mlx-snn: Spiking Neural Networks on Apple Silicon via MLX

2026-03-05 · Source: cs.NE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, long

Summary

mlx-snn is the first spiking neural network (SNN) library developed natively for Apple's MLX framework, addressing a gap for Apple Silicon users previously reliant on PyTorch or custom backends. The library offers six neuron models (LIF, IF, Izhikevich, Adaptive LIF, Synaptic, Alpha), four surrogate gradient functions, and four spike encoding methods, including an EEG-specific encoder. It integrates a complete backpropagation-through-time training pipeline, leveraging MLX's unified memory, lazy evaluation, and composable function transforms like mx.grad and mx.compile for efficiency. Benchmarking on MNIST digit classification shows mlx-snn achieving up to 97.28% accuracy, with 2.0–2.5x faster training and 3–10x lower GPU memory consumption compared to snnTorch on the same M3 Max hardware. The project is open-source under the MIT license and available via PyPI.

Key takeaway

For AI Scientists and Research Scientists developing Spiking Neural Networks on Apple Silicon, mlx-snn offers a significant performance advantage. You can achieve 2.0–2.5x faster training and 3–10x lower GPU memory usage compared to PyTorch-based libraries like snnTorch. Consider adopting mlx-snn to optimize your SNN development workflow and reduce reliance on external GPU infrastructure.

Key insights

mlx-snn provides a native, efficient SNN library for Apple Silicon, outperforming PyTorch-based alternatives in speed and memory.

Principles

• MLX-native tensor operations are crucial for performance.
• Explicit neuron state enables functional transforms.
• SNN research benefits from composable, JIT-compiled functions.

Method

mlx-snn implements a backpropagation-through-time training pipeline using an STE pattern for surrogate gradients, unrolling models over timesteps, and supporting various neuron models and spike encoding methods.

In practice

• Migrate existing snnTorch code using mlx-snn's compatible API.
• Utilize mlx-snn for SNN research on Apple Silicon without NVIDIA GPUs.
• Experiment with different spike encoding methods like rate or delta modulation.

Topics

• Spiking Neural Networks
• Apple MLX Framework
• Surrogate Gradient Learning
• SNN Libraries
• MNIST Classification

Code references

• D-ST-Sword/mlx-snn
• google/jax
• ml-explore/mlx
• lava-nc/lava
• norse/norse

Best for: AI Scientist, Research Scientist, AI Researcher, Machine Learning Engineer, AI Student

Related on AIssential

• Otters++: A Time-to-first-spike Based Energy Efficient Optical Spiking Transformer — Otters++ leverages optoelectronic device decay for energy-efficient time-to-first-spike computation in spiking Transformers.
• Hybrid ANN-SNN Pipeline with Local Plasticity — A hybrid ANN-SNN pipeline combines pretrained ANN embeddings and local SNN plasticity for high accuracy, bypassing end-to-end gradient training.
• Attacking the Spike: On the Transferability and Security of Spiking Neural Networks to Adversarial Examples — Adversarial attacks on SNNs are highly dependent on surrogate gradient estimators, and a new method, MDSE, exploits this for cross-model effectiveness.
• Linear Spiking Focus Network — The LSFNN achieves O(N) computational cost for SNNs by transforming an O(N²) focus field into a linear identity.
• SNN-MLIR: An MLIR Dialect for Compiling Neuromorphic SNNs from NIR to Bare-Metal C — SNN-MLIR unifies SNN compilation from NIR to bare-metal C, supporting both floating-point and quantized data.
• Memory-Augmented Spiking Networks: Synergistic Integration of Complementary Mechanisms for Neuromorphic Vision — Architectural balance of memory mechanisms in SNNs yields synergistic performance gains over individual optimizations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.NE updates on arXiv.org.