meta-learning-biological-plasticity

# Meta-Learning Biologically Plausible Plasticity Rules **Source:** arXiv:2210.16414v5 (February 2023) **Utility:** 0.90 **Authors:** Navid Shervani-Tabar et al. --- ## Description This skill meta-learns interpretable, biologically plausible plasticity rules that improve online learning with fixed random feedback connections (Feedback Alignment). It addresses the gap between backpropagation and biological synaptic plasticity. **Core Insight:** Random feedback alignment + meta-learned plasticity rules enable effective online learning without symmetric backward connections. --- ## Tools Used - `pytorch` - Neural network implementation - `meta_learning_framework` - Meta-optimization (MAML-like) - `feedback_alignment` - Random feedback connections - `plasticity_rules` - Custom learning rule implementations --- ## Instructions for Agents 1. **Define plasticity rule search space** - Parameterized update functions satisfying biological constraints 2. **Setup random feedback architecture** - Fixed random backward weights 3. **Meta-learn plasticity parameters** - Optimize for online learning performance across tasks 4. **Validate biological plausibility** - Check locality, no symmetric backward connections 5. **Test on target domains** - Evaluate in low-data regime, deep models --- ## Examples ### Example 1: Training Deep Network with Biological Constraints **User:** 如何在不使用反向传播的情况下训练深度神经网络？ **Agent:** 使用元学习的生物可塑性规则： 1. 设置固定随机反馈连接（替代反向传播的权重传输） 2. 定义局部可塑性规则（只依赖神经元局部信息） 3. 元学习优化规则参数 4. 结果：在线学习性能接近反向传播，但满足生物约束 **优势：** 无需对称反馈连接，适合在线学习。 ### Example 2: Low-Data Regime Learning **User:** 在数据有限的情况下，如何快速适应新任务？ **Agent:** 生物可塑性规则 + 元学习： - 元学习阶段：在多个任务上优化可塑性规则 - 应用阶段：使用优化后的规则快速适应新任务 - 低数据环境下表现优于标准反向传播 --- ## Activation Keywords - 生物可塑性、biological plasticity - 反馈对齐、feedback alignment - 元学习可塑性规则、meta-learning plasticity - 随机反馈、random feedback - 在线学习、online learning - 神经可塑性、synaptic plasticity --- ## Key Concepts ### 1. Feedback Alignment Problem - **Backpropagation:** Requires symmetric forward/backward weights - **Feedback Alignment:** Uses fixed random backward connections - **Challenge:** Shallow models work, deep models learn slowly ### 2. Meta-Learning Approach **Objective:** Discover plasticity rules that: - Use only local information (biological constraint) - Improve deep model training with random feedback - Enable efficient online learning **Method:** 1. Parameterize plasticity rule (weight update function) 2. Meta-optimize parameters across multiple tasks 3. Resulting rule generalizes to new tasks ### 3. Biological Plausibility Criteria | Criterion | Backprop | Meta-Learned Rule | |-----------|----------|-------------------| | Symmetric feedback | ❌ Required | ✅ Not needed | | Local computation | ❌ Global | ✅ Local | | Online learning | ❌ Batch | ✅ Online | | Weight transport | ❌ Required | ✅ Random fixed | --- ## When to Use 1. **Biologically constrained learning** - Models matching brain constraints 2. **Online learning systems** - Continual adaptation 3. **Low-data regimes** - Quick adaptation with few samples 4. **Neuromorphic hardware** - Local learning rules suitable for hardware --- ## Results (Paper) | Setting | Standard FA | Meta-Learned Rules | |---------|-------------|-------------------| | Shallow models | OK | Better | | Deep models | Slow | Faster convergence | | Online learning | Poor | Good | | Low-data regime | Poor | Good | --- ## Limitations 1. Still not as efficient as backpropagation for large datasets 2. Meta-learning phase requires multiple tasks 3. Random feedback quality affects performance 4. Deep model performance gap remains --- ## Related Skills - `decolle-snn-learning` - Deep continuous local learning - `neuromodulated-synaptic-plasticity` - Neuromodulation-based learning - `noisy-snn-learning` - Learning with noise in SNNs