.12, "text": "course,"}, {"end": 651.16, "start": 650.56, "text": "tells"}, {"end": 651.4, "start": 651.16, "text": "us"}, {"end": 651.88, "start": 651.4, "text": "whether"}, {"end": 651.88, "start": 651.88, "text": "the"}, {"end": 652.28, "start": 651.88, "text": "corresponding"}, {"end": 652.76, "start": 652.28, "text": "component"}, {"end": 652.92, "start": 652.76, "text": "of"}, {"end": 653.12, "start": 652.92, "text": "the"}, {"end": 653.12, "start": 653.12, "text": "input"}, {"end": 653.6, "start": 653.12, "text": "vector"}, {"end": 654.0, "start": 653.6, "text": "should"}, {"end": 654.16, "start": 654.0, "text": "be"}, {"end": 654.56, "start": 654.16, "text": "nudged"}, {"end": 655.72, "start": 654.56, "text": "up"}, {"end": 655.92, "start": 655.72, "text": "or"}, {"end": 656.56, "start": 655.92, "text": "down."}, {"end": 656.76, "start": 656.56, "text": "But"}, {"end": 657.44, "start": 656.76, "text": "importantly,"}, {"end": 657.84, "start": 657.44, "text": "the"}, {"end": 657.96, "start": 657.84, "text": "relative"}, {"end": 658.36, "start": 657.96, "text": "magnitudes"}, {"end": 658.52, "start": 658.36, "text": "of"}, {"end": 658.76, "start": 658.52, "text": "all"}, {"end": 658.92, "start": 658.76, "text": "these"}, {"end": 659.96, "start": 658.92, "text": "components"}], "text": " I'm still showing the picture of a function with two inputs, of course, because nudges in a 13,000 dimensional input space are a little hard to wrap your mind around, but there is actually a nice non-spatial way to think about this. Each component of the negative gradient tells us two things. The sign, of course, tells us whether the corresponding component of the input vector should be nudged up or down. But importantly, the relative magnitudes of all these components"}, {"chunks": [{"end": 660.84, "start": 660.0, "text": "kind"}, {"end": 661.08, "start": 660.84, "text": "of"}, {"end": 661.08, "start": 661.08, "text": "tells"}, {"end": 661.16, "start": 661.08, "text": "you"}, {"end": 661.44, "start": 661.16, "text": "which"}, {"end": 661.92, "start": 661.44, "text": "changes"}, {"end": 662.28, "start": 661.92, "text": "matter"}, {"end": 663.96, "start": 662.28, "text": "more."}, {"end": 664.56, "start": 663.96, "text": "You"}, {"end": 665.56, "start": 664.56, "text": "see,"}, {"end": 665.84, "start": 665.56, "text": "in"}, {"end": 666.12, "start": 665.84, "text": "our"}, {"end": 666.72, "start": 666.12, "text": "network,"}, {"end": 666.84, "start": 666.72, "text": "an"}, {"end": 667.4, "start": 666.84, "text": "adjustment"}, {"end": 667.48, "start": 667.4, "text": "to"}, {"end": 668.0, "start": 667.48, "text": "one"}, {"end": 668.24, "start": 668.0, "text": "of"}, {"end": 668.24, "start": 668.24, "text": "the"}, {"end": 668.4, "start": 668.24, "text": "weights"}, {"end": 668.6, "start": 668.4, "text": "might"}, {"end": 668.84, "start": 668.6, "text": "have"}, {"end": 669.12, "start": 668.84, "text": "a"}, {"end": 669.52, "start": 669.12, "text": "much"}, {"end": 670.0, "start": 669.52, "text": "greater"}, {"end": 670.28, "start": 670.0, "text": "impact"}, {"end": 670.48, "start": 670.28, "text": "on"}, {"end": 670.48, "start": 670.48, "text": "the"}, {"end": 670.84, "start": 670.48, "text": "cost"}, {"end": 671.36, "start": 670.84, "text": "function"}, {"end": 671.52, "start": 671.36, "text": "than"}, {"end": 671.64, "start": 671.52, "text": "the"}, {"end": 672.2, "start": 671.64, "text": "adjustment"}, {"end": 672.32, "start": 672.2, "text": "to"}, {"end": 672.52, "start": 672.32, "text": "some"}, {"end": 672.76, "start": 672.52, "text": "other"}, {"end": 673.48, "start": 672.76, "text": "weight."}, {"end": 673.8, "start": 673.48, "text": "Some"}, {"end": 674.8, "start": 673.8, "text": "of"}, {"end": 675.2, "start": 674.8, "text": "these"}, {"end": 675.8, "start": 675.2, "text": "connections"}, {"end": 676.0, "start": 675.8, "text": "just"}, {"end": 676.2, "start": 676.0, "text": "matter"}, {"end": 676.64, "start": 676.2, "text": "more"}, {"end": 677.2, "start": 676.64, "text": "for"}, {"end": 677.4, "start": 677.2, "text": "our"}, {"end": 677.76, "start": 677.4, "text": "training"}, {"end": 678.24, "start": 677.76, "text": "data."}, {"end": 678.44, "start": 678.24, "text": "So"}, {"end": 678.6, "start": 678.44, "text": "a"}, {"end": 678.8, "start": 678.6, "text": "way"}, {"end": 679.0, "start": 678.8, "text": "that"}, {"end": 679.08, "start": 679.0, "text": "you"}, {"end": 679.72, "start": 679.08, "text": "can"}, {"end": 680.28, "start": 679.72, "text": "think"}, {"end": 680.56, "start": 680.28, "text": "about"}, {"end": 680.68, "start": 680.56, "text": "this"}, {"end": 680.92, "start": 680.68, "text": "gradient"}, {"end": 681.32, "start": 680.92, "text": "vector"}, {"end": 681.68, "start": 681.32, "text": "of"}, {"end": 682.0, "start": 681.68, "text": "our"}, {"end": 682.8, "start": 682.0, "text": "mind-warpingly"}, {"end": 683.2, "start": 682.8, "text": "massive"}, {"end": 683.52, "start": 683.2, "text": "cost"}, {"end": 684.2, "start": 683.52, "text": "function"}, {"end": 684.2, "start": 684.2, "text": "is"}, {"end": 684.2, "start": 684.2, "text": "that"}, {"end": 684.48, "start": 684.2, "text": "it"}, {"end": 685.6, "start": 684.48, "text": "encodes"}, {"end": 685.8, "start": 685.6, "text": "the"}, {"end": 686.4, "start": 685.8, "text": "relative"}, {"end": 687.0, "start": 686.4, "text": "importance"}, {"end": 687.16, "start": 687.0, "text": "of"}, {"end": 687.32, "start": 687.16, "text": "each"}, {"end": 687.6, "start": 687.32, "text": "weight"}, {"end": 687.76, "start": 687.6, "text": "and"}, {"end": 688.32, "start": 687.76, "text": "bias,"}, {"end": 688.68, "start": 688.32, "text": "that"}, {"end": 689.32, "start": 688.68, "text": "is,"}, {"end": 689.6, "start": 689.32, "text": "which"}, {"end": 689.64, "start": 689.6, "text": "of"}, {"end": 689.8, "start": 689.64, "text": "these"}, {"end": 689.96, "start": 689.8, "text": "changes"}], "text": " kind of tells you which changes matter more. You see, in our network, an adjustment to one of the weights might have a much greater impact on the cost function than the adjustment to some other weight. Some of these connections just matter more for our training data. So a way that you can think about this gradient vector of our mind-warpingly massive cost function is that it encodes the relative importance of each weight and bias, that is, which of these changes"}, {"chunks": [{"end": 690.4, "start": 690.0, "text": "is"}, {"end": 690.52, "start": 690.4, "text": "going"}, {"end": 690.68, "start": 690.52, "text": "to"}, {"end": 690.96, "start": 690.68, "text": "carry"}, {"end": 691.12, "start": 690.96, "text": "the"}, {"end": 691.28, "start": 691.12, "text": "most"}, {"end": 691.4, "start": 691.28, "text": "bang"}, {"end": 691.76, "start": 691.4, "text": "for"}, {"end": 692.08, "start": 691.76, "text": "your"}, {"end": 692.28, "start": 692.08, "text": "buck."}, {"end": 692.84, "start": 692.28, "text": "This"}, {"end": 693.96, "start": 692.84, "text": "really"}, {"end": 694.32, "start": 693.96, "text": "is"}, {"end": 694.76, "start": 694.32, "text": "just"}, {"end": 694.92, "start": 694.76, "text": "another"}, {"end": 695.44, "start": 694.92, "text": "way"}, {"end": 695.44, "start": 695.44, "text": "of"}, {"end": 695.88, "start": 695.44, "text": "thinking"}, {"end": 696.16, "start": 695.88, "text": "about"}, {"end": 696.68, "start": 696.16, "text": "direction."}, {"end": 696.72, "start": 696.68, "text": "To"}, {"end": 697.28, "start": 696.72, "text": "take"}, {"end": 697.56, "start": 697.28, "text": "a"}, {"end": 697.96, "start": 697.56, "text": "simpler"}, {"end": 698.4, "start": 697.96, "text": "example,"}, {"end": 698.44, "start": 698.4, "text": "if"}, {"end": 698.6, "start": 698.44, "text": "you"}, {"end": 698.92, "start": 698.6, "text": "have"}, {"end": 699.12, "start": 698.92, "text": "some"}, {"end": 699.4, "start": 699.12, "text": "function"}, {"end": 699.68, "start": 699.4, "text": "with"}, {"end": 700.0, "start": 699.68, "text": "two"}, {"end": 700.72, "start": 700.0, "text": "variables"}, {"end": 700.96, "start": 700.72, "text": "as"}, {"end": 701.12, "start": 700.96, "text": "an"}, {"end": 701.28, "start": 701.12, "text": "input"}, {"end": 701.52, "start": 701.28, "text": "and"}, {"end": 701.76, "start": 701.52, "text": "you"}, {"end": 702.24, "start": 701.76, "text": "compute"}, {"end": 702.6, "start": 702.24, "text": "that"}, {"end": 702.92, "start": 702.6, "text": "its"}, {"end": 703.44, "start": 702.92, "text": "gradient"}, {"end": 703.6, "start": 703.44, "text": "at"}, {"end": 703.96, "start": 703.6, "text": "some"}, {"end": 704.56, "start": 703.96, "text": "particular"}, {"end": 705.2, "start": 704.56, "text": "point"}, {"end": 705.76, "start": 705.2, "text": "comes"}, {"end": 706.04, "start": 705.76, "text": "out"}, {"end": 706.24, "start": 706.04, "text": "as"}, {"end": 706.56, "start": 706.24, "text": "3"}, {"end": 707.0, "start": 706.56, "text": "1,"}, {"end": 707.08, "start": 707.0, "text": "then"}, {"end": 707.32, "start": 707.08, "text": "on"}, {"end": 707.84, "start": 707.32, "text": "the"}, {"end": 708.64, "start": 707.84, "text": "one"}, {"end": 708.64, "start": 708.64, "text": "hand"}, {"end": 708.64, "start": 708.64, "text": "you"}, {"end": 708.72, "start": 708.64, "text": "can"}, {"end": 709.24, "start": 708.72, "text": "interpret"}, {"end": 709.44, "start": 709.24, "text": "that"}, {"end": 709.68, "start": 709.44, "text": "as"}, {"end": 710.52, "start": 709.68, "text": "saying"}, {"end": 710.92, "start": 710.52, "text": "that"}, {"end": 711.28, "start": 710.92, "text": "when"}, {"end": 711.68, "start": 711.28, "text": "you're"}, {"end": 712.04, "start": 711.68, "text": "standing"}, {"end": 712.04, "start": 712.04, "text": "at"}, {"end": 712.04, "start": 712.04, "text": "that"}, {"end": 712.2, "start": 712.04, "text": "input,"}, {"end": 712.56, "start": 712.2, "text": "moving"}, {"end": 712.92, "start": 712.56, "text": "along"}, {"end": 713.12, "start": 712.92, "text": "this"}, {"end": 713.72, "start": 713.12, "text": "direction"}, {"end": 714.28, "start": 713.72, "text": "increases"}, {"end": 714.32, "start": 714.28, "text": "the"}, {"end": 714.72, "start": 714.32, "text": "function"}, {"end": 714.92, "start": 714.72, "text": "most"}, {"end": 715.32, "start": 714.92, "text": "quickly."}, {"end": 715.88, "start": 715.32, "text": "That"}, {"end": 716.12, "start": 715.88, "text": "when"}, {"end": 716.48, "start": 716.12, "text": "you"}, {"end": 716.88, "start": 716.48, "text": "graph"}, {"end": 716.92, "start": 716.88, "text": "the"}, {"end": 717.52, "start": 716.92, "text": "function"}, {"end": 717.68, "start": 717.52, "text": "above"}, {"end": 717.8, "start": 717.68, "text": "the"}, {"end": 718.44, "start": 717.8, "text": "plane"}, {"end": 718.44, "start": 718.44, "text": "of"}, {"end": 718.68, "start": 718.44, "text": "input"}, {"end": 719.2, "start": 718.68, "text": "points,"}, {"end": 719.36, "start": 719.2, "text": "that"}, {"end": 719.96, "start": 719.36, "text": "vector"}], "text": " is going to carry the most bang for your buck. This really is just another way of thinking about direction. To take a simpler example, if you have some function with two variables as an input and you compute that its gradient at some particular point comes out as 3 1, then on the one hand you can interpret that as saying that when you're standing at that input, moving along this direction increases the function most quickly. That when you graph the function above the plane of input points, that vector"}, {"chunks": [{"end": 720.24, "start": 720.0, "text": "is"}, {"end": 720.48, "start": 720.24, "text": "what's"}, {"end": 720.84, "start": 720.48, "text": "giving"}, {"end": 720.84, "start": 720.84, "text": "you"}, {"end": 720.96, "start": 720.84, "text": "the"}, {"end": 721.44, "start": 720.96, "text": "straight"}, {"end": 721.72, "start": 721.44, "text": "uphill"}, {"end": 722.52, "start": 721.72, "text": "direction."