How neural networks learn: multi-layer perceptrons, backpropagation, modern optimisers, and building models in PyTorch.
Takeaway 1. A neural network is layers of weighted sums glued together by nonlinear activations — nothing more mysterious than that.
Takeaway 2. Training means following the gradient of the loss downhill. It works well, but it is not guaranteed: initialization, learning rate, and plateaus all matter.
Takeaway 3. Capacity cuts both ways. Count your parameters, watch a validation set, and regularize with early stopping, weight decay, and dropout.
Takeaway 4. Choose the architecture that matches the structure of your data — and in chemistry, build the symmetry in.