Energy-Based Models for Continual Learning

This project aims at classification continual learning problems using Energy-Based Models. Mainly based on our paper Energy-Based Models for Continual Learning.

• Project Page

• [Code] This code is the basic version of our paper. We will release the final version soon.

---

Requirements

The current version of the code has been tested with:

• pytorch 1.4.0
• torchvision 0.2.1

---

Training (Boundary-aware setting)

Split MNIST:

EBM:

sh scripts/boundary_aware/train_ebm_splitmnist.sh

Softmax-based classifier:

sh scripts/boundary_aware/train_sbc_splitmnist.sh

Permuted MNIST:

EBM:

sh scripts/boundary_aware/train_ebm_permmnist.sh

Softmax-based classifier:

sh scripts/boundary_aware/train_sbc_permmnist.sh

CIFAR-10:

EBM:

sh scripts/boundary_aware/train_ebm_cifar10.sh

Softmax-based classifier:

sh scripts/boundary_aware/train_sbc_cifar10.sh

CIFAR-100:

EBM:

sh scripts/boundary_aware/train_ebm_cifar100.sh

Softmax-based classifier:

sh scripts/boundary_aware/train_sbc_cifar100.sh

---

Training (Boundary-agnostic setting)

Split MNIST:

EBM:

sh scripts/boundary_agnostic/train_ebm_splitmnist.sh

Softmax-based classifier:

sh scripts/boundary_agnostic/train_sbc_splitmnist.sh

Permuted MNIST:

EBM:

sh scripts/boundary_agnostic/train_ebm_permmnist.sh

Softmax-based classifier:

sh scripts/boundary_agnostic/train_sbc_permmnist.sh

CIFAR-10:

EBM:

sh scripts/boundary_agnostic/train_ebm_cifar10.sh

Softmax-based classifier:

sh scripts/boundary_agnostic/train_sbc_cifar10.sh

CIFAR-100:

EBM:

sh scripts/boundary_agnostic/train_ebm_cifar100.sh

Softmax-based classifier:

sh scripts/boundary_agnostic/train_sbc_cifar100.sh

---

Acknowledgements

Parts of the code were based on the implementation of https://github.com/GMvandeVen/continual-learning.

---

Citation

Please consider citing our papers if you use this code in your research:

@article{li2020energy,
  title={Energy-Based Models for Continual Learning},
  author={Li, Shuang and Du, Yilun and van de Ven, Gido M and Torralba, Antonio and Mordatch, Igor},
  journal={arXiv preprint arXiv:2011.12216},
  year={2020}
}