Deep Learning

Five-course specialisation covering MLPs, CNNs, RNNs, optimisation, and structuring ML projects. Audit for free.

The canonical deep learning textbook. Chapters 6–9 cover feedforward networks, regularisation, optimisation, and CNNs. Free HTML version.

Series of YouTube lectures building neural networks from scratch: micrograd → makemore → GPT. The best way to build mechanical understanding.

University of Amsterdam deep learning course with detailed PyTorch tutorial notebooks covering MLPs, attention, normalizing flows, and more.

Official step-by-step tutorials: tensors, autograd, training a classifier, building custom datasets, and deploying models.

High-level PyTorch wrapper eliminating boilerplate. Excellent tutorials for training loops, checkpointing, logging, and multi-GPU training.

Original Adam optimiser paper — still the default for most deep learning tasks.

Introduces batch normalization and explains its effect on training stability and learning rate sensitivity.

Original dropout paper. Explains the co-adaptation problem and why random unit deactivation reduces overfitting.

The dominant deep learning framework. Dynamic computation graph, autograd, and a large ecosystem of libraries (PyG, HuggingFace, Lightning).

Visualization toolkit for training metrics, model graphs, histograms, and embeddings. Works natively with PyTorch via SummaryWriter.

70,000 handwritten digit images — the canonical introductory benchmark for image classification and neural network training.

60,000 32×32 colour images in 10 classes. Standard benchmark for CNNs and regularisation methods such as dropout and batch normalisation.

134k small organic molecules with 19 quantum chemical properties. Ideal chemistry benchmark for testing deep learning architectures beyond standard vision tasks.