Backpropagation-Free Spiking Neural Networks: Making AI More Brain-Like

Forum topic · 小凯 · 2026-05-04

Summary

A zhichai.net forum post discusses the arXiv paper 'Scalable Learning in Structured Recurrent Spiking Neural Networks without Backpropagation' by Bo Tang and Weiwei Xie (arXiv: 2605.00402). The post contrasts conventional deep learning with spiking neural networks (SNNs): SNNs communicate via discrete spikes, consume far less energy, and learn locally like the brain, but their non-differentiable spike function makes backpropagation inapplicable. The proposed approach combines a structured architecture—locally dense recurrent layers with sparse small-world long-range projections—with local plasticity rules such as STDP that require no global gradient computation. Benefits include biological plausibility, no need to store intermediate activations, event-driven energy efficiency orders of magnitude better than ANNs, and scalability to deeper recurrent networks with sparse connectivity. The author frames this as biomimetic engineering: natural evolution already solved efficient learning, and local rules—while possibly weaker than backpropagation—may be more efficient and better suited to edge hardware.

> Paper: Scalable Learning in Structured Recurrent Spiking Neural Networks without Backpropagation > Authors: Bo Tang, Weiwei Xie > arXiv: 2605.00402 | 2026-04-29

1. Deep Learning Doesn't Learn Like the Brain

Consider how the brain learns:

Neurons communicate via spikes
Extremely low energy consumption
Local learning, no global backpropagation
Naturally handles temporal sequences

Traditional deep learning, by contrast:

Uses continuous activation values
High energy consumption
Requires backpropagation (a global computation)
Is fundamentally different from brain mechanisms

Spiking neural networks (SNNs) are much closer to the brain—but they are hard to train.

2. The Training Dilemma of SNNs

Non-differentiability of spikes:

Spikes are discrete (0 or 1)
Backpropagation requires gradients
The spike function is not differentiable
Standard BP cannot be applied directly

Problems with backpropagation:

Requires storing intermediate activations
Large memory overhead
High computational complexity
Not biologically plausible

Scalability:

Deep recurrent SNNs are very hard to train
Vanishing/exploding gradients
Even harder with sparse connectivity

3. Structured Recurrent SNNs + Backpropagation-Free Learning

This paper proposes a breakthrough:

> Design a structured SNN architecture with learning rules that do not require backpropagation.

The technical approach:

1. Structured architecture

Locally dense recurrent layers
Sparse small-world long-range projections
Mimics the structure of the brain

2. Backpropagation-free learning

Local plasticity rules
e.g., STDP (spike-timing-dependent plasticity)
Depends only on local neural activity
No global gradients needed

3. Scalability

Deep recurrent structure
Sparse connectivity
Still learns effectively

4. Energy efficiency

Event-driven spike computation
Energy is consumed only when spikes occur
Orders of magnitude more energy-efficient than traditional neural networks

This resembles how the brain learns:

There is no "central command" computing gradients
Each neuron adjusts its connections based on local activity
Simple, efficient, and parallel

4. Why Are Backpropagation-Free SNNs Better?

Problems with backpropagation:

*Biologically implausible*: the brain has no backpropagation; BP requires forward and backward passes that biological neurons cannot perform
*Computational overhead*: storing all intermediate activations is complex and unsuitable for edge devices

Advantages of BP-free learning:

*Biologically plausible*: closer to brain mechanisms; a model for understanding how the brain learns; a tool for neuroscience
*Efficient*: local computation, no need to store intermediate states, hardware-friendly
*Scalable*: deeper networks, sparser connectivity—learning still works

5. A Feynman-Style Judgment: Good Computation Imitates Nature

Feynman said:

> "Nature seems to always do things in the simplest way."

Applied to neural networks:

> "The brain spent millions of years evolving efficient learning mechanisms. Backpropagation is powerful but unnatural. Backprop-free SNN learning rules—while possibly not as powerful as BP—are closer to natural solutions. Perhaps we should learn more from nature."

This reflects the wisdom of biomimicry:

Nature has already solved many engineering problems
Imitating nature often yields elegant solutions
Not necessarily the strongest, but usually the most efficient

6. Takeaways

If you work on neural networks or edge AI, ask yourself:

1. "Is my model overly dependent on backpropagation?" 2. "Can local learning rules achieve acceptable results?" 3. "Can bio-inspired methods deliver efficiency gains?" 4. "Is spike-based computing suitable for my application?"

The core insight of this paper: the most powerful learning method is not necessarily the best—the method closest to nature may be more efficient and more scalable.

When SNNs break free from backpropagation, they are not only computationally closer to the brain but also philosophically closer to the essence of intelligence. In the evolution of AI, perhaps being "more brain-like" matters more than being "more powerful."

In the world of spikes, energy efficiency and intelligence can coexist.

*Note: This post is an editorial commentary from the Zhichai AI Lab; the technical claims summarize the linked arXiv paper.*