Deep Learning: Core & Tracks

Deep learning is what happens when you stack differentiable building blocks and let optimization do the feature engineering for you. Instead of hand-crafting what matters in the data, you define an architecture, a loss, and a gradient, and the representations emerge from training. This pillar builds the core mechanics first, then branches into the major application tracks.

You will start with the core building blocks (layers, activations, backpropagation, normalization, and regularization), then work through computer vision (CNNs from LeNet to ConvNeXt, detection, and neural rendering), sequence modeling (recurrent networks, attention, and state space models like Mamba), and generative modeling (autoencoders, GANs, and the diffusion models behind modern image generation). The track widens further into self-supervised learning and graph neural networks, and closes with the training systems (mixed precision, distributed training, and compilation) that make billion-parameter models trainable at all.

This is the hinge of the whole tree: NLP, modern RL, and most of AI engineering downstream assume you are fluent here.