Biologically Plausible Learning via Bidirectional Spike-Based Distillation

TL;DR

Bidirectional Spike-Based Distillation (BSD) presents a biologically grounded learning framework that trains two spike-based networks in tandem to perform perception (stimulus-to-concept) and recall (concept-to-stimulus) via mutual distillation. By enforcing five plausibility criteria—including asymmetric forward/backward weights and unsigned, local learning via the Relaxed Contrastive (ReCo) loss—BSD achieves performance on par with backpropagation across classification, sequential prediction, and generation tasks, while preserving spike-based, local computation. The method leverages a three-compartment pyramidal neuron model, discrete spike communication, and FFT-based frequency-aware losses to balance structural fidelity and detail in generation. Empirical results showBSD’s robustness across datasets, strong convergence properties, and a favorable energy profile during inference on SNN hardware, underscoring the practical relevance of spike-driven, biologically faithful learning.

Abstract

Developing biologically plausible learning algorithms that can achieve performance comparable to error backpropagation remains a longstanding challenge. Existing approaches often compromise biological plausibility by entirely avoiding the use of spikes for error propagation or relying on both positive and negative learning signals, while the question of how spikes can represent negative values remains unresolved. To address these limitations, we introduce Bidirectional Spike-based Distillation (BSD), a novel learning algorithm that jointly trains a feedforward and a backward spiking network. We formulate learning as a transformation between two spiking representations (i.e., stimulus encoding and concept encoding) so that the feedforward network implements perception and decision-making by mapping stimuli to actions, while the backward network supports memory recall by reconstructing stimuli from concept representations. Extensive experiments on diverse benchmarks, including image recognition, image generation, and sequential regression, show that BSD achieves performance comparable to networks trained with classical error backpropagation. These findings represent a significant step toward biologically grounded, spike-driven learning in neural networks. Our code is available at https://github.com/alden199/Bidirectional-Spike-Based-Distillation.

Figures (11)

• Figure 1: (a) Overview of the bidirectional spike-based distillation framework. (b) Illustration of feature alignment using the properties of pyramidal neurons, which receive feedforward and feedback signals through their basal and apical dendrites, respectively. (c) Performance comparison of neural networks trained with BSD, dendritic localized learning (DLL), a recently proposed biologically plausible learning algorithm, and backpropagation. The experimental results show that BSD achieves performance comparable to backpropagation.
• Figure 2: Results on image generation tasks using autoencoders trained with backpropagation and BSD. GT denotes the ground truth, BP indicates backpropagation, and BSD refers to our proposed method. The reconstruction quality demonstrates that BSD attains performance comparable to backpropagation on generation tasks, while preserving biological plausibility
• Figure 4: Comparison of image reconstruction on the MNIST dataset by autoencoders trained with Backpropagation (BP) and our proposed Bidirectional Spike-based Distillation (BSD).
• Figure 5: Visual comparison of FashionMNIST reconstructions from autoencoders trained with Backpropagation (BP) and our Bidirectional Spike-based Distillation (BSD) method..
• Figure 6: Visual examples of image reconstructions on the CIFAR-10 dataset, showcasing the performance of autoencoders trained with BSD using various layer-wise loss function configurations.