When to Learn and When to Stop: Quitting at the Optimal Time (Student Abstract)

Diana Vins, Jean Erik Delanois, Maxim BazhenovAAAI 2025 · DOI

Abstract

Artificial Neural Networks (ANNs) often suffer from "catastrophic forgetting" when learning tasks sequentially, as new information typically overwrites the weights used for previous tasks. This paper introduces Optimal Stopping (OS), a strategy that halts training on a new task the moment it begins to degrade performance on previous ones. When combined with Sleep Replay Consolidation (SRC)—a biologically-inspired memory consolidation phase—the proposed method more than doubles the accuracy of standard continual learning models while providing a significantly more stable learning process.

Introduction

A fundamental challenge in machine learning is the ability to integrate new information without wiping out existing knowledge. In standard training scenarios, models are pushed to reach maximum performance on a current task. However, this "perfectionism" is often what destroys old memories. This work explores a "net benefit" philosophy, treating the network's total knowledge as a shared resource and ensuring that new knowledge is only integrated if it benefits the overall system.

Methods

The researchers propose a two-pronged approach to mitigate forgetting:

Optimal Stopping (OS)

Unlike traditional early stopping designed to prevent overfitting on a single task, OS monitors the mean validation accuracy across all tasks (both the current one and all previous ones). If the aggregate performance across the entire curriculum drops after an epoch, the model reverts to its state from the previous epoch and terminates training for that task. This ensures the model accepts "good enough" performance on new data to maintain the "health" of the entire system.

Sleep Replay Consolidation (SRC)

Inspired by biological sleep, SRC involves temporarily converting the ANN into a Spiking Neural Network (SNN). During a simulated "sleep" phase, the network undergoes unsupervised learning using Hebbian plasticity. This process replays activation patterns from previous tasks, effectively "refreshing" the model's memory and consolidating knowledge.

Experimental Setup

The approach was evaluated using the Split MNIST benchmark, where the standard handwritten digit dataset was divided into five sequential tasks. The model architecture was a 4-layer fully connected ANN. Performance was compared across four conditions: Baseline, OS only, Baseline + SRC, and the combined OS + SRC.

Results and Discussion

The experimental results demonstrate that while individual techniques improve performance, their combination offers the best balance of accuracy and stability.

Performance heatmaps for the four experimental conditions across five sequential tasks.

As illustrated in Figure 1, the Baseline model exhibits extreme forgetting, retaining only the most recent task and resulting in a mean accuracy of ~20%. Optimal Stopping (OS) alone improves mean accuracy to ~35% by preventing total erasure, though performance on individual tasks is lower than the baseline's peak.

The OS + SRC combination achieved ~45% mean accuracy. Crucially, as seen in the heatmaps, OS + SRC creates a much smoother and more stable learning curve compared to SRC alone. While SRC alone can reach higher peak performance (~60%), it suffers from high variability, with performance swinging wildly between "awake" and "sleep" phases. The stability provided by OS is vital for real-world applications where a model might be deployed at any point during its learning cycle.

Conclusion

This study highlights that the timing of learning and consolidation is critical for lifelong learning. By implementing a principled stopping criterion based on global network performance, models can better navigate the stability-plasticity dilemma. Future work will explore applying these thresholds to more complex architectures like Transformers to enable truly lifelong learning without memory loss.

Summary updated

BibTeX
@inproceedings{vins2025when,
  author = {Diana Vins and Jean Erik Delanois and Maxim Bazhenov},
  title = {{When to Learn and When to Stop: Quitting at the Optimal Time (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35309},
}
Figure