}, {"end": 723.0, "start": 722.52, "text": "But"}, {"end": 723.52, "start": 723.0, "text": "another"}, {"end": 723.88, "start": 723.52, "text": "way"}, {"end": 723.88, "start": 723.88, "text": "to"}, {"end": 723.88, "start": 723.88, "text": "read"}, {"end": 724.32, "start": 723.88, "text": "that"}, {"end": 724.4, "start": 724.32, "text": "is"}, {"end": 724.44, "start": 724.4, "text": "to"}, {"end": 725.04, "start": 724.44, "text": "say"}, {"end": 725.08, "start": 725.04, "text": "that"}, {"end": 725.52, "start": 725.08, "text": "changes"}, {"end": 725.68, "start": 725.52, "text": "to"}, {"end": 726.08, "start": 725.68, "text": "this"}, {"end": 726.24, "start": 726.08, "text": "first"}, {"end": 727.12, "start": 726.24, "text": "variable"}, {"end": 727.4, "start": 727.12, "text": "have"}, {"end": 727.64, "start": 727.4, "text": "three"}, {"end": 728.04, "start": 727.64, "text": "times"}, {"end": 728.28, "start": 728.04, "text": "the"}, {"end": 728.8, "start": 728.28, "text": "importance"}, {"end": 728.92, "start": 728.8, "text": "as"}, {"end": 729.32, "start": 728.92, "text": "changes"}, {"end": 729.32, "start": 729.32, "text": "to"}, {"end": 729.52, "start": 729.32, "text": "the"}, {"end": 729.84, "start": 729.52, "text": "second"}, {"end": 730.48, "start": 729.84, "text": "variable,"}, {"end": 730.88, "start": 730.48, "text": "that"}, {"end": 731.32, "start": 730.88, "text": "at"}, {"end": 731.88, "start": 731.32, "text": "least"}, {"end": 732.0, "start": 731.88, "text": "in"}, {"end": 732.28, "start": 732.0, "text": "the"}, {"end": 732.48, "start": 732.28, "text": "neighborhood"}, {"end": 732.48, "start": 732.48, "text": "of"}, {"end": 732.52, "start": 732.48, "text": "the"}, {"end": 733.16, "start": 732.52, "text": "relevant"}, {"end": 733.16, "start": 733.16, "text": "input,"}, {"end": 733.8, "start": 733.16, "text": "nudging"}, {"end": 734.24, "start": 733.8, "text": "the"}, {"end": 734.52, "start": 734.24, "text": "x"}, {"end": 735.0, "start": 734.52, "text": "value"}, {"end": 735.72, "start": 735.0, "text": "carries"}, {"end": 735.96, "start": 735.72, "text": "a"}, {"end": 735.96, "start": 735.96, "text": "lot"}, {"end": 736.0, "start": 735.96, "text": "more"}, {"end": 736.2, "start": 736.0, "text": "bang"}, {"end": 736.52, "start": 736.2, "text": "for"}, {"end": 736.76, "start": 736.52, "text": "your"}, {"end": 737.36, "start": 736.76, "text": "buck."}, {"end": 737.56, "start": 737.36, "text": "Alright,"}, {"end": 738.0, "start": 737.56, "text": "let's"}, {"end": 738.4, "start": 738.0, "text": "zoom"}, {"end": 738.76, "start": 738.4, "text": "out"}, {"end": 739.16, "start": 738.76, "text": "and"}, {"end": 739.96, "start": 739.16, "text": "sum"}, {"end": 740.0, "start": 739.96, "text": "up"}, {"end": 740.24, "start": 740.0, "text": "where"}, {"end": 740.56, "start": 740.24, "text": "we"}, {"end": 740.96, "start": 740.56, "text": "are"}, {"end": 742.0, "start": 740.96, "text": "so"}, {"end": 742.32, "start": 742.0, "text": "far."}, {"end": 742.8, "start": 742.32, "text": "The"}, {"end": 743.24, "start": 742.8, "text": "network"}, {"end": 744.0, "start": 743.24, "text": "itself"}, {"end": 744.12, "start": 744.0, "text": "is"}, {"end": 744.36, "start": 744.12, "text": "this"}, {"end": 744.72, "start": 744.36, "text": "function"}, {"end": 744.92, "start": 744.72, "text": "with"}, {"end": 745.88, "start": 744.92, "text": "784"}, {"end": 746.44, "start": 745.88, "text": "inputs"}, {"end": 746.56, "start": 746.44, "text": "and"}, {"end": 747.04, "start": 746.56, "text": "10"}, {"end": 747.48, "start": 747.04, "text": "outputs,"}, {"end": 748.2, "start": 747.48, "text": "defined"}, {"end": 748.28, "start": 748.2, "text": "in"}, {"end": 748.68, "start": 748.28, "text": "terms"}, {"end": 748.92, "start": 748.68, "text": "of"}, {"end": 748.92, "start": 748.92, "text": "all"}, {"end": 749.0, "start": 748.92, "text": "of"}, {"end": 749.16, "start": 749.0, "text": "these"}, {"end": 749.52, "start": 749.16, "text": "weighted"}, {"end": 749.96, "start": 749.52, "text": "sums."}], "text": " is what's giving you the straight uphill direction. But another way to read that is to say that changes to this first variable have three times the importance as changes to the second variable, that at least in the neighborhood of the relevant input, nudging the x value carries a lot more bang for your buck. Alright, let's zoom out and sum up where we are so far. The network itself is this function with 784 inputs and 10 outputs, defined in terms of all of these weighted sums."}, {"chunks": [{"end": 750.28, "start": 750.0, "text": "The"}, {"end": 751.16, "start": 750.28, "text": "cost"}, {"end": 751.64, "start": 751.16, "text": "function"}, {"end": 751.8, "start": 751.64, "text": "is"}, {"end": 752.08, "start": 751.8, "text": "a"}, {"end": 752.4, "start": 752.08, "text": "layer"}, {"end": 752.52, "start": 752.4, "text": "of"}, {"end": 752.92, "start": 752.52, "text": "complexity"}, {"end": 753.04, "start": 752.92, "text": "on"}, {"end": 753.4, "start": 753.04, "text": "top"}, {"end": 753.8, "start": 753.4, "text": "of"}, {"end": 754.04, "start": 753.8, "text": "that."}, {"end": 754.12, "start": 754.04, "text": "It"}, {"end": 754.64, "start": 754.12, "text": "takes"}, {"end": 754.8, "start": 754.64, "text": "the"}, {"end": 755.64, "start": 754.8, "text": "13,000"}, {"end": 756.0, "start": 755.64, "text": "weights"}, {"end": 756.04, "start": 756.0, "text": "and"}, {"end": 756.68, "start": 756.04, "text": "biases"}, {"end": 756.88, "start": 756.68, "text": "as"}, {"end": 757.68, "start": 756.88, "text": "inputs"}, {"end": 757.84, "start": 757.68, "text": "and"}, {"end": 758.2, "start": 757.84, "text": "spits"}, {"end": 758.24, "start": 758.2, "text": "out"}, {"end": 758.44, "start": 758.24, "text": "a"}, {"end": 758.84, "start": 758.44, "text": "single"}, {"end": 759.24, "start": 758.84, "text": "measure"}, {"end": 759.36, "start": 759.24, "text": "of"}, {"end": 759.76, "start": 759.36, "text": "lousiness"}, {"end": 760.12, "start": 759.76, "text": "based"}, {"end": 760.16, "start": 760.12, "text": "on"}, {"end": 760.32, "start": 760.16, "text": "the"}, {"end": 761.12, "start": 760.32, "text": "training"}, {"end": 762.24, "start": 761.12, "text": "examples."}, {"end": 762.44, "start": 762.24, "text": "And"}, {"end": 762.76, "start": 762.44, "text": "the"}, {"end": 763.48, "start": 762.76, "text": "gradient"}, {"end": 763.68, "start": 763.48, "text": "of"}, {"end": 763.8, "start": 763.68, "text": "the"}, {"end": 764.08, "start": 763.8, "text": "cost"}, {"end": 764.76, "start": 764.08, "text": "function"}, {"end": 764.96, "start": 764.76, "text": "is"}, {"end": 765.04, "start": 764.96, "text": "one"}, {"end": 765.52, "start": 765.04, "text": "more"}, {"end": 765.96, "start": 765.52, "text": "layer"}, {"end": 766.08, "start": 765.96, "text": "of"}, {"end": 766.48, "start": 766.08, "text": "complexity"}, {"end": 767.4, "start": 766.48, "text": "still."}, {"end": 767.48, "start": 767.4, "text": "It"}, {"end": 767.8, "start": 767.48, "text": "tells"}, {"end": 768.12, "start": 767.8, "text": "us"}, {"end": 768.32, "start": 768.12, "text": "what"}, {"end": 768.32, "start": 768.32, "text": "nudges"}, {"end": 768.4, "start": 768.32, "text": "to"}, {"end": 768.64, "start": 768.4, "text": "all"}, {"end": 768.76, "start": 768.64, "text": "of"}, {"end": 769.16, "start": 768.76, "text": "these"}, {"end": 769.84, "start": 769.16, "text": "weights"}, {"end": 770.2, "start": 769.84, "text": "and"}, {"end": 771.08, "start": 770.2, "text": "biases"}, {"end": 771.36, "start": 771.08, "text": "cause"}, {"end": 771.56, "start": 771.36, "text": "the"}, {"end": 772.0, "start": 771.56, "text": "fastest"}, {"end": 772.64, "start": 772.0, "text": "change"}, {"end": 772.68, "start": 772.64, "text": "to"}, {"end": 772.84, "start": 772.68, "text": "the"}, {"end": 773.24, "start": 772.84, "text": "value"}, {"end": 773.24, "start": 773.24, "text": "of"}, {"end": 773.36, "start": 773.24, "text": "the"}, {"end": 773.68, "start": 773.36, "text": "cost"}, {"end": 773.92, "start": 773.68, "text": "function,"}, {"end": 774.16, "start": 773.92, "text": "which"}, {"end": 774.24, "start": 774.16, "text": "you"}, {"end": 774.36, "start": 774.24, "text": "might"}, {"end": 775.0, "start": 774.36, "text": "interpret"}, {"end": 775.16, "start": 775.0, "text": "as"}, {"end": 775.68, "start": 775.16, "text": "saying"}, {"end": 775.92, "start": 775.68, "text": "which"}, {"end": 776.16, "start": 775.92, "text": "changes"}, {"end": 776.44, "start": 776.16, "text": "to"}, {"end": 776.68, "start": 776.44, "text": "which"}, {"end": 777.0, "start": 776.68, "text": "weights"}, {"end": 777.28, "start": 777.0, "text": "matter"}, {"end": 777.44, "start": 777.28, "text": "the"}, {"end": 779.96, "start": 777.44, "text": "most."}], "text": " The cost function is a layer of complexity on top of that. It takes the 13,000 weights and biases as inputs and spits out a single measure of lousiness based on the training examples. And the gradient of the cost function is one more layer of complexity still. It tells us what nudges to all of these weights and biases cause the fastest change to the value of the cost function, which you might interpret as saying which changes to which weights matter the most."}, {"chunks": [{"end": 781.96, "start": 780.0, "text": "So,"}, {"end": 782.6, "start": 781.96, "text": "when"}, {"end": 782.76, "start": 782.6, "text": "you"}, {"end": 783.56, "start": 782.76, "text": "initialize"}, {"end": 783.96, "start": 783.56, "text": "the"}, {"end": 784.68, "start": 783.96, "text": "network"}, {"end": 784.88, "start": 784.68, "text": "with"}, {"end": 785.12, "start": 784.88, "text": "random"}, {"end": 785.64, "start": 785.12, "text": "weights"}, {"end": 785.68, "start": 785.64, "text": "and"}, {"end": 786.4, "start": 785.68, "text": "biases,"}, {"end": 786.48, "start": 786.4, "text": "and"}, {"end": 786.76, "start": 786.48, "text": "adjust"}, {"end": 786.8, "start": 786.76, "text": "them"}, {"end": 787.2, "start": 786.8, "text": "many"}, {"end": 787.68, "start": 787.2, "text": "times"}, {"end": 787.88, "start": 787.68, "text": "based"}, {"end": 788.04, "start": 787.88, "text": "on"}, {"end": 788.12, "start": 788.04, "text": "this"}, {"end": 788.52, "start": 788.12, "text": "gradient"}, {"end": 788.84, "start": 788.52, "text": "descent"}, {"end": 789.8, "start": 788.84, "text": "process,"}, {"end": 790.44, "start": 789.8, "text": "how"}, {"end": 790.84, "start": 790.44, "text": "well"}, {"end": 790.96, "start": 790.84, "text": "does"}, {"end": 790.96, "start": 790.96, "text": "it"}, {"end": 791.0, "start": 790.96, "text": "actually"}, {"end": 791.4, "start": 791.0, "text": "perform"}, {"end": 791.68, "start": 791.4, "text": "on"}, {"end": 792.2, "start": 791.68, "text": "images"}, {"end": 792.8, "start": 792.2, "text": "that"}, {"end": 793.0, "start": 792.8, "text": "it's"}, {"end": 793.36, "start": 793.0, "text": "never"}, {"end": 793.52, "start": 793.36, "text": "seen"}, {"end": 793.76, "start": 793.52, "text": "before?"