TDSNN: From Deep Neural Networks to Deep Spike Neural Networks with Temporal-Coding

Abstract

Continuous-valued deep convolutional networks (DNNs) can be converted into accurate rate-coding based spike neural networks (SNNs). However, the substantial computational and energy costs, which is caused by multiple spikes, limit their use in mobile and embedded applications. And recent works have shown that the newly emerged temporal-coding based SNNs converted from DNNs can reduce the computational load effectively. In this paper, we propose a novel method to convert DNNs to temporal-coding SNNs, called TDSNN. Combined with the characteristic of the leaky integrate-andfire (LIF) neural model, we put forward a new coding principle Reverse Coding and design a novel Ticking Neuron mechanism. According to our evaluation, our proposed method achieves 42% total operations reduction on average in large networks comparing with DNNs with no more than 0.5% accuracy loss. The evaluation shows that TDSNN may prove to be one of the key enablers to make the adoption of SNNs widespread.

T2FSNN: Deep Spiking Neural Networks with Time-to-first-spike Coding

“…With the conversion, deep SNNs achieved comparable results to DNNs in various applications, including image classification [23] and multi-object detection [25]. However, conversion approaches are limited in efficiency [31], which is defined with latency and the number of spikes, even if the converted SNNs adopted several temporal coding schemes, such as TTFS coding [16,17].…”

“…Neural coding defines the way of representing the information in the form of spike trains including encoding and decoding function [22]. There have been various types of neural coding, such as rate [23,24,25], phase [26], burst [27], temporal-switching-coding (TSC) [28], and TTFS coding [12,16,17,15]. To maximize the efficiency by fully utilizing the temporal information in spike train, TTFS coding, which is known as latency coding, was introduced in SNNs [12].…”

“…Recently, several studies successfully adopted TTFS coding to deep SNNs [16,17,15]. According to [16,17], to implement TTFS coding in deep SNNs, there should be dependencies between integration and fire phases of subsequent layers. The integration phase, which corresponds to the decoding function, can be implemented with an exponentially decaying synaptic kernel κ, which is described as…”

“…SNNs have a feature of temporal processing with spiking neurons, which requires computations depending on time steps and is executed iteratively as recurrent neural networks (RNNs). Furthermore, in TTFS coding, the inference latency is proportional to the depth of neural networks and is usually more than several hundreds time steps in deep SNNs [16,17]. For these reasons, the direct training with surrogate gradients of SNNs demands more than LT time computations compared to DNNs, where L is the number of layers and T is time window of each layer.…”

“…These approach using approximate differentiable spiking function enables efficiency-aware training of deep SNNs. In another approach to fully utilizing energy-efficient potentials, deep SNNs have adopted temporal coding, such as time-to-firstspike (TTFS) [16,17,15], which represents the information with spike time and has shown superior efficiency with fewer spikes. However, their methods in deep SNNs have been limited due to the lack of consideration for efficiency during training.…”