Predictive Uncertainty Estimates for Catastrophic Forgetting

The repository contains the code to reproduce the results of the paper "How to Leverage Predictive Uncertainty Estimates for Reducing Catastrophic Forgetting in Online Continual Learning").

Requirements

The codebase is implemented in Python 3.11.4 and tested on Linux OS.

• pytorch==2.0.1
• torchvision==0.15.2
• numpy==1.24.3
• scikit-learn==1.5.0
• medmnist==3.0.1
• matplotlib==3.7.1

Structure of the repo

• utils contains all the utilities to preprocess and save the data, to run the experiments, and to save the results.
• models contains the model architectures of our evaluation framework.
• configuration contains the hyperparameters (dataset, memory population strategy, memory sampling strategy, memory size, etc.) to run our experiments.

Run the experiments

The current version of the configuration file (configuration/config_jup.py) enables running an experiment on CIFAR10 over three runs. All the parameters can be changed as desired in the main_CL.ipynb file. The most important hyperparameters are the followings:

• --memory_size: to set the size of the memory buffer
• --dataset_name: name of the dataset to evaluate
• --update_strategy: to select whether we want to use reservoir (reservoir) or class-balanced update (balanced).
• --balanced_update: if balanced is used, we can select whether we want to populate the memory randomly (random) or use the uncertainty (uncertainty).
• --uncertainty_score: if uncertainty is used, we can decide the uncertainty score we want to use among the available ones (bregman, confidence, margin, entropy, rainbow, ratio).
• --balanced_step: if uncertainty is used, we can decide the sorting strategy, i.e., bottomk (bottom-k), step (step-sized), and topk (top-k).

CIFAR10-LT and CIFAR100-LT

To create these two datasets, first run the experiments with CIFAR10 and CIFAR100. After running these experiments, it is possible to create the long-tailed (LT) version of the already preprocessed data. The snippet of code to save the LT version of the datasets is provided at the end of the Jupyter notebook.