}, {"end": 793.76, "start": 793.76, "text": "Well,"}, {"end": 794.04, "start": 793.76, "text": "the"}, {"end": 794.04, "start": 794.04, "text": "one"}, {"end": 794.44, "start": 794.04, "text": "that"}, {"end": 794.6, "start": 794.44, "text": "I've"}, {"end": 794.96, "start": 794.6, "text": "described"}, {"end": 795.36, "start": 794.96, "text": "here,"}, {"end": 795.36, "start": 795.36, "text": "with"}, {"end": 795.4, "start": 795.36, "text": "the"}, {"end": 795.76, "start": 795.4, "text": "two"}, {"end": 796.0, "start": 795.76, "text": "hidden"}, {"end": 796.4, "start": 796.0, "text": "layers"}, {"end": 796.8, "start": 796.4, "text": "of"}, {"end": 797.24, "start": 796.8, "text": "16"}, {"end": 797.8, "start": 797.24, "text": "neurons"}, {"end": 798.04, "start": 797.8, "text": "each,"}, {"end": 798.64, "start": 798.04, "text": "chosen"}, {"end": 799.2, "start": 798.64, "text": "mostly"}, {"end": 799.64, "start": 799.2, "text": "for"}, {"end": 800.4, "start": 799.64, "text": "aesthetic"}, {"end": 800.96, "start": 800.4, "text": "reasons,"}, {"end": 801.16, "start": 800.96, "text": "well,"}, {"end": 801.28, "start": 801.16, "text": "it's"}, {"end": 801.52, "start": 801.28, "text": "not"}, {"end": 802.12, "start": 801.52, "text": "bad."}, {"end": 802.24, "start": 802.12, "text": "It"}, {"end": 802.88, "start": 802.24, "text": "classifies"}, {"end": 803.12, "start": 802.88, "text": "about"}, {"end": 804.28, "start": 803.12, "text": "96%"}, {"end": 804.56, "start": 804.28, "text": "of"}, {"end": 804.56, "start": 804.56, "text": "the"}, {"end": 804.6, "start": 804.56, "text": "new"}, {"end": 804.8, "start": 804.6, "text": "images"}, {"end": 804.8, "start": 804.8, "text": "that"}, {"end": 804.96, "start": 804.8, "text": "it"}, {"end": 805.36, "start": 804.96, "text": "sees"}, {"end": 805.68, "start": 805.36, "text": "correctly."}, {"end": 805.68, "start": 805.68, "text": "And"}, {"end": 805.88, "start": 805.68, "text": "honestly,"}, {"end": 806.08, "start": 805.88, "text": "if"}, {"end": 806.32, "start": 806.08, "text": "you"}, {"end": 806.44, "start": 806.32, "text": "look"}, {"end": 806.44, "start": 806.44, "text": "at"}, {"end": 806.56, "start": 806.44, "text": "some"}, {"end": 806.6, "start": 806.56, "text": "of"}, {"end": 806.8, "start": 806.6, "text": "the"}, {"end": 807.6, "start": 806.8, "text": "examples"}, {"end": 808.2, "start": 807.6, "text": "that"}, {"end": 808.8, "start": 808.2, "text": "it"}, {"end": 809.48, "start": 808.8, "text": "messes"}, {"end": 809.52, "start": 809.48, "text": "up"}, {"end": 809.96, "start": 809.52, "text": "on,"}], "text": " So, when you initialize the network with random weights and biases, and adjust them many times based on this gradient descent process, how well does it actually perform on images that it's never seen before? Well, the one that I've described here, with the two hidden layers of 16 neurons each, chosen mostly for aesthetic reasons, well, it's not bad. It classifies about 96% of the new images that it sees correctly. And honestly, if you look at some of the examples that it messes up on,"}, {"chunks": [{"end": 810.36, "start": 810.0, "text": "You"}, {"end": 810.88, "start": 810.36, "text": "kind"}, {"end": 811.0, "start": 810.88, "text": "of"}, {"end": 811.4, "start": 811.0, "text": "feel"}, {"end": 812.32, "start": 811.4, "text": "compelled"}, {"end": 812.36, "start": 812.32, "text": "to"}, {"end": 812.36, "start": 812.36, "text": "cut"}, {"end": 812.4, "start": 812.36, "text": "it"}, {"end": 812.6, "start": 812.4, "text": "a"}, {"end": 813.48, "start": 812.6, "text": "little"}, {"end": 813.92, "start": 813.48, "text": "slack."}, {"end": 814.44, "start": 813.92, "text": "Now"}, {"end": 815.32, "start": 814.44, "text": "if"}, {"end": 815.48, "start": 815.32, "text": "you"}, {"end": 816.52, "start": 815.48, "text": "play"}, {"end": 817.12, "start": 816.52, "text": "around"}, {"end": 817.48, "start": 817.12, "text": "with"}, {"end": 818.04, "start": 817.48, "text": "the"}, {"end": 818.4, "start": 818.04, "text": "hidden"}, {"end": 818.84, "start": 818.4, "text": "layer"}, {"end": 819.2, "start": 818.84, "text": "structure"}, {"end": 819.2, "start": 819.2, "text": "and"}, {"end": 819.24, "start": 819.2, "text": "make"}, {"end": 819.32, "start": 819.24, "text": "a"}, {"end": 819.32, "start": 819.32, "text": "couple"}, {"end": 819.84, "start": 819.32, "text": "tweaks,"}, {"end": 819.96, "start": 819.84, "text": "you"}, {"end": 819.96, "start": 819.96, "text": "can"}, {"end": 819.96, "start": 819.96, "text": "get"}, {"end": 820.4, "start": 819.96, "text": "this"}, {"end": 820.4, "start": 820.4, "text": "up"}, {"end": 820.6, "start": 820.4, "text": "to"}, {"end": 821.6, "start": 820.6, "text": "98%."}, {"end": 821.92, "start": 821.6, "text": "And"}, {"end": 822.2, "start": 821.92, "text": "that's"}, {"end": 822.8, "start": 822.2, "text": "pretty"}, {"end": 823.08, "start": 822.8, "text": "good."}, {"end": 823.2, "start": 823.08, "text": "It's"}, {"end": 823.36, "start": 823.2, "text": "not"}, {"end": 823.52, "start": 823.36, "text": "the"}, {"end": 823.96, "start": 823.52, "text": "best,"}, {"end": 823.96, "start": 823.96, "text": "you"}, {"end": 824.16, "start": 823.96, "text": "can"}, {"end": 824.72, "start": 824.16, "text": "certainly"}, {"end": 824.88, "start": 824.72, "text": "get"}, {"end": 825.28, "start": 824.88, "text": "better"}, {"end": 825.64, "start": 825.28, "text": "performance"}, {"end": 825.64, "start": 825.64, "text": "by"}, {"end": 825.68, "start": 825.64, "text": "getting"}, {"end": 826.24, "start": 825.68, "text": "more"}, {"end": 827.0, "start": 826.24, "text": "sophisticated"}, {"end": 827.0, "start": 827.0, "text": "than"}, {"end": 827.36, "start": 827.0, "text": "this"}, {"end": 827.68, "start": 827.36, "text": "plain"}, {"end": 828.08, "start": 827.68, "text": "vanilla"}, {"end": 828.48, "start": 828.08, "text": "network,"}, {"end": 828.48, "start": 828.48, "text": "but"}, {"end": 828.6, "start": 828.48, "text": "given"}, {"end": 829.2, "start": 828.6, "text": "how"}, {"end": 829.32, "start": 829.2, "text": "daunting"}, {"end": 829.6, "start": 829.32, "text": "the"}, {"end": 829.92, "start": 829.6, "text": "initial"}, {"end": 830.24, "start": 829.92, "text": "task"}, {"end": 830.92, "start": 830.24, "text": "is,"}, {"end": 831.2, "start": 830.92, "text": "I"}, {"end": 831.52, "start": 831.2, "text": "just"}, {"end": 831.88, "start": 831.52, "text": "think"}, {"end": 832.48, "start": 831.88, "text": "there's"}, {"end": 832.68, "start": 832.48, "text": "something"}, {"end": 833.52, "start": 832.68, "text": "incredible"}, {"end": 833.88, "start": 833.52, "text": "about"}, {"end": 834.16, "start": 833.88, "text": "any"}, {"end": 834.64, "start": 834.16, "text": "network"}, {"end": 834.92, "start": 834.64, "text": "doing"}, {"end": 835.16, "start": 834.92, "text": "this"}, {"end": 835.56, "start": 835.16, "text": "well"}, {"end": 835.72, "start": 835.56, "text": "on"}, {"end": 835.92, "start": 835.72, "text": "images"}, {"end": 835.96, "start": 835.92, "text": "that"}, {"end": 836.28, "start": 835.96, "text": "it's"}, {"end": 836.52, "start": 836.28, "text": "never"}, {"end": 836.8, "start": 836.52, "text": "seen"}, {"end": 837.36, "start": 836.8, "text": "before,"}, {"end": 837.56, "start": 837.36, "text": "given"}, {"end": 837.56, "start": 837.56, "text": "that"}, {"end": 837.72, "start": 837.56, "text": "we"}, {"end": 838.44, "start": 837.72, "text": "never"}, {"end": 839.16, "start": 838.44, "text": "specifically"}, {"end": 839.32, "start": 839.16, "text": "told"}, {"end": 839.36, "start": 839.32, "text": "it"}, {"end": 839.96, "start": 839.36, "text": "what"}], "text": " You kind of feel compelled to cut it a little slack. Now if you play around with the hidden layer structure and make a couple tweaks, you can get this up to 98%. And that's pretty good. It's not the best, you can certainly get better performance by getting more sophisticated than this plain vanilla network, but given how daunting the initial task is, I just think there's something incredible about any network doing this well on images that it's never seen before, given that we never specifically told it what"}, {"chunks": [{"end": 840.36, "start": 840.0, "text": "patterns"}, {"end": 840.44, "start": 840.36, "text": "to"}, {"end": 840.72, "start": 840.44, "text": "look"}, {"end": 842.6, "start": 840.72, "text": "for."}, {"end": 843.12, "start": 842.6, "text": "Originally,"}, {"end": 843.28, "start": 843.12, "text": "the"}, {"end": 843.64, "start": 843.28, "text": "way"}, {"end": 843.84, "start": 843.64, "text": "I"}, {"end": 844.24, "start": 843.84, "text": "motivated"}, {"end": 844.44, "start": 844.24, "text": "this"}, {"end": 844.96, "start": 844.44, "text": "structure"}, {"end": 845.24, "start": 844.96, "text": "was"}, {"end": 845.36, "start": 845.24, "text": "by"}, {"end": 845.96, "start": 845.36, "text": "describing"}, {"end": 846.4, "start": 845.96, "text": "a"}, {"end": 847.04, "start": 846.4, "text": "hope"}, {"end": 847.04, "start": 847.04, "text": "that"}, {"end": 847.36, "start": 847.04, "text": "we"}, {"end": 847.52, "start": 847.36, "text": "might"}, {"end": 847.52, "start": 847.52, "text": "have."}, {"end": 847.64, "start": 847.52, "text": "That"}, {"end": 847.88, "start": 847.64, "text": "the"}, {"end": 848.6, "start": 847.88, "text": "second"}, {"end": 849.32, "start": 848.6, "text": "layer"}, {"end": 849.92, "start": 849.32, "text": "might"}, {"end": 849.96, "start": 849.92, "text": "pick"}, {"end": 850.2, "start": 849.96, "text": "up"}, {"end": 850.28, "start": 850.2, "text": "on"}, {"end": 850.4, "start": 850.28, "text": "little"}, {"end": 850.4, "start": 850.4, "text": "edges,"}, {"end": 850.4, "start": 850.4, "text": "that"}, {"end": 850.44, "start": 850.4, "text": "the"}, {"end": 850.96, "start": 850.44, "text": "third"}, {"end": 851.28, "start": 850.96, "text": "layer"}, {"end": 851.32, "start": 851.28, "text": "would"}, {"end": 851.52, "start": 851.32, "text": "piece"}, {"end": 851.88, "start": 851.52, "text": "together"}, {"end": 852.2, "start": 851.88, "text": "those"}, {"end": 852.4, "start": 852.2, "text": "edges"}, {"end": 852.52, "start": 852.4, "text": "to"}, {"end": 853.04, "start": 852.52, "text": "recognize"}, {"end": 853.52, "start": 853.04, "text": "loops"}, {"end": 853.68, "start": 853.52, "text": "and"}, {"end": 853.8, "start": 853.68, "text": "longer"}, {"end": 854.68, "start": 853.8, "text": "lines,"}, {"end": 854.88, "start": 854.68, "text": "and"}, {"end": 854.92, "start": 854.88, "text": "that"}, {"end": 855.2, "start": 854.92, "text": "those"}, {"end": 855.52, "start": 855.2, "text": "might"}, {"end": 855.52, "start": 855.52, "text": "be"}, {"end": 855.72, "start": 855.52, "text": "pieced"}, {"end": 856.08, "start": 855.72, "text": "together"}, {"end": 856.24, "start": 856.08, "text": "to"}, {"end": 856.72, "start": 856.24, "text": "recognize"}, {"end": 858.08, "start": 856.72, "text": "digits."}, {"end": 858.4, "start": 858.08, "text": "So,"}, {"end": 858.76, "start": 858.4, "text": "is"}, {"end": 859.04, "start": 858.76, "text": "this"}, {"end": 859.08, "start": 859.04, "text": "what"}, {"end": 859.28, "start": 859.08, "text": "our"}, {"end": 859.56, "start": 859.28, "text": "network"}, {"end": 859.72, "start": 859.56, "text": "is"}, {"end": 860.04, "start": 859.72, "text": "actually"}, {"end": 860.36, "start": 860.04, "text": "doing?"}, {"end": 860.36, "start": 860.36, "text": "Well,"}, {"end": 860.36, "start": 860.36, "text": "for"}, {"end": 861.08, "start": 860.36, "text": "this"}, {"end": 861.64, "start": 861.08, "text": "one"}, {"end": 862.08, "start": 861.64, "text": "at"}, {"end": 862.36, "start": 862.08, "text": "least,"}, {"end": 863.0, "start": 862.36, "text": "not"}, {"end": 863.08, "start": 863.0, "text": "at"}, {"end": 863.52, "start": 863.08, "text": "all."