Training to Reduce Catastrophic Forgetting

This project demonstrates continual learning on MNIST using a small CNN and three strategies to mitigate catastrophic forgetting:

• Naive sequential training (baseline)
• Experience replay (rehearsal buffer)
• Elastic Weight Consolidation (EWC) (regularization)

All experiments are contained in the notebook adaptive_learning.ipynb.

Model overview

The model is a simple CNN with two convolutional blocks and a two-layer MLP head:

• Conv(1->16) + ReLU + MaxPool
• Conv(16->32) + ReLU + MaxPool
• Linear(3277->128) + ReLU
• Linear(128->10)

It is trained on two tasks:

1. Task 1: Standard MNIST.
2. Task 2: Permuted MNIST (fixed random pixel permutation).

Methods in the notebook

1. Naive training
   • Train on Task 1, then train on Task 2 without any protection.
2. Experience replay
   • Mix a small buffer of Task 1 samples into Task 2 training.
3. EWC
   • Estimate Fisher information after Task 1 and add a quadratic penalty during Task 2.

Metrics and visualization

The notebook computes:

• Accuracy on Task 1 after Task 2 (forgetting).
• Accuracy on Task 2 after Task 2 (plasticity).
• Average accuracy across tasks.
• Forgetting measure: $\Delta = \text{peak}\text{T1} - \text{acc}\text{T1 after T2}$.

It also produces a plot comparing methods and a summary table.

Requirements

• Python 3.9+
• PyTorch
• torchvision
• matplotlib
• tqdm
• numpy

How to run

1. Open adaptive_learning.ipynb in VS Code or Jupyter.
2. Run cells top-to-bottom.

Notes

• The permutation for Task 2 is fixed with a single random permutation tensor.
• Training uses a small number of epochs per task for quick iteration.