Deep Learning¶
working pillar 9
Architectures, the implementation patterns that recur, and generative modelling. An active growth area.
Pages¶
| Page | Covers |
|---|---|
| Architectures | MLP, CNN, RNN, Transformer, U-Net, GNN — and the prior each encodes |
| PyTorch patterns | Training loops, data pipelines, distributed training, mixed precision |
| Generative models | VAEs, GANs, diffusion, and what each sacrifices |
| LLMs and generative AI | Autoregressive generation, RAG, fine-tuning, agents |
The theme¶
An architecture is an assumption about structure, made cheap. Convolution assumes locality and translation equivariance. Attention assumes anything may be relevant and learns which. Graph networks assume permutation invariance. U-Net assumes multiscale structure with detail worth preserving.
Choosing well means matching the assumption to a symmetry the data actually has. Where it holds, the prior buys data efficiency for free; where it does not, you are paying for a constraint that only limits you — and where you have no idea, attention's learn-the-relation approach is the expensive general answer.
A note on volatility¶
LLMs and generative AI deliberately omits model names, prices and context limits. Those change on a timescale of weeks, and a recap page that carried them would be confidently wrong within a month. Mechanisms and failure modes are what belongs in a durable note; specifics belong in the provider's live documentation.
Connections¶
- Machine Learning — classical methods and the evaluation discipline.
- Computer Vision — where most of these architectures are applied.
- Scientific ML — the same tools with physics as structure.