[Paper Review] LIAF-Net: Leaky Integrate and Analog Fire Network for Lightweight and Efficient Spatiotemporal Information Processing

Zhenzhi Wu, Hehui Zhang+4 more|arXiv (Cornell University)|Nov 12, 2020

Advanced Neural Network ApplicationsComputer Science65 references98 citations

TL;DR

This paper proposes LIAF-Net, a novel spiking neural network using a Leaky Integrate and Analog Fire (LIAF) neuron model that transmits analog values instead of binary spikes, preserving LIF's temporal dynamics while enabling efficient integration with artificial neural network (ANN) layers. LIAF-Net achieves state-of-the-art performance on DVS128 Gesture and other spatiotemporal datasets with significantly lower computational and parameter costs than LSTM, GRU, and ConvLSTM networks.

ABSTRACT

Spiking neural networks (SNNs) based on Leaky Integrate and Fire (LIF) model have been applied to energy-efficient temporal and spatiotemporal processing tasks. Thanks to the bio-plausible neuronal dynamics and simplicity, LIF-SNN benefits from event-driven processing, however, usually faces the embarrassment of reduced performance. This may because in LIF-SNN the neurons transmit information via spikes. To address this issue, in this work, we propose a Leaky Integrate and Analog Fire (LIAF) neuron model, so that analog values can be transmitted among neurons, and a deep network termed as LIAF-Net is built on it for efficient spatiotemporal processing. In the temporal domain, LIAF follows the traditional LIF dynamics to maintain its temporal processing capability. In the spatial domain, LIAF is able to integrate spatial information through convolutional integration or fully-connected integration. As a spatiotemporal layer, LIAF can also be used with traditional artificial neural network (ANN) layers jointly. Experiment results indicate that LIAF-Net achieves comparable performance to Gated Recurrent Unit (GRU) and Long short-term memory (LSTM) on bAbI Question Answering (QA) tasks, and achieves state-of-the-art performance on spatiotemporal Dynamic Vision Sensor (DVS) datasets, including MNIST-DVS, CIFAR10-DVS and DVS128 Gesture, with much less number of synaptic weights and computational overhead compared with traditional networks built by LSTM, GRU, Convolutional LSTM (ConvLSTM) or 3D convolution (Conv3D). Compared with traditional LIF-SNN, LIAF-Net also shows dramatic accuracy gain on all these experiments. In conclusion, LIAF-Net provides a framework combining the advantages of both ANNs and SNNs for lightweight and efficient spatiotemporal information processing.

Motivation & Objective

• To address the performance degradation in traditional LIF-based spiking neural networks (SNNs) caused by binary spike transmission.
• To enable seamless integration between SNNs and artificial neural networks (ANNs) by replacing binary spikes with analog activations.
• To maintain bio-plausible temporal processing dynamics while improving information transmission fidelity.
• To reduce computational and storage overhead in spatiotemporal learning tasks compared to recurrent and 3D convolutional networks.

Proposed method

• Proposes the Leaky Integrate and Analog Fire (LIAF) neuron model, which retains LIF’s temporal dynamics but uses analog values for neural activation.
• Enables analog activation transmission between neurons, improving information fidelity over binary spikes in LIF-SNNs.
• Supports joint training with standard ANN layers (e.g., ReLU, batch normalization, residual connections) via compatible activation formats.
• Integrates spatial processing via convolutional or fully connected layers, enabling end-to-end training with standard deep learning frameworks like TensorFlow.
• Uses time-step-based dynamics where each neuron’s membrane potential evolves according to a leaky integrate-and-fire mechanism with analog output.
• Employs backpropagation through time with differentiable analog outputs, enabling effective training of deep spatiotemporal architectures.

Experimental results

Research questions

• RQ1 Can analog activation transmission improve performance in spiking neural networks without sacrificing temporal processing capability?
• RQ2 Can LIAF-Net achieve high accuracy on spatiotemporal tasks while maintaining low computational and memory costs compared to LSTM and GRU?
• RQ3 How does the integration of LIAF with standard ANN components (e.g., normalization, skip connections) affect training stability and performance?
• RQ4 To what extent can LIAF-Net outperform existing SNNs and hybrid networks on dynamic vision sensor (DVS) datasets?
• RQ5 Is the LIAF model biologically plausible, given evidence of analog action potentials in cortical neurons?

Key findings

• LIAF-Net achieves 97.56% test accuracy on the DVS128 Gesture dataset, outperforming the next best method (Bi et al., 2019) by 2.24 percentage points.
• LIAF-Net reduces computational overhead by 91.6% and parameter storage by 90.0% compared to ConvLSTM-based networks on the same dataset.
• On the bAbI QA dataset, LIAF-Net matches the performance of GRU and LSTM networks while using significantly fewer synaptic weights and less computation.
• LIAF-Net achieves state-of-the-art performance on MNIST-DVS, CIFAR10-DVS, and DVS128 Gesture datasets, surpassing both traditional LIF-SNNs and SNNs with spike-based communication.
• The LIAF model is shown to be biologically plausible, with evidence of analog action potentials in human cortical pyramidal neurons supporting its feasibility.
• LIAF-Net enables direct training with standard deep learning frameworks, eliminating the need for spike coding format conversion and enabling use of modern training techniques.