Paper: *Scalable Learning in Structured Recurrent Spiking Neural Networks without Backpropagation* Authors: Bo Tang, Weiwei Xie arXiv: 2605.00402 | 2026-04-29
Why Deep Learning Doesn't Resemble the Brain
The post contrasts biological learning with conventional deep learning:
- The brain: neurons communicate via spikes, consume very little energy, learn locally without global backpropagation, and naturally handle temporal data.
- Traditional deep learning: continuous activations, high energy consumption, and a global backpropagation pass — fundamentally unlike biological mechanisms.
- Non-differentiability: spikes are discrete (0 or 1), so the spike function has no gradient and standard backpropagation cannot be applied directly.
- Problems with backpropagation: it requires storing all intermediate activations, is memory- and compute-intensive, and is not biologically plausible.
- Scalability: deep recurrent SNNs suffer from vanishing/exploding gradients and become even harder to train with sparse connectivity.
- Biological plausibility: the brain has no backpropagation; local rules offer models for understanding how the brain learns.
- Efficiency: purely local computation, no stored intermediate states, well suited to hardware implementation and edge devices.
- Scalability: deeper networks and sparser connectivity remain trainable.
Spiking neural networks (SNNs) are much closer to the brain — but they are notoriously hard to train.
The Training Dilemma for SNNs
Structured Recurrent SNNs + Backpropagation-Free Learning
The paper's central idea: design a structured SNN architecture that can be trained with learning rules that do not require backpropagation.
Key components:
1. Structured architecture — locally dense recurrent layers combined with sparse "small-world" long-range projections, mimicking brain structure. 2. Backpropagation-free learning — local plasticity rules such as STDP (spike-timing-dependent plasticity) that depend only on local neural activity, with no global gradients. 3. Scalability — deep recurrent structure and sparse connections that still learn effectively. 4. Energy efficiency — event-driven spiking computation consumes energy only when spikes occur, potentially saving orders of magnitude versus conventional networks.
Instead of a "central command" computing gradients, each neuron adjusts its connections based on local activity — simple, efficient, and parallel.
Why Backpropagation-Free SNNs Are Attractive
Takeaways
For researchers in neural networks or edge AI, the post suggests asking:
1. Is my model overly dependent on backpropagation? 2. Could local learning rules achieve acceptable results? 3. Would bio-inspired methods bring efficiency gains? 4. Is spiking computation a fit for my application?
Core message: the strongest learning method is not necessarily the best — the most natural one may be more efficient and scalable. When SNNs break free from backpropagation, they move closer not only to the brain computationally, but to the essence of intelligence philosophically. In the spiking world, energy efficiency and intelligence can coexist.
> Paper: arXiv: 2605.00402 — *Scalable Learning in Structured Recurrent Spiking Neural Networks without Backpropagation*