{"title":"CAT SNN: Conversion Aware Training for High Accuracy and Hardware Friendly Spiking Neural Networks","authors":"Dongwoo Lew;Jongsun Park","doi":"10.1109/TETC.2024.3435135","DOIUrl":null,"url":null,"abstract":"Among the various training algorithms for spiking neural network (SNN), ANN-to-SNN conversion gained popularity due to high accuracy and scalability to deep networks. By converting artificial neural network (ANN) to SNN and employing conversion loss reduction techniques, previous ANN-to-SNN conversion approaches achieved good accuracies. However, previous works do not consider the overheads to implement conversion loss reductions in hardware, thereby limiting its feasibility of hardware implementation. In this paper, we present conversion aware training (CAT), where SNN is simulated as closely as possible during ANN training for obtaining SNN-like ANN. So, our approach does not need any conversion loss reduction techniques after conversion, thus reducing hardware overhead while achieving state-of-the-art accuracies for SNNs using various neural coding methods. In addition, as an application of CAT for obtaining a hardware friendly SNN, we demonstrate a lightweight time-to-first-spike (TTFS) coding that adopts logarithmic computations enabled by CAT. An SNN processor that supports the logarithmic TTFS is implemented in 28nm CMOS process, achieving 91.7/67.9/57.4% accuracy and 486.7/503.6/1426uJ inference energy on CIFAR-10/100/Tiny-ImageNet, when running 5-bit logarithmic weight VGG-16. The key contributions are 1) proposing CAT as an ANN-to-SNN conversion guideline 2) applying CAT on various neural codings 3) presenting co-designed TTFS coding and processor.","PeriodicalId":13156,"journal":{"name":"IEEE Transactions on Emerging Topics in Computing","volume":"13 2","pages":"512-524"},"PeriodicalIF":5.4000,"publicationDate":"2024-08-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"IEEE Transactions on Emerging Topics in Computing","FirstCategoryId":"94","ListUrlMain":"https://ieeexplore.ieee.org/document/10622017/","RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, INFORMATION SYSTEMS","Score":null,"Total":0}
引用次数: 0
Abstract
Among the various training algorithms for spiking neural network (SNN), ANN-to-SNN conversion gained popularity due to high accuracy and scalability to deep networks. By converting artificial neural network (ANN) to SNN and employing conversion loss reduction techniques, previous ANN-to-SNN conversion approaches achieved good accuracies. However, previous works do not consider the overheads to implement conversion loss reductions in hardware, thereby limiting its feasibility of hardware implementation. In this paper, we present conversion aware training (CAT), where SNN is simulated as closely as possible during ANN training for obtaining SNN-like ANN. So, our approach does not need any conversion loss reduction techniques after conversion, thus reducing hardware overhead while achieving state-of-the-art accuracies for SNNs using various neural coding methods. In addition, as an application of CAT for obtaining a hardware friendly SNN, we demonstrate a lightweight time-to-first-spike (TTFS) coding that adopts logarithmic computations enabled by CAT. An SNN processor that supports the logarithmic TTFS is implemented in 28nm CMOS process, achieving 91.7/67.9/57.4% accuracy and 486.7/503.6/1426uJ inference energy on CIFAR-10/100/Tiny-ImageNet, when running 5-bit logarithmic weight VGG-16. The key contributions are 1) proposing CAT as an ANN-to-SNN conversion guideline 2) applying CAT on various neural codings 3) presenting co-designed TTFS coding and processor.
期刊介绍:
IEEE Transactions on Emerging Topics in Computing publishes papers on emerging aspects of computer science, computing technology, and computing applications not currently covered by other IEEE Computer Society Transactions. Some examples of emerging topics in computing include: IT for Green, Synthetic and organic computing structures and systems, Advanced analytics, Social/occupational computing, Location-based/client computer systems, Morphic computer design, Electronic game systems, & Health-care IT.