Minimizing Inference Time: Optimization Methods for Converted Deep Spiking Neural Networks

Mueller, Etienne; Hansjakob, Julius; Auge, Daniel; Knoll, Alois

doi:10.1109/ijcnn52387.2021.9533874

Cited by 5 publications

(2 citation statements)

References 36 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Meanwhile, several studies have explored the effectiveness of using quantization techniques to promote fast SNNs (Bu et al, 2021 ; Mueller et al, 2021 ; Wu et al, 2020 ). However, these methods either fail to scale to ImageNet, or suffer severe accuracy degradation.…”

Section: Related Workmentioning

confidence: 99%

Quantization Framework for Fast Spiking Neural Networks

Ма

Furber

2022

Front. Neurosci.

View full text Add to dashboard Cite

Compared with artificial neural networks (ANNs), spiking neural networks (SNNs) offer additional temporal dynamics with the compromise of lower information transmission rates through the use of spikes. When using an ANN-to-SNN conversion technique there is a direct link between the activation bit precision of the artificial neurons and the time required by the spiking neurons to represent the same bit precision. This implicit link suggests that techniques used to reduce the activation bit precision of ANNs, such as quantization, can help shorten the inference latency of SNNs. However, carrying ANN quantization knowledge over to SNNs is not straightforward, as there are many fundamental differences between them. Here we propose a quantization framework for fast SNNs (QFFS) to overcome these difficulties, providing a method to build SNNs with enhanced latency and reduced loss of accuracy relative to the baseline ANN model. In this framework, we promote the compatibility of ANN information quantization techniques with SNNs, and suppress “occasional noise” to minimize accuracy loss. The resulting SNNs overcome the accuracy degeneration observed previously in SNNs with a limited number of time steps and achieve an accuracy of 70.18% on ImageNet within 8 time steps. This is the first demonstration that SNNs built by ANN-to-SNN conversion can achieve a similar latency to SNNs built by direct training.

show abstract

Section: Related Workmentioning

confidence: 99%

Quantization Framework for Fast Spiking Neural Networks

Ма

Furber

2022

Front. Neurosci.

View full text Add to dashboard Cite

show abstract

“…Since then, this method has been well theorized and expanded for use with convolutional neural networks (CNNs) [11]. It enables the conversion of deep networks for object detection, including YOLO [12], ResNet [13], and RetinaNet [14].…”

Section: Introductionmentioning

confidence: 99%

Spiking Transformer Networks: A Rate Coded Approach for Processing Sequential Data

Mueller

Studenyak

Auge

et al. 2021

2021 7th International Conference on Systems and Informatics (ICSAI)

Self Cite

View full text Add to dashboard Cite

Machine learning applications are steadily increasing in performance, while also being deployed on a growing number of devices with limited energy resources. To minimize this trade-off, researchers are continually looking for more energy efficient solutions. A promising field involves the use of spiking neural networks in combination with neuromorphic hardware, significantly reducing energy consumption since energy is only consumed as information is being processed. However, as their learning algorithms lag behind conventional neural networks trained with backpropagation, not many applications can be found today. The highest levels of accuracy can be achieved by converting networks that are trained with backpropagation to spiking networks. Spiking neural networks can show nearly the same performance in fully connected and convolutional networks. The conversion of recurrent networks has been shown to be challenging. However, recent progress with transformer networks could change this. This type of network not only consists of modules that can easily be converted, but also shows the best accuracy levels for different machine learning tasks. In this work, we present a method to convert the transformer architecture to networks of spiking neurons. With only minimal conversion loss, our approach can be used for processing sequential data with very high accuracy while offering the possibility of reductions in energy consumption.

show abstract