}, {"end": 863.68, "start": 863.52, "text": "Remember"}, {"end": 864.36, "start": 863.68, "text": "how"}, {"end": 864.52, "start": 864.36, "text": "last"}, {"end": 864.6, "start": 864.52, "text": "video"}, {"end": 864.6, "start": 864.6, "text": "we"}, {"end": 864.76, "start": 864.6, "text": "looked"}, {"end": 865.84, "start": 864.76, "text": "at"}, {"end": 866.08, "start": 865.84, "text": "how"}, {"end": 866.36, "start": 866.08, "text": "the"}, {"end": 867.2, "start": 866.36, "text": "weights"}, {"end": 867.24, "start": 867.2, "text": "of"}, {"end": 867.36, "start": 867.24, "text": "the"}, {"end": 868.04, "start": 867.36, "text": "connections"}, {"end": 868.2, "start": 868.04, "text": "from"}, {"end": 868.28, "start": 868.2, "text": "all"}, {"end": 868.64, "start": 868.28, "text": "of"}, {"end": 868.76, "start": 868.64, "text": "the"}, {"end": 869.2, "start": 868.76, "text": "neurons"}, {"end": 869.2, "start": 869.2, "text": "in"}, {"end": 869.36, "start": 869.2, "text": "the"}, {"end": 869.68, "start": 869.36, "text": "first"}, {"end": 869.96, "start": 869.68, "text": "layer"}], "text": " patterns to look for. Originally, the way I motivated this structure was by describing a hope that we might have. That the second layer might pick up on little edges, that the third layer would piece together those edges to recognize loops and longer lines, and that those might be pieced together to recognize digits. So, is this what our network is actually doing? Well, for this one at least, not at all. Remember how last video we looked at how the weights of the connections from all of the neurons in the first layer"}, {"chunks": [{"end": 870.08, "start": 870.0, "text": "to"}, {"end": 870.12, "start": 870.08, "text": "a"}, {"end": 870.28, "start": 870.12, "text": "given"}, {"end": 870.6, "start": 870.28, "text": "neuron"}, {"end": 870.8, "start": 870.6, "text": "in"}, {"end": 871.32, "start": 870.8, "text": "the"}, {"end": 871.68, "start": 871.32, "text": "second"}, {"end": 872.24, "start": 871.68, "text": "layer"}, {"end": 872.6, "start": 872.24, "text": "can"}, {"end": 872.6, "start": 872.6, "text": "be"}, {"end": 873.28, "start": 872.6, "text": "visualized"}, {"end": 873.28, "start": 873.28, "text": "as"}, {"end": 873.36, "start": 873.28, "text": "a"}, {"end": 873.88, "start": 873.36, "text": "given"}, {"end": 874.32, "start": 873.88, "text": "pixel"}, {"end": 874.8, "start": 874.32, "text": "pattern"}, {"end": 874.8, "start": 874.8, "text": "that"}, {"end": 874.88, "start": 874.8, "text": "that"}, {"end": 875.56, "start": 874.88, "text": "second"}, {"end": 875.76, "start": 875.56, "text": "layer"}, {"end": 876.08, "start": 875.76, "text": "neuron"}, {"end": 876.24, "start": 876.08, "text": "is"}, {"end": 876.72, "start": 876.24, "text": "picking"}, {"end": 877.72, "start": 876.72, "text": "up"}, {"end": 878.24, "start": 877.72, "text": "on?"}, {"end": 878.84, "start": 878.24, "text": "Well,"}, {"end": 879.28, "start": 878.84, "text": "when"}, {"end": 879.28, "start": 879.28, "text": "we"}, {"end": 879.36, "start": 879.28, "text": "actually"}, {"end": 879.36, "start": 879.36, "text": "do"}, {"end": 879.36, "start": 879.36, "text": "that"}, {"end": 879.36, "start": 879.36, "text": "for"}, {"end": 879.48, "start": 879.36, "text": "the"}, {"end": 880.04, "start": 879.48, "text": "weights"}, {"end": 880.76, "start": 880.04, "text": "associated"}, {"end": 880.96, "start": 880.76, "text": "with"}, {"end": 881.4, "start": 880.96, "text": "these"}, {"end": 882.0, "start": 881.4, "text": "transitions"}, {"end": 882.04, "start": 882.0, "text": "from"}, {"end": 882.24, "start": 882.04, "text": "the"}, {"end": 882.8, "start": 882.24, "text": "first"}, {"end": 883.16, "start": 882.8, "text": "layer"}, {"end": 883.16, "start": 883.16, "text": "to"}, {"end": 883.16, "start": 883.16, "text": "the"}, {"end": 883.92, "start": 883.16, "text": "next,"}, {"end": 884.24, "start": 883.92, "text": "instead"}, {"end": 884.32, "start": 884.24, "text": "of"}, {"end": 884.64, "start": 884.32, "text": "picking"}, {"end": 884.84, "start": 884.64, "text": "up"}, {"end": 885.04, "start": 884.84, "text": "on"}, {"end": 885.6, "start": 885.04, "text": "isolated"}, {"end": 885.88, "start": 885.6, "text": "little"}, {"end": 886.2, "start": 885.88, "text": "edges"}, {"end": 886.52, "start": 886.2, "text": "here"}, {"end": 886.88, "start": 886.52, "text": "and"}, {"end": 887.32, "start": 886.88, "text": "there,"}, {"end": 887.44, "start": 887.32, "text": "they"}, {"end": 888.44, "start": 887.44, "text": "look,"}, {"end": 889.36, "start": 888.44, "text": "well,"}, {"end": 889.76, "start": 889.36, "text": "almost"}, {"end": 890.68, "start": 889.76, "text": "random,"}, {"end": 890.88, "start": 890.68, "text": "just"}, {"end": 890.96, "start": 890.88, "text": "with"}, {"end": 891.44, "start": 890.96, "text": "some"}, {"end": 891.92, "start": 891.44, "text": "very"}, {"end": 892.16, "start": 891.92, "text": "loose"}, {"end": 892.84, "start": 892.16, "text": "patterns"}, {"end": 893.32, "start": 892.84, "text": "in"}, {"end": 893.68, "start": 893.32, "text": "the"}, {"end": 893.8, "start": 893.68, "text": "middle"}, {"end": 893.84, "start": 893.8, "text": "there."}, {"end": 893.96, "start": 893.84, "text": "It"}, {"end": 894.04, "start": 893.96, "text": "would"}, {"end": 894.04, "start": 894.04, "text": "seem"}, {"end": 894.68, "start": 894.04, "text": "that"}, {"end": 895.2, "start": 894.68, "text": "in"}, {"end": 895.28, "start": 895.2, "text": "the"}, {"end": 896.12, "start": 895.28, "text": "unfathomably"}, {"end": 896.56, "start": 896.12, "text": "large"}, {"end": 897.44, "start": 896.56, "text": "13,000"}, {"end": 897.88, "start": 897.44, "text": "dimensional"}, {"end": 898.28, "start": 897.88, "text": "space"}, {"end": 898.28, "start": 898.28, "text": "of"}, {"end": 898.6, "start": 898.28, "text": "possible"}, {"end": 899.08, "start": 898.6, "text": "weights"}, {"end": 899.08, "start": 899.08, "text": "and"}, {"end": 899.96, "start": 899.08, "text": "biases,"}], "text": " to a given neuron in the second layer can be visualized as a given pixel pattern that that second layer neuron is picking up on? Well, when we actually do that for the weights associated with these transitions from the first layer to the next, instead of picking up on isolated little edges here and there, they look, well, almost random, just with some very loose patterns in the middle there. It would seem that in the unfathomably large 13,000 dimensional space of possible weights and biases,"}, {"chunks": [{"end": 900.28, "start": 900.0, "text": "Our"}, {"end": 900.68, "start": 900.28, "text": "network"}, {"end": 900.96, "start": 900.68, "text": "found"}, {"end": 901.32, "start": 900.96, "text": "itself"}, {"end": 901.52, "start": 901.32, "text": "a"}, {"end": 901.6, "start": 901.52, "text": "happy"}, {"end": 902.2, "start": 901.6, "text": "little"}, {"end": 902.6, "start": 902.2, "text": "local"}, {"end": 902.92, "start": 902.6, "text": "minimum"}, {"end": 903.04, "start": 902.92, "text": "that,"}, {"end": 903.4, "start": 903.04, "text": "despite"}, {"end": 904.0, "start": 903.4, "text": "successfully"}, {"end": 904.72, "start": 904.0, "text": "classifying"}, {"end": 904.88, "start": 904.72, "text": "most"}, {"end": 905.48, "start": 904.88, "text": "images,"}, {"end": 905.48, "start": 905.48, "text": "doesn't"}, {"end": 905.52, "start": 905.48, "text": "exactly"}, {"end": 905.72, "start": 905.52, "text": "pick"}, {"end": 905.76, "start": 905.72, "text": "up"}, {"end": 905.84, "start": 905.76, "text": "on"}, {"end": 905.84, "start": 905.84, "text": "the"}, {"end": 906.2, "start": 905.84, "text": "patterns"}, {"end": 906.6, "start": 906.2, "text": "that"}, {"end": 906.8, "start": 906.6, "text": "we"}, {"end": 907.0, "start": 906.8, "text": "might"}, {"end": 907.28, "start": 907.0, "text": "have"}, {"end": 907.84, "start": 907.28, "text": "hoped"}, {"end": 908.2, "start": 907.84, "text": "for."}, {"end": 908.36, "start": 908.2, "text": "And"}, {"end": 908.48, "start": 908.36, "text": "to"}, {"end": 908.52, "start": 908.48, "text": "really"}, {"end": 908.8, "start": 908.52, "text": "drive"}, {"end": 909.2, "start": 908.8, "text": "this"}, {"end": 910.2, "start": 909.2, "text": "point"}, {"end": 910.76, "start": 910.2, "text": "home,"}, {"end": 911.48, "start": 910.76, "text": "watch"}, {"end": 911.72, "start": 911.48, "text": "what"}, {"end": 912.32, "start": 911.72, "text": "happens"}, {"end": 912.56, "start": 912.32, "text": "when"}, {"end": 912.56, "start": 912.56, "text": "you"}, {"end": 912.6, "start": 912.56, "text": "input"}, {"end": 912.84, "start": 912.6, "text": "a"}, {"end": 913.68, "start": 912.84, "text": "random"}, {"end": 914.32, "start": 913.68, "text": "image."}, {"end": 914.4, "start": 914.32, "text": "If"}, {"end": 914.64, "start": 914.4, "text": "the"}, {"end": 914.96, "start": 914.64, "text": "system"}, {"end": 915.4, "start": 914.96, "text": "was"}, {"end": 915.84, "start": 915.4, "text": "smart,"}, {"end": 915.88, "start": 915.84, "text": "you"}, {"end": 916.2, "start": 915.88, "text": "might"}, {"end": 916.72, "start": 916.2, "text": "expect"}, {"end": 916.72, "start": 916.72, "text": "it"}, {"end": 916.72, "start": 916.72, "text": "to"}, {"end": 917.32, "start": 916.72, "text": "either"}, {"end": 917.68, "start": 917.32, "text": "feel"}, {"end": 918.24, "start": 917.68, "text": "uncertain,"}, {"end": 918.68, "start": 918.24, "text": "maybe"}, {"end": 918.76, "start": 918.68, "text": "not"}, {"end": 919.28, "start": 918.76, "text": "really"}, {"end": 919.84, "start": 919.28, "text": "activating"}, {"end": 920.08, "start": 919.84, "text": "any"}, {"end": 920.08, "start": 920.08, "text": "of"}, {"end": 920.24, "start": 920.08, "text": "those"}, {"end": 920.44, "start": 920.24, "text": "10"}, {"end": 920.72, "start": 920.44, "text": "output"}, {"end": 921.44, "start": 920.72, "text": "neurons"}, {"end": 921.52, "start": 921.44, "text": "or"}, {"end": 921.92, "start": 921.52, "text": "activating"}, {"end": 922.16, "start": 921.92, "text": "them"}, {"end": 922.44, "start": 922.16, "text": "all"}, {"end": 923.24, "start": 922.44, "text": "evenly."}, {"end": 923.64, "start": 923.24, "text": "But"}, {"end": 924.56, "start": 923.64, "text": "instead,"}, {"end": 924.68, "start": 924.56, "text": "it"}, {"end": 925.4, "start": 924.68, "text": "confidently"}, {"end": 925.64, "start": 925.4, "text": "gives"}, {"end": 925.72, "start": 925.64, "text": "you"}, {"end": 925.96, "start": 925.72, "text": "some"}, {"end": 926.6, "start": 925.96, "text": "nonsense"}, {"end": 926.96, "start": 926.6, "text": "answer,"}, {"end": 926.96, "start": 926.96, "text": "as"}, {"end": 927.0, "start": 926.96, "text": "if"}, {"end": 927.04, "start": 927.0, "text": "it"}, {"end": 927.72, "start": 927.04, "text": "feels"}, {"end": 927.92, "start": 927.72, "text": "as"}, {"end": 927.92, "start": 927.92, "text": "sure"}, {"end": 927.92, "start": 927.92, "text": "that"}, {"end": 928.48, "start": 927.92, "text": "this"}, {"end": 928.88, "start": 928.48, "text": "random"}, {"end": 929.72, "start": 928.88, "text": "noise"}, {"end": 929.96, "start": 929.72, "text": "is"}], "text": " Our network found itself a happy little local minimum that, despite successfully classifying most images, doesn't exactly pick up on the patterns that we might have hoped for. And to really drive this point home, watch what happens when you input a random image. If the system was smart, you might expect it to either feel uncertain, maybe not really activating any of those 10 output neurons or activating them all evenly. But instead, it confidently gives you some nonsense answer, as if it feels as sure that this random noise is"}, {"chunks": [{"end": 930.12, "start": 930.0, "text": "the"}, {"end": 930.64, "start": 930.12, "text": "5"}, {"end": 930.72, "start": 930.64, "text": "as"}, {"end": 930.92, "start": 930.72, "text": "it"}, {"end": 931.16, "start": 930.92, "text": "does"}, {"end": 931.44, "start": 931.16, "text": "that"}, {"end": 931.72, "start": 931.44, "text": "an"}, {"end": 932.24, "start": 931.72, "text": "actual"}, {"end": 932.88, "start": 932.24, "text": "image"}, {"end": 933.12, "start": 932.88, "text": "of"}, {"end": 933.12, "start": 933.12, "text": "a"}, {"end": 933.16, "start": 933.12, "text": "5"}, {"end": 933.32, "start": 933.16, "text": "is"}, {"end": 933.44, "start": 933.32, "text": "a"}, {"end": 934.48, "start": 933.44, "text": "5."