This page is due for review. Re-read the key points and answer the self-check questions before recording your video.
ML/DL · Hub
Deep learning
Twenty-five pages, in learning order: the neuron and its activations, how networks are trained (forward pass, backpropagation, losses, optimisers), how they are regularized, and the main architectures: CNNs, RNNs, LSTMs, GRUs, attention and Transformers, transfer learning, autoencoders, GANs and VAEs.
Why it matters: every modern AI system, from image recognition to language models, is built from these pieces. Knowing what each one computes, with real numbers, makes the big models far less mysterious.
5 random questions mixed from all twenty-five deep-learning pages.
Mixed review
Answer each one out loud before opening the sample answer.
flowchart TD
A{"What is your data?"} -->|"images"| C["CNN + pooling (or transfer learning if data is small)"]
A -->|"sequences: text, audio, time series"| S{"Long sequences, lots of data?"}
S -->|yes| T["Transformer (attention)"]
S -->|"no, small and simple"| R["LSTM / GRU"]
A -->|"tables of features"| M["MLP (or gradient boosting)"]
A -->|"want to generate new data"| G{"Sharp samples or a smooth latent space?"}
G -->|sharp| GAN["GAN"]
G -->|smooth latent space| VAE["VAE"]
A -->|"unlabelled, want compression or anomalies"| AE["Autoencoder"]
Key points
A network is layers of neurons because each layer computes $g(Wx + b)$ on the previous layer's output.
Training is gradient descent because backpropagation gives the gradient of the loss for every weight.
Adam is the usual default because it combines momentum with per-weight step sizes.
Dropout, batch norm and early stopping fight overfitting because each stops the network from memorising its training data.
Architectures match data shapes because CNNs exploit local image structure, RNNs and Transformers sequence order.