A New Image Classification Architecture Inspired by Working Memory.

Neuroscience and Artificial Intelligence (AI) have a long and intertwined history. The two fields promote each other and develop together. At present, many famous AI algorithms are inspired by the biological brains and most of them have got better performance. Working memory is regarded as a cognitive system which has been demonstrated playing an important role in the algorithms involving long term data storage, such as natural language processing and path finding. However, if the algorithms of image classification, which is a kind of cognitive task, can be optimized by this human mechanism is unclear. In this paper, we aim to first investigate the cognitive task execution mechanism of working memory. Second we will explore a theoretical model from the physical cognitive mechanism and apply it on the image classification tasks. We will show that, our model overturns the way to feature extraction of previous models, it uses an unsupervised method instead of supervised methods and the accuracy is improved.