}, {"end": 934.8, "start": 934.48, "text": "Phrased"}, {"end": 935.2, "start": 934.8, "text": "differently,"}, {"end": 935.36, "start": 935.2, "text": "even"}, {"end": 935.72, "start": 935.36, "text": "if"}, {"end": 936.32, "start": 935.72, "text": "this"}, {"end": 936.68, "start": 936.32, "text": "network"}, {"end": 936.92, "start": 936.68, "text": "can"}, {"end": 937.48, "start": 936.92, "text": "recognize"}, {"end": 937.76, "start": 937.48, "text": "digits"}, {"end": 937.88, "start": 937.76, "text": "pretty"}, {"end": 938.28, "start": 937.88, "text": "well,"}, {"end": 938.76, "start": 938.28, "text": "it"}, {"end": 939.16, "start": 938.76, "text": "has"}, {"end": 939.6, "start": 939.16, "text": "no"}, {"end": 940.12, "start": 939.6, "text": "idea"}, {"end": 940.92, "start": 940.12, "text": "how"}, {"end": 941.24, "start": 940.92, "text": "to"}, {"end": 941.32, "start": 941.24, "text": "draw"}, {"end": 941.36, "start": 941.32, "text": "them."}, {"end": 941.44, "start": 941.36, "text": "A"}, {"end": 941.8, "start": 941.44, "text": "lot"}, {"end": 942.08, "start": 941.8, "text": "of"}, {"end": 942.32, "start": 942.08, "text": "this"}, {"end": 942.48, "start": 942.32, "text": "is"}, {"end": 942.88, "start": 942.48, "text": "because"}, {"end": 943.2, "start": 942.88, "text": "it's"}, {"end": 943.32, "start": 943.2, "text": "such"}, {"end": 943.36, "start": 943.32, "text": "a"}, {"end": 943.8, "start": 943.36, "text": "tightly"}, {"end": 944.4, "start": 943.8, "text": "constrained"}, {"end": 944.8, "start": 944.4, "text": "training"}, {"end": 945.76, "start": 944.8, "text": "setup."}, {"end": 946.04, "start": 945.76, "text": "I"}, {"end": 946.36, "start": 946.04, "text": "mean,"}, {"end": 946.76, "start": 946.36, "text": "put"}, {"end": 947.56, "start": 946.76, "text": "yourself"}, {"end": 947.64, "start": 947.56, "text": "in"}, {"end": 947.64, "start": 947.64, "text": "the"}, {"end": 948.08, "start": 947.64, "text": "network's"}, {"end": 948.44, "start": 948.08, "text": "shoes"}, {"end": 948.72, "start": 948.44, "text": "here."}, {"end": 949.08, "start": 948.72, "text": "From"}, {"end": 949.16, "start": 949.08, "text": "its"}, {"end": 949.2, "start": 949.16, "text": "point"}, {"end": 949.24, "start": 949.2, "text": "of"}, {"end": 949.24, "start": 949.24, "text": "view,"}, {"end": 949.28, "start": 949.24, "text": "the"}, {"end": 950.08, "start": 949.28, "text": "entire"}, {"end": 950.6, "start": 950.08, "text": "universe"}, {"end": 951.32, "start": 950.6, "text": "consists"}, {"end": 951.48, "start": 951.32, "text": "of"}, {"end": 951.72, "start": 951.48, "text": "nothing"}, {"end": 951.88, "start": 951.72, "text": "but"}, {"end": 952.36, "start": 951.88, "text": "clearly"}, {"end": 952.84, "start": 952.36, "text": "defined"}, {"end": 953.36, "start": 952.84, "text": "unmoving"}, {"end": 953.92, "start": 953.36, "text": "digits"}, {"end": 954.08, "start": 953.92, "text": "centered"}, {"end": 954.32, "start": 954.08, "text": "in"}, {"end": 954.6, "start": 954.32, "text": "a"}, {"end": 954.96, "start": 954.6, "text": "tiny"}, {"end": 955.28, "start": 954.96, "text": "grid,"}, {"end": 955.52, "start": 955.28, "text": "and"}, {"end": 955.72, "start": 955.52, "text": "its"}, {"end": 956.04, "start": 955.72, "text": "cost"}, {"end": 956.44, "start": 956.04, "text": "function"}, {"end": 956.64, "start": 956.44, "text": "just"}, {"end": 957.04, "start": 956.64, "text": "never"}, {"end": 957.36, "start": 957.04, "text": "gave"}, {"end": 957.36, "start": 957.36, "text": "it"}, {"end": 957.76, "start": 957.36, "text": "any"}, {"end": 958.24, "start": 957.76, "text": "incentive"}, {"end": 958.36, "start": 958.24, "text": "to"}, {"end": 958.56, "start": 958.36, "text": "be"}, {"end": 959.0, "start": 958.56, "text": "anything"}, {"end": 959.28, "start": 959.0, "text": "but"}, {"end": 959.96, "start": 959.28, "text": "utterly"}], "text": " the 5 as it does that an actual image of a 5 is a 5. Phrased differently, even if this network can recognize digits pretty well, it has no idea how to draw them. A lot of this is because it's such a tightly constrained training setup. I mean, put yourself in the network's shoes here. From its point of view, the entire universe consists of nothing but clearly defined unmoving digits centered in a tiny grid, and its cost function just never gave it any incentive to be anything but utterly"}, {"chunks": [{"end": 960.24, "start": 960.0, "text": "So"}, {"end": 960.4, "start": 960.24, "text": "with"}, {"end": 960.96, "start": 960.4, "text": "this"}, {"end": 961.52, "start": 960.96, "text": "as"}, {"end": 961.64, "start": 961.52, "text": "the"}, {"end": 961.72, "start": 961.64, "text": "image"}, {"end": 961.76, "start": 961.72, "text": "of"}, {"end": 962.04, "start": 961.76, "text": "what"}, {"end": 962.48, "start": 962.04, "text": "those"}, {"end": 963.08, "start": 962.48, "text": "second"}, {"end": 963.36, "start": 963.08, "text": "layer"}, {"end": 963.88, "start": 963.36, "text": "neurons"}, {"end": 964.12, "start": 963.88, "text": "are"}, {"end": 964.8, "start": 964.12, "text": "really"}, {"end": 965.12, "start": 964.8, "text": "doing,"}, {"end": 965.12, "start": 965.12, "text": "you"}, {"end": 965.32, "start": 965.12, "text": "might"}, {"end": 965.56, "start": 965.32, "text": "wonder"}, {"end": 965.84, "start": 965.56, "text": "why"}, {"end": 966.0, "start": 965.84, "text": "I"}, {"end": 966.56, "start": 966.0, "text": "would"}, {"end": 967.04, "start": 966.56, "text": "introduce"}, {"end": 967.24, "start": 967.04, "text": "this"}, {"end": 967.68, "start": 967.24, "text": "network"}, {"end": 967.76, "start": 967.68, "text": "with"}, {"end": 967.92, "start": 967.76, "text": "the"}, {"end": 968.52, "start": 967.92, "text": "motivation"}, {"end": 968.52, "start": 968.52, "text": "of"}, {"end": 968.88, "start": 968.52, "text": "picking"}, {"end": 968.92, "start": 968.88, "text": "up"}, {"end": 969.12, "start": 968.92, "text": "on"}, {"end": 969.44, "start": 969.12, "text": "edges"}, {"end": 969.48, "start": 969.44, "text": "and"}, {"end": 970.12, "start": 969.48, "text": "patterns."}, {"end": 970.36, "start": 970.12, "text": "I"}, {"end": 970.36, "start": 970.36, "text": "mean,"}, {"end": 970.48, "start": 970.36, "text": "that's"}, {"end": 970.8, "start": 970.48, "text": "just"}, {"end": 970.8, "start": 970.8, "text": "not"}, {"end": 970.92, "start": 970.8, "text": "at"}, {"end": 972.0, "start": 970.92, "text": "all"}, {"end": 972.32, "start": 972.0, "text": "what"}, {"end": 972.4, "start": 972.32, "text": "it"}, {"end": 972.44, "start": 972.4, "text": "ends"}, {"end": 972.44, "start": 972.44, "text": "up"}, {"end": 973.4, "start": 972.44, "text": "doing."}, {"end": 973.68, "start": 973.4, "text": "Well,"}, {"end": 973.96, "start": 973.68, "text": "this"}, {"end": 974.2, "start": 973.96, "text": "is"}, {"end": 974.8, "start": 974.2, "text": "not"}, {"end": 975.24, "start": 974.8, "text": "meant"}, {"end": 975.24, "start": 975.24, "text": "to"}, {"end": 975.28, "start": 975.24, "text": "be"}, {"end": 975.44, "start": 975.28, "text": "our"}, {"end": 975.72, "start": 975.44, "text": "end"}, {"end": 975.96, "start": 975.72, "text": "goal,"}, {"end": 976.04, "start": 975.96, "text": "but"}, {"end": 976.56, "start": 976.04, "text": "instead"}, {"end": 976.88, "start": 976.56, "text": "a"}, {"end": 977.08, "start": 976.88, "text": "starting"}, {"end": 977.52, "start": 977.08, "text": "point."}, {"end": 978.0, "start": 977.52, "text": "Frankly,"}, {"end": 978.08, "start": 978.0, "text": "this"}, {"end": 978.48, "start": 978.08, "text": "is"}, {"end": 978.84, "start": 978.48, "text": "old"}, {"end": 979.28, "start": 978.84, "text": "technology,"}, {"end": 979.36, "start": 979.28, "text": "the"}, {"end": 979.84, "start": 979.36, "text": "kind"}, {"end": 980.44, "start": 979.84, "text": "researched"}, {"end": 980.88, "start": 980.44, "text": "in"}, {"end": 981.08, "start": 980.88, "text": "the"}, {"end": 981.68, "start": 981.08, "text": "80s"}, {"end": 981.8, "start": 981.68, "text": "and"}, {"end": 981.8, "start": 981.8, "text": "90s."}, {"end": 981.8, "start": 981.8, "text": "And"}, {"end": 981.8, "start": 981.8, "text": "you"}, {"end": 981.84, "start": 981.8, "text": "do"}, {"end": 981.88, "start": 981.84, "text": "need"}, {"end": 982.28, "start": 981.88, "text": "to"}, {"end": 983.04, "start": 982.28, "text": "understand"}, {"end": 983.48, "start": 983.04, "text": "it"}, {"end": 983.76, "start": 983.48, "text": "before"}, {"end": 983.84, "start": 983.76, "text": "you"}, {"end": 984.12, "start": 983.84, "text": "can"}, {"end": 984.68, "start": 984.12, "text": "understand"}, {"end": 984.68, "start": 984.68, "text": "more"}, {"end": 985.16, "start": 984.68, "text": "detailed"}, {"end": 985.48, "start": 985.16, "text": "modern"}, {"end": 986.28, "start": 985.48, "text": "variants,"}, {"end": 986.28, "start": 986.28, "text": "and"}, {"end": 986.32, "start": 986.28, "text": "it"}, {"end": 986.84, "start": 986.32, "text": "clearly"}, {"end": 987.12, "start": 986.84, "text": "is"}, {"end": 987.48, "start": 987.12, "text": "capable"}, {"end": 987.68, "start": 987.48, "text": "of"}, {"end": 988.04, "start": 987.68, "text": "solving"}, {"end": 988.32, "start": 988.04, "text": "some"}, {"end": 988.68, "start": 988.32, "text": "interesting"}, {"end": 989.96, "start": 988.68, "text": "problems."}], "text": " So with this as the image of what those second layer neurons are really doing, you might wonder why I would introduce this network with the motivation of picking up on edges and patterns. I mean, that's just not at all what it ends up doing. Well, this is not meant to be our end goal, but instead a starting point. Frankly, this is old technology, the kind researched in the 80s and 90s. And you do need to understand it before you can understand more detailed modern variants, and it clearly is capable of solving some interesting problems."}, {"chunks": [{"end": 990.28, "start": 990.0, "text": "The"}, {"end": 990.28, "start": 990.28, "text": "more"}, {"end": 990.28, "start": 990.28, "text": "you"}, {"end": 990.64, "start": 990.28, "text": "dig"}, {"end": 991.0, "start": 990.64, "text": "into"}, {"end": 991.4, "start": 991.0, "text": "what"}, {"end": 992.0, "start": 991.4, "text": "those"}, {"end": 992.12, "start": 992.0, "text": "hidden"}, {"end": 992.6, "start": 992.12, "text": "layers"}, {"end": 992.76, "start": 992.6, "text": "are"}, {"end": 992.8, "start": 992.76, "text": "really"}, {"end": 992.8, "start": 992.8, "text": "doing,"}, {"end": 992.92, "start": 992.8, "text": "the"}, {"end": 993.24, "start": 992.92, "text": "less"}, {"end": 993.72, "start": 993.24, "text": "intelligent"}, {"end": 993.84, "start": 993.72, "text": "it"}, {"end": 998.72, "start": 993.84, "text": "seems."}, {"end": 999.12, "start": 998.72, "text": "Shifting"}, {"end": 999.12, "start": 999.12, "text": "the"}, {"end": 999.44, "start": 999.12, "text": "focus"}, {"end": 999.88, "start": 999.44, "text": "for"}, {"end": 999.88, "start": 999.88, "text": "a"}, {"end": 999.96, "start": 999.88, "text": "moment"}, {"end": 1000.12, "start": 999.96, "text": "from"}, {"end": 1000.44, "start": 1000.12, "text": "how"}, {"end": 1000.84, "start": 1000.44, "text": "networks"}, {"end": 1001.12, "start": 1000.84, "text": "learn"}, {"end": 1001.16, "start": 1001.12, "text": "to"}, {"end": 1001.68, "start": 1001.16, "text": "how"}, {"end": 1002.16, "start": 1001.68, "text": "you"}, {"end": 1002.8, "start": 1002.16, "text": "learn?"