mistake-gated-continual-learning

# Mistake Gating for Energy and Memory Efficient Continual Learning ## Description 'Mistorized mistake-gated learning' -- a biologically plausible plasticity rule where synaptic updates are strictly gated by current and past classification errors. Based on Pache & van Rossum 2026 (arXiv:2604.14336v1). Inspired by human negativity bias and error-related negativity (ERN) in EEG. ## Core Innovation ### Biological Inspiration - **Human Negativity Bias**: Humans learn more from negative experiences - **Error-Related Negativity (ERN)**: Brain signal associated with error detection - **Metabolic Efficiency**: Animals update models without exhausting energy ### Problem Addressed Standard neural network training: - Updates parameters on every sample - Even correctly classified samples trigger updates - Inefficient for continual learning ### Solution **Mistake-Gated Learning**: Only update on errors - Reduces updates by **50-80%** - No hyperparameters added - Negligible computational overhead ## Methodology ### Update Rule ``` Standard: Δw = η * ∇L Mistake-Gated: Δw = η * ∇L * I(error) Where I(error) = 1 if prediction ≠ target, else 0 ``` ### Memorized Version ``` Update on current OR past errors: Δw = η * ∇L * I(error_t OR error_{t-k} for k in buffer) ``` ### Algorithm ```python def mistake_gated_update(weights, gradient, prediction, target): if prediction != target: # Current error weights -= lr * gradient elif memory_buffer_has_error(): # Past error weights -= lr * gradient # Otherwise: no update ``` ## Benefits ### 1. Energy Efficiency - **50-80% fewer updates** - Synaptic plasticity is metabolically expensive - Critical for edge/neuromorphic deployment ### 2. Memory Efficiency - Reduces storage buffer requirements - Only store samples with errors - Enables larger replay buffers ### 3. Continual Learning Well-suited for: - **Incremental learning**: New knowledge on pre-existing background - **Online learning**: Data stored for later replay - **Non-stationary data**: Adapting to distribution shifts ## Implementation ### Basic Implementation ```python class MistakeGatedOptimizer: def __init__(self, base_optimizer, memory_size=100): self.base_optimizer = base_optimizer self.error_memory = deque(maxlen=memory_size) def step(self, loss, pred, target): current_error = (pred != target).any() past_error = len(self.error_memory) > 0 if current_error or past_error: self.base_optimizer.step() # Update if current_error: self.error_memory.append((input, target)) else: pass # Skip update ``` ### Integration - Can be added to any optimizer in few lines - Compatible with SGD, Adam, etc. - Works with backpropagation ## Technical Specifications ### Performance - **Update Reduction**: 50-80% - **Accuracy**: Maintained or improved - **Overhead**: Negligible (<1% compute) ### Hyperparameters - **None added** to base optimizer - Optional: Memory buffer size - Optional: Past error lookback ## Applications ### Continual Learning Scenarios 1. **Class-Incremental Learning**: New classes added over time 2. **Task-Incremental Learning**: Different tasks sequentially 3. **Domain-Incremental Learning**: Same task, different distributions ### Hardware Deployment - **Neuromorphic systems**: Event-driven updates - **Edge devices**: Energy-constrained learning - **Real-time systems**: Low-latency inference ### Biological Plausibility - Implements error-driven learning - Consistent with ERN literature - Energy-efficient synaptic updates ## Comparison with Standard Methods | Method | Updates | Energy | Memory | Implementation | |--------|---------|--------|--------|----------------| | Standard | 100% | Baseline | Baseline | Simple | | Mistake-Gated | 20-50% | **-50-80%** | **Reduced** | **Simple** | | EWC | 100% | High | High | Complex | | Replay | 100% | Medium | High | Medium | ## Activation Keywords - mistake gating - continual learning - error-gated plasticity - energy efficient learning - error-related negativity - negativity bias - synaptic update reduction - biological plasticity ## Related Papers - Pache & van Rossum 2026: "Mistake gating leads to energy and memory efficient continual learning" (arXiv:2604.14336v1) ## References ```bibtex @article{pache2026mistake, title={Mistake gating leads to energy and memory efficient continual learning}, author={Pache, Aaron and van Rossum, Mark CW}, journal={arXiv preprint arXiv:2604.14336}, year={2026} } ``` --- _Last updated: 2026-04-17_