Skip to content

EVCL-WC: Elastic Variational Continual Learning with Weight Consolidation

Notifications You must be signed in to change notification settings

hunarbatra/elastic-variational-continual-learning

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

64 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

EVCL-WC: Elastic Variational Continual Learning with Weight Consolidation

Abstract

Continual learning aims to train models on sequential tasks while balancing adaptation to new data and retention of existing knowledge. This work introduces Elastic Variational Continual Learning with Weight Consolidation (EVCL-WC), a novel hybrid model that integrates the variational posterior approximation mechanism of Variational Continual Learning (VCL) with the regularization-based parameter protection strategy of Elastic Weight Consolidation (EWC). By combining the strengths of both methods, EVCL-WC effectively mitigates catastrophic forgetting and enables better capture of dependencies between model parameters and task-specific data. Evaluated on five discriminative tasks, EVCL-WC consistently outperforms existing baselines in both domain-incremental and task-incremental learning scenarios for deep discriminative models.

Steps to run:

EVCL-WC:

python3 run_evcl.py \
        --num_tasks=5 \
        --num_epochs=100 \ 
        --task_config='split_mnist' \ # supported task_configs: ['permuted_mnist', 'split_mnist', 'no_mnist', 'fashion_mnist', 'split_cifar']
        --experiment_name='experiment-name' \ # runs saved under f'experiments/{task_config}/{experiment_name}.csv'
        --ewc_lambda=100 # EWC regularization term λ

VCL and VCL + Coresets:

python3 run_vcl.py \
        --num_tasks=5 \
        --num_epochs=100 \
        --task_config='split_mnist' \ # supported task_configs: ['permuted_mnist', 'split_mnist', 'no_mnist', 'fashion_mnist', 'split_cifar']
        --coreset_size=200 \ # set to 0 if w/o coreset
        --coreset_method='random' \ # set to None if w/o coresets; supported coreset_methods: ['random', 'k-center', 'pca-k-center', 'class_balanced']
        --experiment_name='experiment-name' # runs saved under f'experiments/{task_config}/{experiment_name}.csv'

Note: set coreset_size = 0, and coreset_method = None, if running VCL without coreset.

EWC:

python3 run_ewc.py \
        --num_tasks=5 \
        --num_epochs=100 \
        --task_config='split_mnist' \ # supported task_configs: ['permuted_mnist', 'split_mnist', 'no_mnist', 'fashion_mnist', 'split_cifar']
        --experiment_name='experiment-name' \ # runs saved under f'experiments/{task_config}/{experiment_name}.csv'
        --ewc_lambda=100 \ # EWC regularization term λ

Coreset Only:

python3 coreset.py \
  --num_tasks=5 \
  --num_epochs=100 \
  --task_config='split_mnist' \
  --coreset_size=200 \
  --coreset_method='random' # supported coreset_methods: ['random', 'k-center', 'pca-k-center', 'class_balanced']

Results

PermutedMNIST:

Permuted MNIST Plot

SplitMNIST:

Split MNIST Plot

SplitNotMNIST:

Split Not MNIST Plot

SplitFashionMNIST:

Split Fashion MNIST Plot

SplitCIFAR:

Split CIFAR Plot

About

EVCL-WC: Elastic Variational Continual Learning with Weight Consolidation

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages