neurotrain-local-learning-snn-benchmarking

name

neurotrain-local-learning-snn-benchmarking

description

NeuroTrain: Survey and benchmarking framework for SNN local learning rules. Comprehensive taxonomy of SNN training algorithms spanning surrogate-gradient backpropagation, local/three-factor learning, biologically inspired plasticity, ANN-to-SNN conversion, and non-standard optimization. Includes open-source benchmarking framework built on snnTorch for reproducible cross-method comparison. Activation: neurotrain, SNN training survey, spiking neural network benchmark, local learning rules, SNN taxonomy, snnTorch benchmarking framework, eligibility traces, STDP-inspired learning, surrogate gradient SNN

NeuroTrain: Local Learning Rules for SNNs

Paper: arXiv:2605.15058v1 (May 14, 2026) Authors: Alessio Caviglia, Filippo Marostica, Roberta Bardini, Alessandro Savino, Stefano Di Carlo (Politecnico di Torino) GitHub: https://github.com/smilies-polito/neurotrain

Core Problem

SNN training is challenging due to: (1) non-differentiable spike events requiring surrogate gradients, (2) temporal credit assignment across many timesteps, (3) computational/memory costs of BPTT that map poorly to neuromorphic hardware. The field lacks a unified taxonomy and standardized benchmarking framework.

Taxonomy Architecture

Primary Axis: Training Strategy

1. Direct Training — learn while simulating spiking dynamics (surrogate-gradient BPTT, local/three-factor rules, STDP variants)
2. ANN-to-SNN Conversion — train ANN first, then map to SNN
3. Evolutionary/Population-based — black-box optimization without gradients

Secondary Axis: Learning Signal

• Supervised — labeled targets
• Unsupervised — input structure-driven objectives
• Reinforcement Learning — reward-driven delayed signals

Locality Dimensions

• Temporal Locality — updates depend only on current timestep info (+ bounded state like eligibility traces)
• Spatial Locality — updates use only synapse-local variables (pre/post activity + modulatory signal)

Recurring Mechanisms (cross-cutting)

1. Eligibility Traces — Δw ∝ e · M (three-factor rule form)
2. Direct Feedback Alignment (DFA) — random projection of output error to hidden layers
3. Direct Random Target Projection (DRTP) — direct target projection, no output error needed
4. Auxiliary Local Classifiers — per-layer loss/readout heads
5. STDP-Inspired — spike-timing correlations adapted to supervised/RL objectives
6. Spatial BP + Online Temporal — spatial backprop per timestep, temporal credit online

Key SNN Training Algorithms

Supervised Direct Training

Algorithm	Temporal Loc	Spatial Loc	Traces	STDP	DFA/DRTP	Local Clf
BPTT	✗	✗
E-prop	✓	✗	✓
DECOLLE	✓	✓	✓			✓
OTTT	✓	✗	✓
ESD-RTRL	✓	✓	✓
Target Propagation	✗	✗			✓
ETLP	✓	✓	✓
STDP-like supervised	✓	✓		✓

Unsupervised

Algorithm	Temporal Loc	Spatial Loc	Traces	STDP
STDP variants	✓	✓		✓
Hebbian learning	✓	✓		✓
BCM rule	✓	✓

NeuroTrain Framework Architecture

Repository

• GitHub: https://github.com/smilies-polito/neurotrain
• License: CC BY 4.0

Three modular components built on snnTorch + PyTorch:

1. Dataloaders — standardized interface for neuromorphic/rate-encoded datasets (NeuroBench, Tonic)
2. Models — library of benchmark SNNs (FC, Conv, Recurrent, VGG-style, LIF-based)
3. Trainers — Python classes implementing specific learning rules with common interface

Execution Modes

• Campaign mode — orthogonal combination of trainers × models × datasets with auto-hyperparameter tuning (Optuna)
• Custom mode — predefined single experiment for targeted evaluation

Reported Metrics

• Test/train accuracy, loss, execution time, time per epoch
• Parameters count, memory footprint, activation sparsity

Benchmark Results Summary

Representative results across 8 datasets (MNIST, Fashion-MNIST, CIFAR-10, SVHN, N-MNIST, DVS Gesture, DVS CIFAR-10, SHD):

• Best performers: BPTT and Target Propagation on simple tasks (MNIST ~98%)
• Most hardware-friendly: DECOLLE, E-prop, OSTL (local in time+space)
• Trade-off: Local rules sacrifice some accuracy for neuromorphic compatibility
• ~850 GPU hours for comprehensive campaign

Key Insights

1. Many training rules exhibit limited portability across datasets/architectures
2. Fair benchmarking requires unified hyperparameter optimization, not fixed settings
3. Stronger convergence needed between algorithmic benchmarking and hardware-aware evaluation
4. NeuroTrain designed as living, community-driven resource

Activation Context

Use this skill when:

• Designing or comparing SNN training algorithms
• Understanding locality trade-offs in spiking network learning
• Building reproducible SNN benchmarks
• Selecting training methods for neuromorphic hardware deployment
• Studying three-factor learning rules, eligibility traces, or STDP variants