# Generative replay with feedback connections as a general strategy for continual learning

Source: [Generative replay with feedback connections as a general strategy for continual learning](<https://www.sota2.com/research/paper/68fa659f1e209131cea0b52b>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-25T21:30:43.473Z

Authors: Gido M. van de Ven, Andreas S. Tolias

Published: 2018-09-27T16:55:58.000Z

## Abstract

A major obstacle to developing artificial intelligence applications capable of true lifelong learning is that artificial neural networks quickly or catastrophically forget previously learned tasks when trained on a new one. Numerous methods for alleviating catastrophic forgetting are currently being proposed, but differences in evaluation protocols make it difficult to directly compare their performance. To enable more meaningful comparisons, here we identified three distinct scenarios for continual learning based on whether task identity is known and, if it is not, whether it needs to be inferred. Performing the split and permuted MNIST task protocols according to each of these scenarios, we found that regularization-based approaches (e.g., elastic weight consolidation) failed when task identity needed to be inferred. In contrast, generative replay combined with distillation (i.e., using class probabilities as "soft targets") achieved superior performance in all three scenarios. Addressing the issue of efficiency, we reduced the computational cost of generative replay by integrating the generative model into the main model by equipping it with generative feedback or backward connections. This Replay-through-Feedback approach substantially shortened training time with no or negligible loss in performance. We believe this to be an important first step towards making the powerful technique of generative replay scalable to real-world continual learning applications.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Incremental Task Learning on split-MNIST (test)](<https://www.sota2.com/research/sota/incremental-task-learning-on-split-mnist-test.md>): Incremental Task Learning (ITL); split-MNIST (test); Retained Accuracy: 99.66.

- [Incremental Domain Learning (IDL) on permuted-MNIST (test)](<https://www.sota2.com/research/sota/incremental-domain-learning-idl-on-permuted-mnist-test.md>): Incremental Domain Learning (IDL); Permuted MNIST (test); Retained Accuracy: 97.06.

- [Incremental Task Learning (ITL) on permuted-MNIST (test)](<https://www.sota2.com/research/sota/incremental-task-learning-itl-on-permuted-mnist-test.md>): Incremental Task Learning (ITL); Permuted MNIST (test); Retained Accuracy: 97.31.

- [Incremental Domain Learning (IDL) on split-MNIST (test)](<https://www.sota2.com/research/sota/incremental-domain-learning-idl-on-split-mnist-test.md>): Incremental Domain Learning (IDL); split-MNIST (test); Retained Accuracy: 97.31.

- [Incremental Learning on MNIST](<https://www.sota2.com/research/sota/incremental-learning-on-mnist.md>): Incremental Learning; MNIST; Average Accuracy: 0.9256.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/1809.10635>)

[PDF](<https://arxiv.org/pdf/1809.10635.pdf>)
