← Dashboard
MODULE 7 · ABOUT 6 HOURS

Building an LSTM

Memory with valves. Build one LSTM cell, then train a full model.

This module is an outline. The full lecture, widget, quiz and lab are on the way.

Lecture

What you’ll learn

  • Why order matters in a sequence
  • Recurrent networks, unrolled through time
  • Why plain recurrent networks forget
  • The LSTM’s gates: forget, input and output
  • Training curves, overfitting and early stopping
  • Comparing a trained model against baselines, fairly

Lecture outline

  1. Data with a memory
  2. Recurrent networks
  3. The forgetting problem
  4. Inside an LSTM cell
  5. Training a Keras LSTM
  6. Did it beat the baseline?

Discover: Memory Valves

Toggle the forget, input and output gates and watch the memory change along a sequence.

Interactive widget coming soon.

Quiz

8 to 10 questions. 80% to pass, unlimited retakes.

Quiz coming soon.

Lab: From one cell to a trained model

7a builds one LSTM cell in NumPy in the browser. 7b trains a Keras LSTM in Google Colab and compares it to the baselines.

Part A in your browser, part B in Google Colab

Notebook coming soon.