TL;DR
This paper introduces "forget-and-relearn," a paradigm that reframes forgetting not as a failure, but as a beneficial step in neural network training. By selectively forgetting "undesirable" information and then relearning, models can achieve better generalization. This framework unifies existing iterative training methods and offers a path to designing more effective learning algorithms.
What Problem Does This Paper Address?
In both human and machine learning, forgetting is typically viewed as a negative phenomenon. In machine learning, catastrophic forgetting is a well-known issue where learning new tasks leads to a decline in performance on previously learned tasks. However, research in psychology suggests a symbiotic relationship between forgetting and learning, where forgetting can actually enhance recall and learning. This paper bridges this gap by proposing that forgetting, when strategically applied, can be a powerful tool for improving the generalization capabilities of artificial neural networks, moving beyond simply avoiding catastrophic forgetting to actively leveraging forgetting for better learning.
Methods
The core contribution is the forget-and-relearn paradigm. This involves an iterative process with two stages:
- Forgetting Stage: This stage selectively removes "undesirable" information from the model. The paper defines this as any process that leads to a decrease in training accuracy, while still retaining enough information to allow for subsequent relearning (i.e., accuracy remains above chance).
- Relearning Stage: The model is then retrained, reinforcing features that are consistently useful and were not forgotten.
The authors propose that many existing iterative training algorithms (like iterative magnitude pruning, knowledge evolution, and iterated learning) can be understood as instances of this forget-and-relearn process. They also introduce partial weight perturbation as a mechanism for targeted forgetting, where specific weights are modified (e.g., reinitialized or zeroed out) based on criteria like magnitude or layer importance.
Datasets
The paper uses several standard datasets for evaluation:
- Image Classification:
- MNIST (for initial experiments with MLPs)
- CIFAR-10
- CIFAR-100
- Tiny-ImageNet
- ImageNet
- Flower
- CUB (Caltech-UCSD Birds-200-2011)
- Aircraft
- MIT67
- Stanford-Dogs
- Language Emergence:
- Custom datasets for the Lewis game, designed to test compositional language learning.
Key Insights and Ideas
- Forgetting as a Feature Enhancer: The core idea is that forgetting isn't just about removing bad stuff; it's about making space to better learn and amplify the good stuff. By removing information that leads to overfitting or memorization of specific examples, the model is forced to focus on more generalizable features.
- Unifying Existing Methods: The forget-and-relearn framework provides a cohesive explanation for why various iterative training techniques, which might seem disparate, actually work. They all share the underlying principle of strategic forgetting and relearning.
- Targeted Forgetting is Key: The success of this paradigm hinges on what is forgotten. The paper demonstrates that disproportionately forgetting information associated with difficult or mislabeled examples leads to better generalization.
- "Goldilocks Zone" of Forgetting: There's an optimal balance. Forgetting too little means the model might not escape its current (potentially suboptimal) solution. Forgetting too much could lead to losing crucial information. Finding this sweet spot is an area for future research.
Key Results
- Improved Generalization: The proposed targeted forgetting methods, particularly "Later Layer Forgetting" (LLF), consistently outperform baselines and existing iterative methods like Knowledge Evolution (KE) on various image classification datasets. For example, on Flower, LLF achieved significantly higher accuracy than KE and other baselines.
- Targeted Forgetting Effectiveness: Experiments showed that weight perturbations disproportionately affected difficult or mislabeled examples more than easy or correctly labeled ones, supporting the hypothesis that undesirable information is selectively forgotten.
- Compositionality in Language: In language emergence tasks, forget-and-relearn mechanisms (like partial balanced forgetting) led to more compositional languages, which are crucial for out-of-distribution generalization. Visualizations showed that compositional mappings were retained and strengthened across generations.
- LLF Outperforms LW: The proposed LLF method was shown to outperform concurrent work on layer-wise reinitialization (LW) across several datasets.
Limitations and Future Work
- Hyperparameter Tuning: The LLF method introduces an additional hyperparameter (the layer threshold L), which requires tuning. While the paper suggests a heuristic for this, it adds complexity.
- Understanding the "Why": While the paper provides strong empirical evidence and a theoretical framework, further work is needed to fully understand the precise mechanisms by which iterative retraining distills features and leads to better generalization.
- Exploring the "Goldilocks Zone": Future research could focus on precisely defining and finding this optimal balance of forgetting for different tasks and architectures.
- Connections to Other Areas: The authors suggest exploring connections to invariant prediction and information bottleneck theory, as well as investigating how forgetting might be leveraged in continual learning settings. They also point to self-distillation methods as a potential area where a forget-and-relearn perspective could offer new insights.
Summary updated