Skip to content

RNN / LSTM

Phase: 3 | Status: ✅ Complete | Prerequisites: 13 Neural Networks

Overview

Recurrent neural networks for sequential data: vanilla RNN with parameter sharing across time, backpropagation through time (BPTT), vanishing/exploding gradient problem, LSTM gating mechanism (forget/input/output gates, cell state as gradient highway), and GRU as a simplified alternative.

Contents

# File Type Description
1 theory.md Theory RNN equations, BPTT derivation, vanishing gradients, LSTM/GRU gates, failure cases
2 first_principles.ipynb Computation From-scratch RNN and LSTM cells, vanishing gradient demo, long-dependency comparison, PyTorch verification
3 exercises.ipynb Practice Hand calculation of RNN forward step, RNN forward pass implementation, LSTM gradient highway analysis

Connections