}, {"end": 1003.16, "start": 1002.8, "text": "That'll"}, {"end": 1003.24, "start": 1003.16, "text": "only"}, {"end": 1003.68, "start": 1003.24, "text": "happen"}, {"end": 1003.68, "start": 1003.68, "text": "if"}, {"end": 1003.8, "start": 1003.68, "text": "you"}, {"end": 1004.28, "start": 1003.8, "text": "engage"}, {"end": 1004.76, "start": 1004.28, "text": "actively"}, {"end": 1004.92, "start": 1004.76, "text": "with"}, {"end": 1005.0, "start": 1004.92, "text": "the"}, {"end": 1005.56, "start": 1005.0, "text": "material"}, {"end": 1005.76, "start": 1005.56, "text": "here"}, {"end": 1007.0, "start": 1005.76, "text": "somehow."}, {"end": 1007.28, "start": 1007.0, "text": "One"}, {"end": 1007.8, "start": 1007.28, "text": "pretty"}, {"end": 1008.36, "start": 1007.8, "text": "simple"}, {"end": 1009.08, "start": 1008.36, "text": "thing"}, {"end": 1009.52, "start": 1009.08, "text": "that"}, {"end": 1009.52, "start": 1009.52, "text": "I"}, {"end": 1009.52, "start": 1009.52, "text": "want"}, {"end": 1009.52, "start": 1009.52, "text": "you"}, {"end": 1009.52, "start": 1009.52, "text": "to"}, {"end": 1009.52, "start": 1009.52, "text": "do"}, {"end": 1009.52, "start": 1009.52, "text": "is"}, {"end": 1009.6, "start": 1009.52, "text": "just"}, {"end": 1010.24, "start": 1009.6, "text": "pause"}, {"end": 1010.44, "start": 1010.24, "text": "right"}, {"end": 1011.12, "start": 1010.44, "text": "now"}, {"end": 1011.44, "start": 1011.12, "text": "and"}, {"end": 1011.64, "start": 1011.44, "text": "think"}, {"end": 1012.04, "start": 1011.64, "text": "deeply"}, {"end": 1012.44, "start": 1012.04, "text": "for"}, {"end": 1012.52, "start": 1012.44, "text": "a"}, {"end": 1012.64, "start": 1012.52, "text": "moment"}, {"end": 1013.08, "start": 1012.64, "text": "about"}, {"end": 1013.4, "start": 1013.08, "text": "what"}, {"end": 1014.24, "start": 1013.4, "text": "changes"}, {"end": 1014.48, "start": 1014.24, "text": "you"}, {"end": 1014.48, "start": 1014.48, "text": "might"}, {"end": 1014.48, "start": 1014.48, "text": "make"}, {"end": 1014.48, "start": 1014.48, "text": "to"}, {"end": 1014.96, "start": 1014.48, "text": "this"}, {"end": 1015.6, "start": 1014.96, "text": "system"}, {"end": 1015.76, "start": 1015.6, "text": "and"}, {"end": 1016.08, "start": 1015.76, "text": "how"}, {"end": 1016.12, "start": 1016.08, "text": "it"}, {"end": 1016.72, "start": 1016.12, "text": "perceives"}, {"end": 1017.2, "start": 1016.72, "text": "images"}, {"end": 1017.4, "start": 1017.2, "text": "if"}, {"end": 1017.56, "start": 1017.4, "text": "you"}, {"end": 1018.16, "start": 1017.56, "text": "wanted"}, {"end": 1018.16, "start": 1018.16, "text": "it"}, {"end": 1018.36, "start": 1018.16, "text": "to"}, {"end": 1018.76, "start": 1018.36, "text": "better"}, {"end": 1018.84, "start": 1018.76, "text": "pick"}, {"end": 1018.96, "start": 1018.84, "text": "up"}, {"end": 1019.16, "start": 1018.96, "text": "on"}, {"end": 1019.68, "start": 1019.16, "text": "things"}, {"end": 1019.8, "start": 1019.68, "text": "like"}, {"end": 1019.96, "start": 1019.8, "text": "edge"}], "text": " The more you dig into what those hidden layers are really doing, the less intelligent it seems. Shifting the focus for a moment from how networks learn to how you learn? That'll only happen if you engage actively with the material here somehow. One pretty simple thing that I want you to do is just pause right now and think deeply for a moment about what changes you might make to this system and how it perceives images if you wanted it to better pick up on things like edge"}, {"chunks": [{"end": 1020.32, "start": 1020.0, "text": "But"}, {"end": 1020.68, "start": 1020.32, "text": "better"}, {"end": 1020.72, "start": 1020.68, "text": "than"}, {"end": 1021.28, "start": 1020.72, "text": "that,"}, {"end": 1021.8, "start": 1021.28, "text": "to"}, {"end": 1022.72, "start": 1021.8, "text": "actually"}, {"end": 1023.52, "start": 1022.72, "text": "engage"}, {"end": 1023.8, "start": 1023.52, "text": "with"}, {"end": 1024.0, "start": 1023.8, "text": "the"}, {"end": 1024.76, "start": 1024.0, "text": "material,"}, {"end": 1025.12, "start": 1024.76, "text": "I"}, {"end": 1025.6, "start": 1025.12, "text": "highly"}, {"end": 1026.04, "start": 1025.6, "text": "recommend"}, {"end": 1026.2, "start": 1026.04, "text": "the"}, {"end": 1026.44, "start": 1026.2, "text": "book"}, {"end": 1026.6, "start": 1026.44, "text": "by"}, {"end": 1026.96, "start": 1026.6, "text": "Michael"}, {"end": 1027.48, "start": 1026.96, "text": "Nielsen"}, {"end": 1027.68, "start": 1027.48, "text": "on"}, {"end": 1027.72, "start": 1027.68, "text": "Deep"}, {"end": 1027.92, "start": 1027.72, "text": "Learning"}, {"end": 1028.2, "start": 1027.92, "text": "and"}, {"end": 1028.68, "start": 1028.2, "text": "Neural"}, {"end": 1029.64, "start": 1028.68, "text": "Networks."}, {"end": 1029.96, "start": 1029.64, "text": "In"}, {"end": 1030.04, "start": 1029.96, "text": "it,"}, {"end": 1030.24, "start": 1030.04, "text": "you"}, {"end": 1030.6, "start": 1030.24, "text": "can"}, {"end": 1030.6, "start": 1030.6, "text": "find"}, {"end": 1030.6, "start": 1030.6, "text": "the"}, {"end": 1030.92, "start": 1030.6, "text": "code"}, {"end": 1031.04, "start": 1030.92, "text": "and"}, {"end": 1031.36, "start": 1031.04, "text": "the"}, {"end": 1031.88, "start": 1031.36, "text": "data"}, {"end": 1032.04, "start": 1031.88, "text": "to"}, {"end": 1032.68, "start": 1032.04, "text": "download"}, {"end": 1033.08, "start": 1032.68, "text": "and"}, {"end": 1033.16, "start": 1033.08, "text": "play"}, {"end": 1033.2, "start": 1033.16, "text": "with"}, {"end": 1033.56, "start": 1033.2, "text": "for"}, {"end": 1033.84, "start": 1033.56, "text": "this"}, {"end": 1034.24, "start": 1033.84, "text": "exact"}, {"end": 1034.72, "start": 1034.24, "text": "example,"}, {"end": 1034.84, "start": 1034.72, "text": "and"}, {"end": 1034.88, "start": 1034.84, "text": "the"}, {"end": 1035.12, "start": 1034.88, "text": "book"}, {"end": 1035.52, "start": 1035.12, "text": "will"}, {"end": 1035.84, "start": 1035.52, "text": "walk"}, {"end": 1036.0, "start": 1035.84, "text": "you"}, {"end": 1036.24, "start": 1036.0, "text": "through"}, {"end": 1036.48, "start": 1036.24, "text": "step"}, {"end": 1036.64, "start": 1036.48, "text": "by"}, {"end": 1037.12, "start": 1036.64, "text": "step"}, {"end": 1037.56, "start": 1037.12, "text": "what"}, {"end": 1037.92, "start": 1037.56, "text": "that"}, {"end": 1038.68, "start": 1037.92, "text": "code"}, {"end": 1038.96, "start": 1038.68, "text": "is"}, {"end": 1039.16, "start": 1038.96, "text": "doing."}, {"end": 1039.52, "start": 1039.16, "text": "What's"}, {"end": 1040.12, "start": 1039.52, "text": "awesome"}, {"end": 1040.36, "start": 1040.12, "text": "is"}, {"end": 1040.36, "start": 1040.36, "text": "that"}, {"end": 1040.4, "start": 1040.36, "text": "this"}, {"end": 1040.56, "start": 1040.4, "text": "book"}, {"end": 1040.76, "start": 1040.56, "text": "is"}, {"end": 1040.76, "start": 1040.76, "text": "free"}, {"end": 1040.76, "start": 1040.76, "text": "and"}, {"end": 1041.32, "start": 1040.76, "text": "publicly"}, {"end": 1042.08, "start": 1041.32, "text": "available,"}, {"end": 1042.28, "start": 1042.08, "text": "so"}, {"end": 1042.32, "start": 1042.28, "text": "if"}, {"end": 1042.4, "start": 1042.32, "text": "you"}, {"end": 1042.6, "start": 1042.4, "text": "do"}, {"end": 1042.72, "start": 1042.6, "text": "get"}, {"end": 1043.04, "start": 1042.72, "text": "something"}, {"end": 1043.04, "start": 1043.04, "text": "out"}, {"end": 1043.12, "start": 1043.04, "text": "of"}, {"end": 1043.4, "start": 1043.12, "text": "it,"}, {"end": 1043.76, "start": 1043.4, "text": "consider"}, {"end": 1044.36, "start": 1043.76, "text": "joining"}, {"end": 1044.6, "start": 1044.36, "text": "me"}, {"end": 1044.6, "start": 1044.6, "text": "in"}, {"end": 1045.6, "start": 1044.6, "text": "making"}, {"end": 1045.92, "start": 1045.6, "text": "a"}, {"end": 1046.44, "start": 1045.92, "text": "donation"}, {"end": 1046.76, "start": 1046.44, "text": "towards"}, {"end": 1047.36, "start": 1046.76, "text": "Nielsen's"}, {"end": 1047.92, "start": 1047.36, "text": "efforts."}, {"end": 1048.36, "start": 1047.92, "text": "I've"}, {"end": 1048.84, "start": 1048.36, "text": "also"}, {"end": 1048.96, "start": 1048.84, "text": "linked"}, {"end": 1048.96, "start": 1048.96, "text": "a"}, {"end": 1049.16, "start": 1048.96, "text": "couple"}, {"end": 1049.2, "start": 1049.16, "text": "other"}, {"end": 1049.96, "start": 1049.2, "text": "resources"}], "text": " But better than that, to actually engage with the material, I highly recommend the book by Michael Nielsen on Deep Learning and Neural Networks. In it, you can find the code and the data to download and play with for this exact example, and the book will walk you through step by step what that code is doing. What's awesome is that this book is free and publicly available, so if you do get something out of it, consider joining me in making a donation towards Nielsen's efforts. I've also linked a couple other resources"}, {"chunks": [{"end": 1050.52, "start": 1050.0, "text": "that"}, {"end": 1050.64, "start": 1050.52, "text": "I"}, {"end": 1050.84, "start": 1050.64, "text": "like"}, {"end": 1050.88, "start": 1050.84, "text": "a"}, {"end": 1051.04, "start": 1050.88, "text": "lot"}, {"end": 1051.04, "start": 1051.04, "text": "in"}, {"end": 1051.12, "start": 1051.04, "text": "the"}, {"end": 1051.68, "start": 1051.12, "text": "description,"}, {"end": 1052.28, "start": 1051.68, "text": "including"}, {"end": 1052.36, "start": 1052.28, "text": "the"}, {"end": 1052.88, "start": 1052.36, "text": "phenomenal"}, {"end": 1053.16, "start": 1052.88, "text": "and"}, {"end": 1053.52, "start": 1053.16, "text": "beautiful"}, {"end": 1053.8, "start": 1053.52, "text": "blog"}, {"end": 1054.16, "start": 1053.8, "text": "posts"}, {"end": 1054.28, "start": 1054.16, "text": "by"}, {"end": 1054.6, "start": 1054.28, "text": "Chris"}, {"end": 1054.92, "start": 1054.6, "text": "Ola"}, {"end": 1055.0, "start": 1054.92, "text": "and"}, {"end": 1055.4, "start": 1055.0, "text": "the"}, {"end": 1055.96, "start": 1055.4, "text": "articles"}, {"end": 1056.04, "start": 1055.96, "text": "in"}, {"end": 1056.56, "start": 1056.04, "text": "Distil."}, {"end": 1056.92, "start": 1056.56, "text": "To"}, {"end": 1057.48, "start": 1056.92, "text": "close"}, {"end": 1057.96, "start": 1057.48, "text": "things"}, {"end": 1058.64, "start": 1057.96, "text": "off"}, {"end": 1059.08, "start": 1058.64, "text": "here"}, {"end": 1059.32, "start": 1059.08, "text": "for"}, {"end": 1059.64, "start": 1059.32, "text": "the"}, {"end": 1059.88, "start": 1059.64, "text": "last"}, {"end": 1060.08, "start": 1059.88, "text": "few"}, {"end": 1060.56, "start": 1060.08, "text": "minutes,"}, {"end": 1060.8, "start": 1060.56, "text": "I"}, {"end": 1061.0, "start": 1060.8, "text": "want"}, {"end": 1061.0, "start": 1061.0, "text": "to"}, {"end": 1061.08, "start": 1061.0, "text": "jump"}, {"end": 1061.36, "start": 1061.08, "text": "back"}, {"end": 1061.96, "start": 1061.36, "text": "into"}, {"end": 1062.36, "start": 1061.96, "text": "a"}, {"end": 1062.56, "start": 1062.36, "text": "snippet"}, {"end": 1063.0, "start": 1062.56, "text": "of"}, {"end": 1063.36, "start": 1063.0, "text": "the"}, {"end": 1063.88, "start": 1063.36, "text": "interview"}, {"end": 1063.96, "start": 1063.88, "text": "that"}, {"end": 1064.08, "start": 1063.96, "text": "I"}, {"end": 1064.12, "start": 1064.08, "text": "had"}, {"end": 1064.12, "start": 1064.12, "text": "with"}, {"end": 1064.12, "start": 1064.12, "text": "Leisha"}, {"end": 1064.12, "start": 1064.12, "text": "Lee."