Self-Supervised Pseudodata Filtering for Improved Replay with Sub-Optimal Generators
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: continual learning, catastrophic forgetting, generative replay, bayesian neural networks, deep learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Continual learning on a sequence of tasks without forgetting previously acquired knowledge is one of the main challenges faced by modern deep neural networks. In the class-incremental scenario, one of the most difficult continual learning problems, new classes are presented to a classifier over time. The model needs to be able to learn and recognize these new classes while also retaining its knowledge of previously witnessed ones. To achieve this, the model has to revisit previous classes in some form, either by analysing stored exemplars or by using artificially generated samples. The latter approach, Generative Replay, usually relies on a separate generator trained alongside the main classifier. Since the generator also needs to learn continually, it is retrained on every task, using its own generated samples as training data representing older classes. This can lead to error propagation and accumulating features unimportant or confusing for the classifier, reducing the overall performance for larger numbers of tasks. We propose a simple filtering mechanism for mitigating this issue – whenever pseudodata is generated for a new task, the classifier can reject samples it is not able to classify with sufficient confidence, thus preventing itself from retraining on poor-quality data. We tested this mechanism using combinations of Bayesian neural classifiers and two different generators: a Variational Autoencoder and Real-value Non-Volume Preserving Normalizing Flow. We show that the improvement in the classification accuracy grows with the number of tasks, suggesting this approach is particularly useful for the most challenging continual learning scenarios, where very many tasks are learned in a sequence.
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: 7296
Loading