Introduction
Continual learning enables neural networks to learn tasks sequentially without forgetting previously acquired knowledge. However, modern architectures frequently suffer from catastrophic forgetting, a phenomenon where learning a new task overwrites the weights essential for earlier ones.
Existing solutions typically fall into two categories:
- Memory-based methods: These store a "buffer" of old data to replay later, which raises privacy concerns and consumes significant storage.
- Dynamic architectures: These add new neurons for every new task, inevitably leading to "parameter explosion" and making the model too large for practical deployment.
To address these limitations, researchers introduce Shapley Neuron Valuation (SNV), a principled framework grounded in cooperative game theory. SNV selectively identifies and freezes the most important neurons for a specific task while keeping others plastic. By leveraging the inherent redundancy in over-parameterized neural networks, SNV enables state-of-the-art, buffer-free continual learning without expanding the model's architecture.
Methodology: Shapley Neuron Valuation (SNV)
In game theory, Shapley values provide a mathematically rigorous way to fairly distribute a "payout" among players based on their individual contributions to a team's success. SNV adapts this concept to neural networks:
- Players: Individual neurons (specifically, convolutional filters).
- Payout: The model's performance (accuracy or loss).
- The Goal: Calculate a value ($\phi_i$) for every neuron to quantify its contribution to a specific task.
Once the "expert" neurons for Task 1 are identified, they are frozen. When the model moves to Task 2, it only updates the "plastic" neurons that were deemed non-critical for Task 1.
The TMAB-Shapley Pipeline
Calculating exact Shapley values is computationally prohibitive because it requires evaluating every possible combination of neurons. To make SNV feasible for real-world AI, the authors developed the TMAB-Shapley pipeline, which relies on three efficiency mechanisms:
- Monte Carlo Estimation: Samples random subsets of neurons instead of exhaustively checking every combination.
- Truncation: Halts calculations early if a subset of neurons is so small that the model's performance drops to near zero.
- Multi-Armed Bandit (MAB): Focuses computational effort on neurons whose importance is still uncertain, avoiding wasted cycles on neurons already definitively classified as "important" or "useless."
Experimental Evaluation
The authors evaluated SNV on three standard computer vision benchmarks: CIFAR-100, Tiny-ImageNet, and ImageNet-1k (the "gold standard" large-scale dataset with 1,000 classes).
Testing encompassed two primary scenarios:
- Class Incremental Learning (CIL): A highly challenging setup where the model does not know which task a sample belongs to during inference.
- Task Incremental Learning (TIL): A setup where task identity is provided.
Key Findings
1. State-of-the-Art Performance
SNV consistently outperforms existing buffer-free methods. In the rigorous 10-task CIL scenario on ImageNet-1k, SNV achieved 41.30% accuracy, beating the next best buffer-free method (NFL+) by +2.88%. In the TIL scenario on ImageNet-1k, SNV improved accuracy by +6.46% over the second-best baseline.
2. Zero Forgetting
In TIL scenarios, SNV maintained zero backward transfer. This means it perfectly preserved its knowledge of the first task even after sequentially learning 50 additional tasks.
3. Quality Over Quantity
Surprisingly, SNV—which uses absolutely no stored data—outperformed several "memory-based" methods that stored up to 20,000 historical images. This suggests that mathematically protecting the right neurons is fundamentally more effective than simply replaying old data.
4. The "Stable vs. Plastic" Hierarchy
Unlike previous methods that rely on simple heuristics (like gradient magnitude) to guess neuron importance, SNV's principled approach yields highly stable neuron selection. The framework naturally discovers that early layers (which capture basic edges and textures) are often shared across tasks, while later layers become highly task-specific.
5. Computational Efficiency
While Shapley estimation sounds slow, the TMAB-Shapley pipeline is highly efficient. On CIFAR-100, it adds only about 1.3 minutes to a standard 5-minute training phase—a modest 26% increase in computational time for a massive gain in model stability.
Limitations and Future Directions
The primary limitation of SNV is its reliance on a post-task estimation step. Currently, the model must finish training on a given task before it can calculate which neurons to freeze, adding a distinct "valuation phase" to the workflow.
Future work will focus on online Shapley estimation, where neuron importance is calculated during the training process. This would eliminate the extra phase and allow the model to adapt its freezing strategy in real-time. Additionally, as models grow significantly wider, the number of permutations to sample will scale up, potentially requiring even more aggressive approximation techniques to maintain efficiency.
Summary updated