Biologically Inspired Sleep Algorithm for Reducing Catastrophic Forgetting in Neural Networks

Timothy Tadros, Giri P. Krishnan, Ramyaa Ramyaa, Maxim BazhenovAAAI 2020 · DOI

Introduction

Artificial Neural Networks (ANNs) face a significant challenge known as catastrophic forgetting. When a network trained on Task A is subsequently trained on Task B, the new information often overwrites the weights associated with the first task, leading to a total loss of prior knowledge. This limitation hinders the development of "continual learning" systems capable of accumulating knowledge over time.

In contrast, biological brains excel at incremental learning, largely due to the role of sleep in memory consolidation. During sleep, spontaneous neural activity, or "replay," helps stabilize and integrate new memories without erasing old ones. This research explores an algorithm that borrows these biological mechanisms to help artificial networks "remember" what they have learned.

Methodology: The Sleep Algorithm

To address catastrophic forgetting, the researchers proposed a multi-step "Sleep Algorithm" that bridges the gap between static ANNs and dynamic biological brains:

  1. ANN to SNN Conversion: Standard feed-forward ANNs lack the temporal dynamics necessary for biological-like replay. The researchers first convert a trained ANN into a Spiking Neural Network (SNN).
  2. The Sleep Phase: The network undergoes a sleep phase where it is stimulated with "noisy" versions of the average input seen across all tasks, rather than new training data.
  3. Synaptic Plasticity (STDP): During sleep, weights are updated using Spike-Timing Dependent Plasticity (STDP). This biological rule strengthens connections between neurons that fire in sequence and weakens those that do not.
  4. Simulated Replay: By scaling weights to induce high firing rates (simulating biological "up-states"), the network triggers a replay of previously learned patterns, reactivating specific neural pathways without requiring the original training data.

Experimental Results

The algorithm was evaluated on incremental MNIST and CUB-200 datasets. In the MNIST task, the network learned pairs of digits sequentially. Typically, accuracy on the first pair would drop to near zero after learning subsequent pairs. However, the introduction of the sleep phase allowed for significant performance recovery.

Accuracy recovery across sequential tasks after sleep phases.

As shown in the figure above, Task 1 accuracy often plummeted to 0% during the training of new tasks (labeled 'T') but was restored to high levels (e.g., 77%) following the sleep phase (labeled 'S'). This demonstrates that the internal representations of earlier tasks are preserved and can be "re-awakened."

Mechanistic Insights: Replay and Decorrelation

The success of the sleep algorithm stems from two key phenomena: replay without storage and representational decorrelation. The network does not need to store a database of old images; instead, its internal dynamics naturally reactivate the pathways associated with old tasks when stimulated by noise.

Furthermore, sleep clarifies these memories. Analysis of the hidden layers reveals that STDP during sleep helps decorrelate neural activations. By making the representations of different tasks more distinct, the network reduces the interference that causes forgetting.

Activation correlations in Layer 2 before and after the sleep phase.

The correlation matrices indicate a significant reduction in off-diagonal values after sleep, proving that the network's internal "concepts" become more distinct and less prone to overlap.

Conclusion

This biologically inspired approach provides a promising proof-of-concept for mitigating catastrophic forgetting. By leveraging SNNs and simulated sleep, networks can retain and even clarify old memories while learning new ones. Future work will focus on scaling this approach to larger datasets and exploring more complex biological sleep stages, such as REM and NREM, to further enhance generalization and memory stability.

Summary updated

BibTeX
@inproceedings{tadros2020biologically,
  author = {Timothy Tadros and Giri P. Krishnan and Ramyaa Ramyaa and Maxim Bazhenov},
  title = {{Biologically Inspired Sleep Algorithm for Reducing Catastrophic Forgetting in Neural Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i10.7239},
}
Figure