English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

The tremendous energy consumption of deep neural networks (DNNs) has become aserious problem in deep learning. Spiking neural networks (SNNs), which mimicthe operations in the human brain, have been studied as prominentenergy-efficient neural networks. Due to their event-driven andspatiotemporally sparse operations, SNNs show possibilities forenergy-efficient processing. To unlock their potential, deep SNNs have adoptedtemporal coding such as time-to-first-spike (TTFS)coding, which represents theinformation between neurons by the first spike time. With TTFS coding, eachneuron generates one spike at most, which leads to a significant improvement inenergy efficiency. Several studies have successfully introduced TTFS coding indeep SNNs, but they showed restricted efficiency improvement owing to the lackof consideration for efficiency during training. To address the aforementionedissue, this paper presents training methods for energy-efficient deep SNNs withTTFS coding. We introduce a surrogate DNN model to train the deep SNN in afeasible time and analyze the effect of the temporal kernel on trainingperformance and efficiency. Based on the investigation, we proposestochastically relaxed activation and initial value-based regularization forthe temporal kernel parameters. In addition, to reduce the number of spikeseven further, we present temporal kernel-aware batch normalization. With theproposed methods, we could achieve comparable training results withsignificantly reduced spikes, which could lead to energy-efficient deep SNNs.

Training Energy-Efficient Deep Spiking Neural Networks with Time-to-First-Spike Coding