}, {"end": 1064.12, "start": 1064.12, "text": "You"}, {"end": 1064.12, "start": 1064.12, "text": "might"}, {"end": 1064.12, "start": 1064.12, "text": "remember"}, {"end": 1064.28, "start": 1064.12, "text": "her"}, {"end": 1064.76, "start": 1064.28, "text": "from"}, {"end": 1065.24, "start": 1064.76, "text": "the"}, {"end": 1065.6, "start": 1065.24, "text": "last"}, {"end": 1065.88, "start": 1065.6, "text": "video,"}, {"end": 1066.04, "start": 1065.88, "text": "she"}, {"end": 1066.04, "start": 1066.04, "text": "did"}, {"end": 1066.32, "start": 1066.04, "text": "her"}, {"end": 1066.72, "start": 1066.32, "text": "PhD"}, {"end": 1067.2, "start": 1066.72, "text": "work"}, {"end": 1067.32, "start": 1067.2, "text": "in"}, {"end": 1067.32, "start": 1067.32, "text": "deep"}, {"end": 1067.84, "start": 1067.32, "text": "learning."}, {"end": 1067.88, "start": 1067.84, "text": "And"}, {"end": 1067.88, "start": 1067.88, "text": "in"}, {"end": 1068.24, "start": 1067.88, "text": "this"}, {"end": 1068.88, "start": 1068.24, "text": "little"}, {"end": 1069.24, "start": 1068.88, "text": "snippet,"}, {"end": 1069.36, "start": 1069.24, "text": "she"}, {"end": 1069.68, "start": 1069.36, "text": "talks"}, {"end": 1069.92, "start": 1069.68, "text": "about"}, {"end": 1070.16, "start": 1069.92, "text": "two"}, {"end": 1070.44, "start": 1070.16, "text": "recent"}, {"end": 1070.84, "start": 1070.44, "text": "papers"}, {"end": 1070.84, "start": 1070.84, "text": "that"}, {"end": 1071.04, "start": 1070.84, "text": "really"}, {"end": 1071.32, "start": 1071.04, "text": "dig"}, {"end": 1071.68, "start": 1071.32, "text": "into"}, {"end": 1072.12, "start": 1071.68, "text": "how"}, {"end": 1072.32, "start": 1072.12, "text": "some"}, {"end": 1072.32, "start": 1072.32, "text": "of"}, {"end": 1072.44, "start": 1072.32, "text": "the"}, {"end": 1072.68, "start": 1072.44, "text": "more"}, {"end": 1073.12, "start": 1072.68, "text": "modern"}, {"end": 1073.48, "start": 1073.12, "text": "image"}, {"end": 1074.04, "start": 1073.48, "text": "recognition"}, {"end": 1074.56, "start": 1074.04, "text": "networks"}, {"end": 1074.84, "start": 1074.56, "text": "are"}, {"end": 1075.4, "start": 1074.84, "text": "actually"}, {"end": 1076.0, "start": 1075.4, "text": "learning."}, {"end": 1076.32, "start": 1076.0, "text": "Just"}, {"end": 1076.32, "start": 1076.32, "text": "to"}, {"end": 1076.72, "start": 1076.32, "text": "set"}, {"end": 1076.72, "start": 1076.72, "text": "up"}, {"end": 1076.96, "start": 1076.72, "text": "where"}, {"end": 1076.96, "start": 1076.96, "text": "we"}, {"end": 1077.28, "start": 1076.96, "text": "were"}, {"end": 1077.28, "start": 1077.28, "text": "in"}, {"end": 1077.48, "start": 1077.28, "text": "the"}, {"end": 1078.12, "start": 1077.48, "text": "conversation,"}, {"end": 1078.28, "start": 1078.12, "text": "the"}, {"end": 1078.52, "start": 1078.28, "text": "first"}, {"end": 1078.96, "start": 1078.52, "text": "paper"}, {"end": 1079.16, "start": 1078.96, "text": "took"}, {"end": 1079.64, "start": 1079.16, "text": "one"}, {"end": 1079.76, "start": 1079.64, "text": "of"}, {"end": 1079.96, "start": 1079.76, "text": "these"}], "text": " that I like a lot in the description, including the phenomenal and beautiful blog posts by Chris Ola and the articles in Distil. To close things off here for the last few minutes, I want to jump back into a snippet of the interview that I had with Leisha Lee. You might remember her from the last video, she did her PhD work in deep learning. And in this little snippet, she talks about two recent papers that really dig into how some of the more modern image recognition networks are actually learning. Just to set up where we were in the conversation, the first paper took one of these"}, {"chunks": [{"end": 1080.12, "start": 1080.0, "text": "It's"}, {"end": 1080.4, "start": 1080.12, "text": "a"}, {"end": 1080.64, "start": 1080.4, "text": "particularly"}, {"end": 1080.68, "start": 1080.64, "text": "deep"}, {"end": 1081.0, "start": 1080.68, "text": "neural"}, {"end": 1081.52, "start": 1081.0, "text": "network"}, {"end": 1082.04, "start": 1081.52, "text": "that's"}, {"end": 1082.52, "start": 1082.04, "text": "really"}, {"end": 1082.6, "start": 1082.52, "text": "good"}, {"end": 1082.8, "start": 1082.6, "text": "at"}, {"end": 1083.12, "start": 1082.8, "text": "image"}, {"end": 1083.52, "start": 1083.12, "text": "recognition,"}, {"end": 1083.72, "start": 1083.52, "text": "and"}, {"end": 1084.04, "start": 1083.72, "text": "instead"}, {"end": 1084.12, "start": 1084.04, "text": "of"}, {"end": 1084.76, "start": 1084.12, "text": "training"}, {"end": 1084.88, "start": 1084.76, "text": "it"}, {"end": 1084.88, "start": 1084.88, "text": "on"}, {"end": 1084.88, "start": 1084.88, "text": "a"}, {"end": 1085.36, "start": 1084.88, "text": "properly"}, {"end": 1085.68, "start": 1085.36, "text": "labeled"}, {"end": 1086.32, "start": 1085.68, "text": "dataset,"}, {"end": 1086.64, "start": 1086.32, "text": "it"}, {"end": 1087.04, "start": 1086.64, "text": "shuffled"}, {"end": 1087.24, "start": 1087.04, "text": "all"}, {"end": 1087.24, "start": 1087.24, "text": "of"}, {"end": 1087.28, "start": 1087.24, "text": "the"}, {"end": 1087.76, "start": 1087.28, "text": "labels"}, {"end": 1088.0, "start": 1087.76, "text": "around"}, {"end": 1088.36, "start": 1088.0, "text": "before"}, {"end": 1089.44, "start": 1088.36, "text": "training."}, {"end": 1089.84, "start": 1089.44, "text": "Obviously,"}, {"end": 1089.96, "start": 1089.84, "text": "the"}, {"end": 1090.32, "start": 1089.96, "text": "testing"}, {"end": 1090.84, "start": 1090.32, "text": "accuracy"}, {"end": 1091.08, "start": 1090.84, "text": "here"}, {"end": 1091.4, "start": 1091.08, "text": "was"}, {"end": 1091.56, "start": 1091.4, "text": "going"}, {"end": 1091.6, "start": 1091.56, "text": "to"}, {"end": 1091.72, "start": 1091.6, "text": "be"}, {"end": 1091.88, "start": 1091.72, "text": "no"}, {"end": 1092.2, "start": 1091.88, "text": "better"}, {"end": 1092.48, "start": 1092.2, "text": "than"}, {"end": 1092.76, "start": 1092.48, "text": "random,"}, {"end": 1093.2, "start": 1092.76, "text": "since"}, {"end": 1093.64, "start": 1093.2, "text": "everything's"}, {"end": 1093.88, "start": 1093.64, "text": "just"}, {"end": 1094.48, "start": 1093.88, "text": "randomly"}, {"end": 1094.8, "start": 1094.48, "text": "labeled."}, {"end": 1095.04, "start": 1094.8, "text": "But"}, {"end": 1095.2, "start": 1095.04, "text": "it"}, {"end": 1095.84, "start": 1095.2, "text": "was"}, {"end": 1096.2, "start": 1095.84, "text": "still"}, {"end": 1096.32, "start": 1096.2, "text": "able"}, {"end": 1096.44, "start": 1096.32, "text": "to"}, {"end": 1096.88, "start": 1096.44, "text": "achieve"}, {"end": 1096.96, "start": 1096.88, "text": "the"}, {"end": 1097.28, "start": 1096.96, "text": "same"}, {"end": 1098.0, "start": 1097.28, "text": "training"}, {"end": 1098.64, "start": 1098.0, "text": "accuracy"}, {"end": 1098.68, "start": 1098.64, "text": "as"}, {"end": 1098.92, "start": 1098.68, "text": "you"}, {"end": 1099.44, "start": 1098.92, "text": "would"}, {"end": 1099.68, "start": 1099.44, "text": "on"}, {"end": 1099.68, "start": 1099.68, "text": "a"}, {"end": 1100.08, "start": 1099.68, "text": "properly"}, {"end": 1100.4, "start": 1100.08, "text": "labeled"}, {"end": 1100.92, "start": 1100.4, "text": "dataset."}, {"end": 1102.08, "start": 1100.92, "text": "Basically,"}, {"end": 1102.12, "start": 1102.08, "text": "the"}, {"end": 1102.4, "start": 1102.12, "text": "millions"}, {"end": 1102.44, "start": 1102.4, "text": "of"}, {"end": 1103.12, "start": 1102.44, "text": "weights"}, {"end": 1103.4, "start": 1103.12, "text": "for"}, {"end": 1103.8, "start": 1103.4, "text": "this"}, {"end": 1104.32, "start": 1103.8, "text": "particular"}, {"end": 1104.84, "start": 1104.32, "text": "network"}, {"end": 1105.08, "start": 1104.84, "text": "were"}, {"end": 1105.52, "start": 1105.08, "text": "enough"}, {"end": 1105.68, "start": 1105.52, "text": "for"}, {"end": 1105.68, "start": 1105.68, "text": "it"}, {"end": 1105.8, "start": 1105.68, "text": "to"}, {"end": 1106.16, "start": 1105.8, "text": "just"}, {"end": 1106.8, "start": 1106.16, "text": "memorize"}, {"end": 1106.96, "start": 1106.8, "text": "the"}, {"end": 1107.32, "start": 1106.96, "text": "random"}, {"end": 1107.8, "start": 1107.32, "text": "data,"}, {"end": 1108.32, "start": 1107.8, "text": "which"}, {"end": 1108.88, "start": 1108.32, "text": "kind"}, {"end": 1108.92, "start": 1108.88, "text": "of"}, {"end": 1109.0, "start": 1108.92, "text": "raises"}, {"end": 1109.0, "start": 1109.0, "text": "the"}, {"end": 1109.48, "start": 1109.0, "text": "question"}, {"end": 1109.64, "start": 1109.48, "text": "for"}, {"end": 1109.96, "start": 1109.64, "text": "whether"}], "text": " It's a particularly deep neural network that's really good at image recognition, and instead of training it on a properly labeled dataset, it shuffled all of the labels around before training. Obviously, the testing accuracy here was going to be no better than random, since everything's just randomly labeled. But it was still able to achieve the same training accuracy as you would on a properly labeled dataset. Basically, the millions of weights for this particular network were enough for it to just memorize the random data, which kind of raises the question for whether"}, {"chunks": [{"end": 1111.0, "start": 1110.0, "text": "Minimizing"}, {"end": 1111.84, "start": 1111.0, "text": "this"}, {"end": 1112.76, "start": 1111.84, "text": "cost"}, {"end": 1114.92, "start": 1112.76, "text": "function"}, {"end": 1116.64, "start": 1114.92, "text": "actually"}, {"end": 1118.52, "start": 1116.64, "text": "corresponds"}, {"end": 1119.28, "start": 1118.52, "text": "to"}, {"end": 1120.2, "start": 1119.28, "text": "any"}, {"end": 1121.96, "start": 1120.2, "text": "sort"}, {"end": 1123.44, "start": 1121.96, "text": "of"}, {"end": 1126.96, "start": 1123.44, "text": "structure"}, {"end": 1127.44, "start": 1126.96, "text": "in"}, {"end": 1128.2, "start": 1127.44, "text": "the"}, {"end": 1128.68, "start": 1128.2, "text": "image,"}, {"end": 1129.56, "start": 1128.68, "text": "or"}, {"end": 1130.2, "start": 1129.56, "text": "is"}, {"end": 1131.48, "start": 1130.2, "text": "it"}, {"end": 1133.64, "start": 1131.48, "text": "just,"}, {"end": 1134.88, "start": 1133.64, "text": "you"}, {"end": 1136.88, "start": 1134.88, "text": "know,"}, {"end": 1139.96, "start": 1136.88, "text": "memorization?"}], "text": " Minimizing this cost function actually corresponds to any sort of structure in the image, or is it just, you know, memorization?"}, {"chunks": [{"end": 1140.56, "start": 1140.0, "text": "went"}, {"end": 1141.04, "start": 1140.56, "text": "down"}, {"end": 1141.72, "start": 1141.04, "text": "very,"}, {"end": 1142.08, "start": 1141.72, "text": "you"}, {"end": 1142.48, "start": 1142.08, "text": "know,"}, {"end": 1142.56, "start": 1142.48, "text": "very"}, {"end": 1142.6, "start": 1142.56, "text": "slowly"}, {"end": 1142.76, "start": 1142.6, "text": "in"}, {"end": 1143.32, "start": 1142.76, "text": "almost"}, {"end": 1143.88, "start": 1143.32, "text": "kind"}, {"end": 1144.28, "start": 1143.88, "text": "of"}, {"end": 1144.56, "start": 1144.28, "text": "a"}, {"end": 1144.92, "start": 1144.56, "text": "linear"}, {"end": 1145.48, "start": 1144.92, "text": "fashion."}, {"end": 1146.08, "start": 1145.48, "text": "So"}, {"end": 1146.16, "start": 1146.08, "text": "you're"}, {"end": 1146.36, "start": 1146.16, "text": "really"}, {"end": 1147.08, "start": 1146.36, "text": "struggling"}, {"end": 1147.12, "start": 1147.08, "text": "to"}, {"end": 1147.68, "start": 1147.12, "text": "find"}, {"end": 1147.68, "start": 1147.68, "text": "that"}, {"end": 1147.68, "start": 1147.68, "text": "local"}, {"end": 1147.68, "start": 1147.68, "text": "minima"}, {"end": 1147.68, "start": 1147.68, "text": "of"}, {"end": 1147.68, "start": 1147.68, "text": "possible,"}, {"end": 1147.68, "start": 1147.68, "text": "you"}, {"end": 1147.68, "start": 1147.68, "text": "know,"}, {"end": 1147.72, "start": 1147.68, "text": "the"}, {"end": 1148.92, "start": 1147.72, "text": "right"}, {"end": 1148.92, "start": 1148.92, "text": "weights"}, {"end": 1148.92, "start": 1148.92, "text": "that"}, {"end": 1148.92, "start": 1148.92, "text": "would"}, {"end": 1148.92, "start": 1148.92, "text": "get"}, {"end": 1148.92, "start": 1148.92, "text": "you"}, {"end": 1149.12, "start": 1148.92, "text": "that"}, {"end": 1149.44, "start": 1149.12, "text": "accuracy."