Abstract:Spiking neural network (SNN) is a brain-inspired model with more spatio-temporal information processing capacity and computational energy efficiency. However, with the increasing depth of SNNs, the memory problem caused by the weights of SNNs has gradually attracted attention. In this study, we propose an ultra-low latency adaptive local binary spiking neural network (ALBSNN) with accuracy loss estimators, which dynamically selects the network layers to be binarized to ensure a balance between quantization degree and classification accuracy by evaluating the error caused by the binarized weights during the network learning process. At the same time, to accelerate the training speed of the network, the global average pooling (GAP) layer is introduced to replace the fully connected layers by combining convolution and pooling. Finally, to further reduce the error caused by the binary weight, we propose binary weight optimization (BWO), which updates the overall weight by directly adjusting the binary weight. This method further reduces the loss of the network that reaches the training bottleneck. The combination of the above methods balances the network's quantization and recognition ability, enabling the network to maintain the recognition capability equivalent to the full precision network and reduce the storage space by more than 20%. So, SNNs can use a small number of time steps to obtain better recognition accuracy. In the extreme case of using only a one-time step, we still can achieve 93.39, 92.12, and 69.55% testing accuracy on three traditional static datasets, Fashion- MNIST, CIFAR-10, and CIFAR-100, respectively. At the same time, we evaluate our method on neuromorphic N-MNIST, CIFAR10-DVS, and IBM DVS128 Gesture datasets and achieve advanced accuracy in SNN with binary weights. Our network has greater advantages in terms of storage resources and training time.

Weightless for neural networks based on multi-valued probabilistic logic for node for pattern recognition

Sparse Bayesian Approach to Fast Learning Network for Multiclassification.

WeightNet: Revisiting the Design Space of Weight Networks

Structure and Weight Optimization of Neural Network Based on CPA-MLR and Its Application in Naphtha Dry Point Soft Sensor.

Deep Spiking Neural Networks with Binary Weights for Object Recognition

A Closed-form Solution for Weight Optimization in Fully-connected Feed-forward Neural Networks

Explore the Knowledge contained in Network Weights to Obtain Sparse Neural Networks

A Lightweight Multi-Dendritic Pyramidal Neuron Model with Neural Plasticity on Image Recognition

Differentiable Weightless Neural Networks

Analytical Uncertainty-Based Loss Weighting in Multi-Task Learning

Weightless: Lossy Weight Encoding For Deep Neural Network Compression

Multi-gate Weighted Fusion Network for neuronal morphology classification

Unbiased Weight Maximization

Securing Neural Networks with Knapsack Optimization

ALBSNN: ultra-low latency adaptive local binary spiking neural network with accuracy loss estimator

Learning to Modulate Random Weights: Neuromodulation-inspired Neural Networks For Efficient Continual Learning

A Probabilistic Formulation for Meta-Weight-Net

Neural Network with Multiple Connection Weights.

MLDS: A Dataset for Weight-Space Analysis of Neural Networks

STDP Based Pruning of Connections and Weight Quantization in Spiking Neural Networks for Energy Efficient Recognition

Pattern Recognition with Weighted Complex Networks.