PC-SNN: Predictive Coding-based Local Hebbian Plasticity Learning in Spiking Neural Networks

TL;DR

The paper addresses the challenge of training Spiking Neural Networks with biologically plausible, locally computed learning rules. It introduces PC-SNN, which integrates predictive coding with local Hebbian plasticity to enable end-to-end learning without backpropagation, and extends this framework to regression via ASRNN-PCTT. Empirical results on Caltech Face/Motorbike, MNIST, NM-NIST, CIFAR-10, and event-camera angular-velocity data demonstrate competitive classification performance and robust regression accuracy, while highlighting improvements in learning efficiency and hardware suitability. The work advances neuromorphic computing by providing a theoretically grounded, biologically plausible alternative to backpropagation that preserves performance and enables scalable, parallelizable implementations.

Abstract

Spiking Neural Networks (SNNs), regarded as the third generation of neural networks, emulate the brain's information processing with unparalleled biological plausibility compared to traditional neural networks. However, their non-linear, event-driven dynamics pose significant challenges for training, and existing methods often deviate from neuroscientific principles of cortical learning. Drawing inspiration from predictive coding theory-a leading model of brain information processing-we propose PC-SNN, a novel learning framework that integrates predictive coding with SNNs to enable biologically plausible, local Hebbian plasticity without reliance on backpropagation. Unlike conventional SNN training approaches, PC-SNN leverages only local computations, aligning with the brain's distributed processing and overcoming the biological implausibility of global error propagation. Our classification model achieves competitive performance on the benchmark datasets, including Caltech Face/Motorbike, MNIST, and CIFAR10, surpassing state-of-the-art multi-layer SNNs. Furthermore, our predictive coding-based regression model outperforms backpropagation-based methods while adhering to local plasticity constraints, offering a scalable and biologically grounded alternative for SNN training. PC-SNN drives progress in neuromorphic computing through validating the adaptability of bio-inspired algorithms within spiking neural architectures, but also unveils novel understandings of neurocognitive learning processes, presenting a conceptual framework distinguished by its theoretical originality and functional efficacy.

Figures (8)

• Figure 1: (A): SNN trained with BP (BP-SNN). It shows a two-layer SNN, and the input layer encodes pixel intensity values into spike trains using a temporal encoding scheme. IF neurons within the hidden and output layers process these received spike trains to transmit information. The network learns via a backpropagation algorithm that compares the actual output firing times against target firing times and propagates the resulting error gradient backward through the network's layers. The effective backpropagation of the intermediate gradient signal from $\delta_{k}^{l+1}$ to $\delta_{j}^{l}$ necessitates the construction of a feedback structure. This implies a critical constraint: each weight in the feedback network must correspond symmetrically to its counterpart in the feed-forward network. (B): Predictive coding structure of SNN (PC-SNN). The update mechanism for both predictions and prediction errors operates in a parallel manner, exclusively utilizing local information. Note: Excitation denotes additive operator and Inhibition denotes subtraction operator.
• Figure 2: An diagram illustration between ASRNN-BPTT and ASRNN-PCTT.
• Figure 3: The temporal dynamics of membrane potentials for output neurons were recorded in response to a subset of face and motorbike images from the Caltech101 dataset. Arrows denote the precise timing of action potentials for each corresponding neuron.
• Figure 4: This figure presents a comparative analysis of the firing time distributions for two output neurons in response to face and motorbike stimuli. The data are depicted for two distinct conditions: (A) prior to training and (B) subsequent to training.
• Figure 5: This figure presents a histogram comparing the firing time distributions for two distinct neuronal populations in response to the training dataset. The blue and red bars represent the winner and non-winner neurons, respectively. Additionally, the mean firing time for each respective neuron group is denoted by a dashed line.