The Learning Revolution (Decoding the Language Machine Ep. 3)

How do you program a computer to do something you can’t even explain in words? Discover how backpropagation revolutionized AI and allowed machines to learn on their own.

For seventy years, the world’s smartest engineers tried to teach machines how to read human handwriting using rigid rules, and they failed. Then came the connectionist movement: an audacious shift from writing code to learning from data.

In this episode of “Decoding the Language Machine,” we demystify the core science of how neural networks actually learn. We track the journey from Frank Rosenblatt’s 1957 mechanical Perceptron, through the mathematical “wall” of the XOR problem, to the realization of modern deep learning. You’ll see how the landscape of loss works, how backpropagation uses calculus to correct mistakes, and how AI independently rediscovers the biological statistical structures of the mammalian brain.

Test these neural networks right in your browser and watch them learn in real time: • Play with the XOR Problem in the Google TensorFlow Playground:playground.tensorflow.org • Watch an MNIST Neural Network train live via Andrej Karpathy's ConvNetJS:cs.stanford.edu/people/karpathy/convnetjs