Deep Learning

CNNs, RNNs & Transformers

CNNs exploit local spatial patterns, RNNs process sequences through state, and Transformers use attention so tokens can interact directly. Each architecture family encodes a different assumption about the data.

CNNs are built for locality and translation
RNNs process sequences step by step
LSTMs and GRUs gate memory
Transformers use attention over tokens
Architecture choice is an inductive-bias choice
Transformers are not automatically best for everything

Architecture-family cheat sheet
Family	Assumption	Typical use
CNN	Nearby values form local patterns	Images, spectrograms, grids
RNN/LSTM/GRU	Order matters and state evolves	Sequences, time series, older NLP
Transformer	Tokens need flexible pairwise interaction	LLMs, modern NLP, multimodal models
MLP	No special structure assumed	Dense features, simple baselines

Different ways to use structure

Architecture is how the model exploits known structure in the input.