Introduction
Artificial Neural Networks (ANNs) typically struggle with "continual learning," where data is presented sequentially rather than in a shuffled, independent and identically distributed (IID) format. This often leads to catastrophic forgetting (CF), where learning new tasks (e.g., identifying cats) causes the network to overwrite weights associated with previously learned tasks (e.g., identifying dogs).
A common mitigation strategy is generative replay, where a model—such as a Variational Autoencoder (VAE)—learns to generate "fake" samples of old classes to rehearse them alongside new data. However, as the number of classes grows, these generated samples often become blurry or confused with similar-looking new classes. This research introduces a novel approach inspired by the human brain's ability to exaggerate differences between similar memories to prevent confusion.
Methods: Reconstruction Repulsion and Attraction
The authors build upon the Brain-Inspired Replay (BI-R) model by introducing two new loss functions applied during the replay phase:
- Reconstruction Repulsion (RR) Loss: This loss identifies "competing classes" that the model currently confuses with the target sample. It forces the VAE to maximize the difference (using Binary Cross-Entropy) between the generated sample and the mean feature vector of those competing classes.
- Reconstruction Attraction (RA) Loss: This loss encourages samples of the same class to move closer to their shared mean feature vector, tightening internal representations.
To identify competitors, the model uses an adaptive repulsion factor ($f$). It monitors its own softmax output; if a non-target class has a probability higher than a dynamic threshold (the top class probability divided by $f$), it is deemed a competitor and "pushed away."
Key Insights: Embracing Distortion
While generative models usually aim for perfect accuracy, this paper argues that intentional distortion—exaggerating differences—is a feature of high-functioning biological memory. By making a replayed "dog" look intentionally "less like a cat" than it does in reality, the model maintains more distinct boundaries between classes. This mechanism uses the model's own interference (confusion) as a teaching signal to determine which features require exaggeration.
Results and Performance
The study primarily evaluated the approach on CIFAR100 in a class-incremental setting. The "RR+RA" adaptation showed significant improvements over the standard BI-R baseline, particularly for the earliest learned classes which are typically the most vulnerable to forgetting.
Key findings include:
- Early Class Protection: For the first 20% of classes encountered, the model achieved a 15.79% relative improvement in precision.
- Mid-stream Stability: For the first 50% of classes, the model saw a 15.30% relative increase.
- Efficiency: These gains were achieved without storing original images, preserving the memory efficiency of generative replay.
Limitations and Future Work
The primary limitation is the "Correct Label Assumption." The RR+RA logic relies on the classifier correctly identifying the class of a replayed sample. If a replayed image is mislabeled, the repulsion logic may exaggerate the wrong features. Future work aims to incorporate model confidence into the RR loss to mitigate the impact of mislabeled samples and apply this logic to non-image data types.
Summary updated