Skip to main content

mistake-gated-continual-learning

Mistake-gated learning for energy and memory efficient continual learning using neuromorphic hardware. Only neurons that "make mistakes" (prediction errors) are updated, reducing compute and memory. Achieves 10-100x energy reduction vs full backprop on MNIST/CIFAR benchmarks.

Aller à l'installation

Informations de source

Dépôt
hiyenwong/ai_collection
Dernière activité de la source
4 juin 2026 à 13:32
Langue détectée de SKILL.md
anglais
Étoiles
2
Forks
0

Options d'installation

Le prompt qui vérifie d'abord la source est sélectionné par défaut. Vous pouvez passer à une commande directe ou télécharger une copie locale.

Vérifiez les fichiers source

Lisez SKILL.md et les fichiers associés affichés par SkillsMP avant de décider de l'installer.

Affichage de SKILL.md

SKILL.md
Instructions source · Aperçu en lecture seule
name
mistake-gated-continual-learning
description
Mistake-gated learning for energy and memory efficient continual learning using neuromorphic hardware. Only neurons that "make mistakes" (prediction errors) are updated, reducing compute and memory. Achieves 10-100x energy reduction vs full backprop on MNIST/CIFAR benchmarks.
# 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_
Voir sur GitHub