}, {"end": 1150.0, "start": 1149.44, "text": "Whereas"}, {"end": 1150.0, "start": 1150.0, "text": "if"}, {"end": 1150.56, "start": 1150.0, "text": "you're"}, {"end": 1150.76, "start": 1150.56, "text": "actually"}, {"end": 1151.24, "start": 1150.76, "text": "training"}, {"end": 1151.56, "start": 1151.24, "text": "on"}, {"end": 1151.92, "start": 1151.56, "text": "a"}, {"end": 1153.04, "start": 1151.92, "text": "structured"}, {"end": 1153.64, "start": 1153.04, "text": "data"}, {"end": 1154.12, "start": 1153.64, "text": "set,"}, {"end": 1154.36, "start": 1154.12, "text": "one"}, {"end": 1154.76, "start": 1154.36, "text": "that"}, {"end": 1155.16, "start": 1154.76, "text": "has"}, {"end": 1155.36, "start": 1155.16, "text": "the"}, {"end": 1155.76, "start": 1155.36, "text": "right"}, {"end": 1156.12, "start": 1155.76, "text": "labels,"}, {"end": 1156.44, "start": 1156.12, "text": "you"}, {"end": 1156.52, "start": 1156.44, "text": "know,"}, {"end": 1156.56, "start": 1156.52, "text": "you"}, {"end": 1157.04, "start": 1156.56, "text": "fiddle"}, {"end": 1158.24, "start": 1157.04, "text": "around"}, {"end": 1158.48, "start": 1158.24, "text": "a"}, {"end": 1158.48, "start": 1158.48, "text": "little"}, {"end": 1158.48, "start": 1158.48, "text": "bit"}, {"end": 1158.48, "start": 1158.48, "text": "in"}, {"end": 1158.56, "start": 1158.48, "text": "the"}, {"end": 1158.6, "start": 1158.56, "text": "beginning,"}, {"end": 1158.8, "start": 1158.6, "text": "but"}, {"end": 1159.08, "start": 1158.8, "text": "then"}, {"end": 1159.08, "start": 1159.08, "text": "you"}, {"end": 1159.08, "start": 1159.08, "text": "kind"}, {"end": 1159.12, "start": 1159.08, "text": "of"}, {"end": 1159.72, "start": 1159.12, "text": "dropped"}, {"end": 1160.16, "start": 1159.72, "text": "very"}, {"end": 1161.12, "start": 1160.16, "text": "fast"}, {"end": 1161.6, "start": 1161.12, "text": "to"}, {"end": 1161.84, "start": 1161.6, "text": "get"}, {"end": 1161.84, "start": 1161.84, "text": "to"}, {"end": 1161.88, "start": 1161.84, "text": "that"}, {"end": 1162.8, "start": 1161.88, "text": "accuracy"}, {"end": 1163.08, "start": 1162.8, "text": "level."}, {"end": 1163.72, "start": 1163.08, "text": "And"}, {"end": 1164.36, "start": 1163.72, "text": "so"}, {"end": 1164.56, "start": 1164.36, "text": "in"}, {"end": 1164.56, "start": 1164.56, "text": "some"}, {"end": 1164.56, "start": 1164.56, "text": "sense,"}, {"end": 1164.6, "start": 1164.56, "text": "it"}, {"end": 1164.64, "start": 1164.6, "text": "was"}, {"end": 1164.64, "start": 1164.64, "text": "easier"}, {"end": 1165.2, "start": 1164.64, "text": "to"}, {"end": 1165.92, "start": 1165.2, "text": "find"}, {"end": 1166.8, "start": 1165.92, "text": "that"}, {"end": 1167.44, "start": 1166.8, "text": "local"}, {"end": 1168.04, "start": 1167.44, "text": "maxima."}, {"end": 1168.32, "start": 1168.04, "text": "And"}, {"end": 1168.32, "start": 1168.32, "text": "so"}, {"end": 1168.36, "start": 1168.32, "text": "it"}, {"end": 1168.88, "start": 1168.36, "text": "was"}, {"end": 1169.16, "start": 1168.88, "text": "also"}, {"end": 1169.6, "start": 1169.16, "text": "interesting"}, {"end": 1169.76, "start": 1169.6, "text": "about"}, {"end": 1169.96, "start": 1169.76, "text": "that"}], "text": " went down very, you know, very slowly in almost kind of a linear fashion. So you're really struggling to find that local minima of possible, you know, the right weights that would get you that accuracy. Whereas if you're actually training on a structured data set, one that has the right labels, you know, you fiddle around a little bit in the beginning, but then you kind of dropped very fast to get to that accuracy level. And so in some sense, it was easier to find that local maxima. And so it was also interesting about that"}, {"chunks": [{"end": 1170.32, "start": 1170.0, "text": "as"}, {"end": 1170.32, "start": 1170.32, "text": "it"}, {"end": 1170.32, "start": 1170.32, "text": "brings"}, {"end": 1170.32, "start": 1170.32, "text": "into"}, {"end": 1170.64, "start": 1170.32, "text": "light"}, {"end": 1171.16, "start": 1170.64, "text": "another"}, {"end": 1171.56, "start": 1171.16, "text": "paper"}, {"end": 1173.0, "start": 1171.56, "text": "from"}, {"end": 1173.96, "start": 1173.0, "text": "actually"}, {"end": 1174.48, "start": 1173.96, "text": "a"}, {"end": 1175.08, "start": 1174.48, "text": "couple"}, {"end": 1175.16, "start": 1175.08, "text": "of"}, {"end": 1175.24, "start": 1175.16, "text": "years"}, {"end": 1175.88, "start": 1175.24, "text": "ago,"}, {"end": 1176.16, "start": 1175.88, "text": "which"}, {"end": 1176.88, "start": 1176.16, "text": "has"}, {"end": 1177.04, "start": 1176.88, "text": "a"}, {"end": 1177.04, "start": 1177.04, "text": "lot"}, {"end": 1177.08, "start": 1177.04, "text": "more"}, {"end": 1177.96, "start": 1177.08, "text": "simplifications"}, {"end": 1178.4, "start": 1177.96, "text": "about"}, {"end": 1178.4, "start": 1178.4, "text": "the"}, {"end": 1178.4, "start": 1178.4, "text": "network"}, {"end": 1178.4, "start": 1178.4, "text": "layers."}, {"end": 1178.48, "start": 1178.4, "text": "But"}, {"end": 1179.0, "start": 1178.48, "text": "one"}, {"end": 1179.24, "start": 1179.0, "text": "of"}, {"end": 1179.64, "start": 1179.24, "text": "the"}, {"end": 1180.04, "start": 1179.64, "text": "results"}, {"end": 1180.24, "start": 1180.04, "text": "was"}, {"end": 1180.64, "start": 1180.24, "text": "saying"}, {"end": 1180.64, "start": 1180.64, "text": "how"}, {"end": 1180.64, "start": 1180.64, "text": "if"}, {"end": 1180.64, "start": 1180.64, "text": "you"}, {"end": 1180.64, "start": 1180.64, "text": "look"}, {"end": 1180.68, "start": 1180.64, "text": "at"}, {"end": 1181.48, "start": 1180.68, "text": "the"}, {"end": 1182.48, "start": 1181.48, "text": "optimization"}, {"end": 1183.0, "start": 1182.48, "text": "landscape,"}, {"end": 1183.32, "start": 1183.0, "text": "the"}, {"end": 1183.8, "start": 1183.32, "text": "local"}, {"end": 1183.84, "start": 1183.8, "text": "minima"}, {"end": 1185.08, "start": 1183.84, "text": "that"}, {"end": 1185.16, "start": 1185.08, "text": "these"}, {"end": 1186.04, "start": 1185.16, "text": "networks"}, {"end": 1186.68, "start": 1186.04, "text": "tend"}, {"end": 1187.08, "start": 1186.68, "text": "to"}, {"end": 1187.12, "start": 1187.08, "text": "learn"}, {"end": 1187.48, "start": 1187.12, "text": "are"}, {"end": 1188.16, "start": 1187.48, "text": "actually"}, {"end": 1188.28, "start": 1188.16, "text": "of"}, {"end": 1188.28, "start": 1188.28, "text": "equal"}, {"end": 1188.68, "start": 1188.28, "text": "quality."}, {"end": 1189.04, "start": 1188.68, "text": "So"}, {"end": 1189.52, "start": 1189.04, "text": "in"}, {"end": 1189.64, "start": 1189.52, "text": "some"}, {"end": 1190.08, "start": 1189.64, "text": "sense,"}, {"end": 1190.48, "start": 1190.08, "text": "if"}, {"end": 1190.96, "start": 1190.48, "text": "your"}, {"end": 1190.96, "start": 1190.96, "text": "data"}, {"end": 1191.0, "start": 1190.96, "text": "set"}, {"end": 1191.4, "start": 1191.0, "text": "is"}, {"end": 1192.28, "start": 1191.4, "text": "structured,"}, {"end": 1192.56, "start": 1192.28, "text": "you"}, {"end": 1192.72, "start": 1192.56, "text": "should"}, {"end": 1192.76, "start": 1192.72, "text": "be"}, {"end": 1192.76, "start": 1192.76, "text": "able"}, {"end": 1192.76, "start": 1192.76, "text": "to"}, {"end": 1193.16, "start": 1192.76, "text": "find"}, {"end": 1193.48, "start": 1193.16, "text": "that"}, {"end": 1194.04, "start": 1193.48, "text": "much"}, {"end": 1194.64, "start": 1194.04, "text": "more"}, {"end": 1194.92, "start": 1194.64, "text": "easily."}, {"end": 1195.44, "start": 1194.92, "text": "My"}, {"end": 1198.72, "start": 1195.44, "text": "thanks,"}, {"end": 1199.24, "start": 1198.72, "text": "as"}, {"end": 1199.64, "start": 1199.24, "text": "always,"}, {"end": 1199.64, "start": 1199.64, "text": "to"}, {"end": 1199.96, "start": 1199.64, "text": "those"}], "text": " as it brings into light another paper from actually a couple of years ago, which has a lot more simplifications about the network layers. But one of the results was saying how if you look at the optimization landscape, the local minima that these networks tend to learn are actually of equal quality. So in some sense, if your data set is structured, you should be able to find that much more easily. My thanks, as always, to those"}, {"chunks": [{"end": 1200.8, "start": 1200.0, "text": "of"}, {"end": 1201.6, "start": 1200.8, "text": "you"}, {"end": 1202.4, "start": 1201.6, "text": "supporting"}, {"end": 1202.68, "start": 1202.4, "text": "on"}, {"end": 1203.56, "start": 1202.68, "text": "Patreon."}, {"end": 1204.48, "start": 1203.56, "text": "I've"}, {"end": 1205.08, "start": 1204.48, "text": "said"}, {"end": 1205.76, "start": 1205.08, "text": "before"}, {"end": 1206.08, "start": 1205.76, "text": "just"}, {"end": 1206.4, "start": 1206.08, "text": "what"}, {"end": 1206.52, "start": 1206.4, "text": "a"}, {"end": 1207.12, "start": 1206.52, "text": "game"}, {"end": 1207.48, "start": 1207.12, "text": "changer"}, {"end": 1208.08, "start": 1207.48, "text": "Patreon"}, {"end": 1208.48, "start": 1208.08, "text": "is,"}, {"end": 1208.88, "start": 1208.48, "text": "but"}, {"end": 1209.08, "start": 1208.88, "text": "these"}, {"end": 1209.56, "start": 1209.08, "text": "videos"}, {"end": 1210.28, "start": 1209.56, "text": "really"}, {"end": 1210.92, "start": 1210.28, "text": "would"}, {"end": 1211.4, "start": 1210.92, "text": "not"}, {"end": 1211.8, "start": 1211.4, "text": "be"}, {"end": 1212.56, "start": 1211.8, "text": "possible"}, {"end": 1213.36, "start": 1212.56, "text": "without"}, {"end": 1213.68, "start": 1213.36, "text": "you."}, {"end": 1214.08, "start": 1213.68, "text": "I"}, {"end": 1214.96, "start": 1214.08, "text": "also"}, {"end": 1215.32, "start": 1214.96, "text": "want"}, {"end": 1215.8, "start": 1215.32, "text": "to"}, {"end": 1216.28, "start": 1215.8, "text": "give"}, {"end": 1216.48, "start": 1216.28, "text": "a"}, {"end": 1217.04, "start": 1216.48, "text": "special"}, {"end": 1218.32, "start": 1217.04, "text": "thanks"}, {"end": 1218.76, "start": 1218.32, "text": "to"}, {"end": 1219.36, "start": 1218.76, "text": "the"}, {"end": 1219.48, "start": 1219.36, "text": "VC"}, {"end": 1220.0, "start": 1219.48, "text": "firm"}, {"end": 1220.44, "start": 1220.0, "text": "Amplify"}, {"end": 1220.96, "start": 1220.44, "text": "Partners"}, {"end": 1221.24, "start": 1220.96, "text": "in"}, {"end": 1221.88, "start": 1221.24, "text": "their"}, {"end": 1223.44, "start": 1221.88, "text": "support"}, {"end": 1223.92, "start": 1223.44, "text": "of"}, {"end": 1223.96, "start": 1223.92, "text": "these"}, {"end": 1224.44, "start": 1223.96, "text": "initial"}, {"end": 1227.64, "start": 1224.44, "text": "videos"}, {"end": 1228.8, "start": 1227.64, "text": "in"}, {"end": 1229.52, "start": 1228.8, "text": "the"}, {"end": 1229.96, "start": 1229.52, "text": "series."}], "text": " of you supporting on Patreon. I've said before just what a game changer Patreon is, but these videos really would not be possible without you. I also want to give a special thanks to the VC firm Amplify Partners in their support of these initial videos in the series."}, {"chunks": [{"end": 1232.92, "start": 1230.0, "text": "you"}], "text": " you"}]}}