Meet the AI that can rewind and replay its own learning journey
Neural networks can now remember old tasks by reconstructing their own past
This paper introduces ReCL (Reconstruction from Continual Learning), a framework that prevents neural networks from forgetting previously learned tasks by reconstructing old training data from the model's own weights, leveraging the implicit bias of gradient-based training towards margin maximization points.
-----
https://arxiv.org/abs/2411.06916
🤔 Original Problem:
→ In continual learning, neural networks suffer from catastrophic forgetting - they forget previously learned tasks when trained on new data.
→ Existing solutions either require storing old data (memory-based) or modify learning objectives (memory-free), both having significant drawbacks.
-----
🔧 Solution in this Paper:
→ ReCL exploits how gradient-based neural networks naturally converge to margin maximization points.
→ When a new task arrives, ReCL reconstructs samples from previous tasks using the model's weights.
→ The reconstruction uses three key losses: reconstruction loss to optimize candidates, lambda loss to constrain scaling coefficients, and prior loss to enforce value ranges.
→ These reconstructed samples are combined with new task data during training.
-----
💡 Key Insights:
→ Neural networks can act as their own memory buffers, eliminating need for external storage
→ The framework is flexible and can enhance existing continual learning methods
→ Works across different scenarios: class incremental, domain incremental, and task incremental learning
-----
📊 Results:
→ Improves accuracy by up to 57.09% compared to baseline methods
→ Reduces forgetting (backward transfer) by up to 152.37%
→ Effective across MNIST and CIFAR10 datasets with both MLPs and CNNs
Share this post