Continual Learning with Learning without Prejudice (LwP)

This project implements a continual learning system from scratch to tackle the problem of catastrophic forgetting. A custom Convolutional Neural Network (CNN) is trained sequentially on a stream of 20 datasets derived from CIFAR-10 without storing past data (except for a small buffer in Task 2).

The core strategy is based on Learning without Prejudice (LwP), which uses knowledge distillation and pseudo-labeling to learn new tasks while preserving previously acquired knowledge.

---

Project Structure

The project is divided into two main parts, corresponding to the two scripts:

1. task1.py: Implements the continual learning solution for Task 1, where there is no domain shift between datasets D1 to D10. It uses the standard LwP strategy.
2. task2.py: Implements the solution for Task 2, which involves a domain shift in datasets D11 to D20. This script enhances the LwP strategy with Experience Replay to mitigate the negative effects of the changing data distribution.
3. data/: This directory should contain the datasets after downloading and unzipping.
   • task1_no_domain_shift/
   • task2_domain_shift/

---

Requirements

• Python 3.x
• PyTorch

No other external libraries like numpy are required.

---

Setup & How to Run

1. Download the Dataset: Download the data from the URL provided in the project description: https://tinyurl.com/cs771-mp2-data

2. Unzip the Data: Unzip the file to create a data directory with the two sub-folders (task1_no_domain_shift and task2_domain_shift).

3. Place Scripts: Ensure task1.py and task2.py are in the same root directory as the data folder.

4. Run the Experiments: Execute the scripts from your terminal.

   • To run Task 1:

     python task1.py

   • To run Task 2:

     python task2.py

---

Methodology

Task 1: No Domain Shift

The model is trained sequentially on datasets D1 through D10.

• Initial Training: A standard supervised training is performed on the labeled dataset D1.
• Continual Update: For subsequent tasks (D2-D10), which are unlabeled, the model is updated using the LwP strategy. A copy of the model from the previous step (the "teacher") generates pseudo-labels for the new data. The current model (the "student") is then trained with a combined loss function consisting of:
  1. Pseudo-Label Loss: A standard classification loss on the pseudo-labels.
  2. Distillation Loss: A KL-Divergence loss that encourages the student's output distribution to match the teacher's, preserving past knowledge.

Task 2: With Domain Shift

The model is trained on datasets D11 through D20, which have changing data distributions. This task builds on the LwP strategy by adding Experience Replay.

• Memory Buffer: A small buffer stores a few correctly labeled examples from each task after it is learned.
• Combined Training: During the update step, the model's total loss is a combination of the LwP losses (from the new, unlabeled data) and a standard supervised loss calculated on a small batch of "replayed" examples from the memory buffer. This replay mechanism acts as an anchor to past knowledge, making the model more robust to domain shifts.

---

Output

Each script will train the model sequentially and print the progress. At the end of the run, it will display the final 10x10 accuracy matrix, which shows the model's performance on all previously seen tasks after each update.