Generate to Discriminate: Expert Routing for Continual Learning
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Domain Incremental Learning, Continual Learning, Distribution Shift, Transfer Learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: Expert routing for continual learning using generative models for domain discrimination
Abstract: In many real-world settings, norms, regulations, or economic incentives
permit the sharing of models but not data across environments.
Prominent examples arise in healthcare
due to regulatory concerns.
In this scenario, the practitioner wishes to adapt the model to each new environment
but faces the danger of losing performance on previous environments
due to the well-known problem of catastrophic forgetting.
In this paper, we propose Generate-to-Discriminate (G2D), a novel approach that leverages recent advancements in generative models to alleviate the catastrophic forgetting problem in continual learning.
Unlike previous approaches based on generative models
that primarily use synthetic data
for training the label classifier,
we use synthetic data to train a domain discriminator.
Our method involves the following steps:
For each domain, (i) fine-tune the classifier and adapt a
generative model to the current domain data;
(ii) train a domain discriminator to distinguish synthetic samples
from past versus current domain data;
and (iii) during inference, route samples to the respective classifier.
We compare G2D to an alternative approach, where we simply replay the generated synthetic data, and, surprisingly, we find that training a domain discriminator is significantly more effective than augmenting the training data with the same synthetic samples. We consistently outperform previous state-of-the-art domain-incremental learning algorithms
by up to $7.6$ and $6.2$ points across three standard
domain incremental learning benchmarks in the vision and language modalities, respectively,
and $10.0$ points on a challenging real-world dermatology medical imaging task.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8161
Loading