Sara Beery teaches the second lecture of MIT's 6.7960 Deep Learning course, covering the mechanics of training a neural network. She works through stochastic gradient descent, backpropagation, and automatic differentiation, framing all three as pieces of a broader differentiable programming approach. The lecture builds from the basic optimization problem of minimizing a loss function through the chain rule computations that make backpropagation efficient, then connects those computations to how modern frameworks implement automatic differentiation under the hood. Running eighty minutes, it is a full classroom session rather than a highlight reel, aimed at students who already have the setup from lecture one and need the computational engine that makes deep learning trainable at scale.