Generative Latent Replay

Method overview

Repo for Generative Latent Replay (GLR) - a continual learning method which aleviates catastophic forgetting through strict regularisation of low level data representation and synthetic latent replay. Explicitly GLR:

1. Freezes the backbone of a network after initial training
2. Builds generative models of the backbone-output latent representations of each dataset encountered by the model
3. Samples latent pseudo-examples from these generators for replay during subsequent training (to mitigate catastrophic forgetting)

Features

Generative latent replay overcomes two issues encountered in traditional replay strategies:

1. High memory footprint:
   • replays can be sampled ad hoc
   • caches [compressed] latent representations
2. Privacy concerns
   • data is synthetic

Continual Learning Method	Replay based	Low memory	Privacy
Naive	❌	❌	✅
Replay	✅	❌	❌
Latent Replay	✅	✅	❌
Generative Latent Replay	✅	✅	✅

Experiments

Description

We compare generative latent replay against the above methods on the following datasets:

• Permuted MNIST
• Rotated MNIST
• CoRE50

We also explore the effect of different:

• generative models (GMM, etc)
• network freeze depths
• replay buffer sizes
• replay sampling strategies

Reproducing experiments

To run experiments, first create and activate a virtual environment:

uv venv
source .venv/bin/activate

Then run the appropriate notebooks detailing the experiments.

Alternatively you can run the notebook directly in Google Colab:

Porting method

Our implementation is fully compatible with the Avalanche continual learning library, and can be imported as a plugin in the same way as other Avalanche strategies:

from avalanche.training.plugins import StrategyPlugin
from glr.strategies import GenerativeLatentReplay

Citation

Important

If you use any of this code in your work, please reference us:

@misc{armstrong2022generative,
      title={Generative Latent Replay for Continual Learning}, 
      author={J. Armstrong and A. Thakur and D. Clifton},
      year={2022},
      howpublished = "\url{https://github.com/iacobo/generative-latent-replay/blob/main/Generative_Latent_Replay.pdf?raw}",
}