ScieNet: Deep learning with spike-assisted contextual information extraction. (October 2021)

• Record Type: Journal Article
• Authors: She, Xueyuan; Long, Yun; Kim, Daehyun; Mukhopadhyay, Saibal
• Is Part Of: Pattern recognition. Volume 118(2021)
• Publication Date: 2021-10
• Subjects: Deep learning -- Noise robustness -- Spiking neural network; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2021.108002
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A new deep learning architecture that integrates spiking neural network (SNN) for contextual information extraction is proposed. The proposed design shows improved robustness for both random and structured input perturbation during inference. SNN is implemented with a novel frequency-dependent stochastic spike-timing-dependent-plasticity learning rule. Abstract: Spiking neural network (SNN) is a type of artificial neural network that uses biologically inspired neuron models and learning rules to develop artificial intelligence with capability parallel to human brain. Deep neural networks (DNNs), on the other hand, uses less biologically plausible neurons and training methods such as gradient descent, and has shown good accuracy in computer vision tasks. However, human brain can still outperform DNN in certain scenarios. For example, DNN experiences significant performance degradation when perturbation from various sources is present in the input, which makes DNN less reliable for systems interacting with physical world. In this paper, we present a hybrid deep net work architecture with s pike-assisted c ontextual i nformation e xtraction (ScieNet ) as a solution to the problem. ScieNet integrates a front-end SNN with a novel stochastic spike-timing-dependent plasticity (STDP) algorithm that extracts visual context from images. The back-end DNN is trained for classification given the contextual information. The integrated network demonstrates high resilience to input perturbations without relying on pre-training on perturbed inputs . We demonstrate ScieNet with various back-end DNNs for image classification using different datasets and considering both stochastic and structured input perturbations. Experimental results demonstrate significant improvement in accuracy on perturbed images, while maintaining state-of-the-art accuracy on clean images.