Online Pseudo-Zeroth-Order Training of Neuromorphic Spiking Neural Networks

The paper proposes OPZO, a training method for spiking neural networks that uses noise injection and direct top-down signals for spatial credit assignment, avoiding symmetric weight transport and separate forward-backward phases. It sits in the 'Zeroth-Order and Feedback-Based Approximations' leaf, which contains only three papers total, indicating a relatively sparse research direction within the broader field of biologically plausible SNN training. This leaf focuses specifically on methods that replace explicit gradient backpropagation with noise-based or feedback mechanisms, distinguishing it from the more populated gradient-adjustment approaches in the sibling leaf.

The taxonomy reveals that OPZO's parent branch, 'Gradient-Based and Backpropagation Alternatives', contains two main directions: spatiotemporal gradient adjustment methods and zeroth-order/feedback approaches. The sibling leaf on gradient adjustment includes techniques that still compute explicit gradients but modify their flow, whereas OPZO's leaf emphasizes avoiding gradient computation entirely. Neighboring branches in 'Learning Rule Design' include local Hebbian mechanisms and reinforcement learning methods, which differ fundamentally by relying on unsupervised correlation-based rules or reward signals rather than supervised error-driven updates. The taxonomy's scope notes clarify that OPZO belongs here because it uses feedback signals without explicit gradient backpropagation, not in the Hebbian category.

Among the three contributions analyzed across twenty-eight candidate papers, the pseudo-zeroth-order formulation examined eight candidates with none providing clear refutation, while the OPZO training method with momentum feedback examined ten candidates, also without refutation. The biologically plausible on-chip training framework examined ten candidates and found one that appears to provide overlapping prior work. This suggests that the core algorithmic innovations appear relatively novel within the limited search scope, while the on-chip training framing may have more substantial precedent. The analysis explicitly covers top-K semantic matches and citation expansion, not an exhaustive literature review.

Based on the limited search of twenty-eight candidates, the work appears to occupy a sparsely populated research direction with modest prior overlap. The core pseudo-zeroth-order formulation and momentum feedback mechanisms show no clear refutation among examined candidates, though the on-chip training motivation has at least one overlapping prior work. The taxonomy structure suggests this is an emerging area within biologically plausible SNN training, though definitive novelty claims would require broader literature coverage beyond the semantic search scope employed here.