LIAF-Net: Leaky Integrate and Analog Fire Network for Lightweight and Efficient Spatiotemporal Information...

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking neural networks (SNNs) using the Leaky Integrate and Fire (LIF) model offer energy-efficient processing but often suffer from reduced performance due to spike-based communication.
• Traditional SNNs face challenges in achieving performance comparable to Artificial Neural Networks (ANNs) for complex spatiotemporal tasks.

Purpose of the Study:

• To introduce a novel Leaky Integrate and Analog Fire (LIAF) neuron model that transmits analog values, enhancing information processing capabilities.
• To develop LIAF-Net, a deep network utilizing the LIAF model for efficient spatiotemporal data processing.
• To enable direct training of the network using backpropagation through time (BPTT), avoiding performance degradation from ANN-to-SNN conversion.

Main Methods:

• Proposed a Leaky Integrate and Analog Fire (LIAF) neuron model capable of integrating spatial information via convolutional or fully connected layers.
• Developed LIAF-Net, a deep network architecture incorporating the LIAF neuron model.
• Trained LIAF-Net directly using backpropagation through time (BPTT) for spatiotemporal tasks.

Main Results:

• LIAF-Net demonstrated comparable performance to Gated Recurrent Units (GRUs) and Long Short-Term Memory (LSTMs) on question answering tasks.
• Achieved state-of-the-art results on spatiotemporal datasets like MNIST-DVS, CIFAR10-DVS, and DVS128 Gesture.
• Showcased significantly reduced synaptic weights and computational overhead compared to LSTM, GRU, ConvLSTM, and Conv3D networks.
• Exhibited substantial accuracy improvements over traditional LIF-SNNs on all tested experiments.

Conclusions:

• LIAF-Net effectively combines the strengths of ANNs and SNNs, providing a framework for lightweight and efficient spatiotemporal information processing.
• The proposed LIAF model and LIAF-Net architecture represent a significant advancement in SNN performance and applicability.
• This approach offers a promising alternative for energy-efficient and high-performance spatiotemporal data analysis.