Backpropagation-free Spiking Neural Networks with the Forward-Forward Algorithm

TL;DR

This paper tackles the challenge of training Spiking Neural Networks (SNNs) without backpropagation by adopting the Forward-Forward (FF) algorithm, which uses two forward passes to create layer-wise, local learning objectives. It introduces label-embedding to generate positive and negative samples, leverages leaky integrate-and-fire (LIF) neurons with per-layer normalization, and optimizes a Swish-inspired loss based on layer-wise goodness scores computed from spike counts. Empirical results on static benchmarks (MNIST, Fashion-MNIST, Kuzushiji-MNIST) and spiking benchmarks (Neuromorphic-MNIST, SHD) show competitive or superior accuracy with lighter architectures and as few as 10 time steps, often outperforming FF-based ANNs and approaching BP-trained SNN performance. The findings indicate that FF training can deliver biologically plausible, energy-efficient SNNs suitable for neuromorphic hardware, with strong potential for online learning and broader applicability.

Abstract

Spiking Neural Networks (SNNs) offer a biologically inspired computational paradigm that emulates neuronal activity through discrete spike-based processing. Despite their advantages, training SNNs with traditional backpropagation (BP) remains challenging due to computational inefficiencies and a lack of biological plausibility. This study explores the Forward-Forward (FF) algorithm as an alternative learning framework for SNNs. Unlike backpropagation, which relies on forward and backward passes, the FF algorithm employs two forward passes, enabling layer-wise localized learning, enhanced computational efficiency, and improved compatibility with neuromorphic hardware. We introduce an FF-based SNN training framework and evaluate its performance across both non-spiking (MNIST, Fashion-MNIST, Kuzushiji-MNIST) and spiking (Neuro-MNIST, SHD) datasets. Experimental results demonstrate that our model surpasses existing FF-based SNNs on evaluated static datasets with a much lighter architecture while achieving accuracy comparable to state-of-the-art backpropagation-trained SNNs. On more complex spiking tasks such as SHD, our approach outperforms other SNN models and remains competitive with leading backpropagation-trained SNNs. These findings highlight the FF algorithm's potential to advance SNN training methodologies by addressing some key limitations of backpropagation.

Figures (5)

• Figure 1: Different options for overlaying a label on a sample of a 10-class dataset for positive and negative sample generation.
• Figure 2: The architecture of the proposed spiking forward-forward algorithm. The diagram highlights the input forward path (solid arrows) and the error propagation path (dotted arrows).
• Figure 3: Loss function $\frac{-\alpha\Delta}{1 + e^{\alpha\Delta}}$ plotted against $\Delta$ for different scaling factor values of $\alpha$.
• Figure 4: Loss trend during training of our proposed FF-based spiking neural network model through epochs. (a) Loss changes for non-spiking datasets such as MNIST, F-MNIST, and K-MNIST. (b) Loss changes for spiking datasets such as N-MNIST and SHD.
• Figure 5: Comparison of the accuracy and parameter count between our proposed SNN model and other ANN models based on the Forward-Forward algorithm. (a) Accuracy of our proposed model compared to other FF-based models trained with ANN. (b) Parameters count of our proposed model compared to other FF-based models trained with ANN.