BTSP-CAM: A Brain-Inspired Geometric Memory for Class-Incremental Learning

Gradient-based optimization in class-incremental learning (CIL) often faces the plasticity–stability dilemma, since continuous weight updates can distort decision boundaries learned from earlier tasks. We revisit this problem from the viewpoint of stochastic geometric memory allocation and propose BTSP-CAM, a gradient-free memory system that instantiates theoretical insights from the hippocampal simpleBTSP model into a practical algorithm. Rather than fine-tuning a frozen encoder by backpropagation, BTSP-CAM externalizes plasticity into a binary synaptic matrix that evolves through local stochastic bit-flip updates. A trace-gated plateau process, driven by eligibility traces together with familiarity and collision signals, modulates when and where synapses are rewritten and suppresses cross-class interference in Hamming space. The resulting geometric memory states are mapped to semantic logits through a CA1-like competitive layer and a closed-form ridge readout, enabling fast consolidation after each task. Empirically, BTSP-CAM rivals gradient-based methods in a strictly exemplar-free setting and consistently boosts SOTA baselines as a lightweight plugin. Mechanistic analysis validates our geometric theory, confirming that stochastic repulsion actively bounds class overlap and stabilizes decision margins.

Artificial intelligence systems often need to learn new categories over time without keeping old training images. A fundamental problem is that updating a model to learn new concepts can overwrite what it learned before. To address this problem, we introduce BTSP-CAM, a lightweight memory method inspired by how biological memory can quickly store and separate experiences. Instead of constantly adjusting the internal settings of the whole vision model, our approach keeps the main visual network fixed and writes new information into a compact binary memory. This memory can be viewed as a digital board of simple on-and-off switches: when the model sees new information, it flips selected switches to form a memory trace. When a new category is too similar to an old one, the system helps separate them to reduce confusion and protect older memories. Experiments on multiple image classification benchmarks show that BTSP-CAM performs strongly as a standalone classifier and can also be added to existing continual learning methods to improve their ability to remember past information.