Bridging the gap between Deep Learning and Neuroscience

No Thumbnail Available
Journal Title
Journal ISSN
Volume Title
In recent years, Deep Learning has achieved superhuman abilities in many tasks such as visual object recognition. Nevertheless, the brain outperforms Deep Networks in its ability to generalize to distorted images. Yet the exact mechanisms used to achieve this invariance are still not understood completely. The interplay between neuroscience and Deep Learning could both advance the knowledge on the processes that occur in the brain and help the development of more efficient artificial networks. The aim of the present paper is to study the link between the brain and artificial neural models by comparing the behavior of a Convolutional Neural Network to our knowledge of the processing of visual information in the human cortex. The network’s recognition ability under invariance conditions was tested when presenting input images that were different from the images employed for the training of the network. The test images were modified either with geometric deformations, by varying the rotation, position and size of the objects within the image, or by compromising the extent of visual information transmitted from the input when changing the quality, contrast and amount of noise. The results are compared to neural data obtained from behavioral and neuroimaging studies in which the subject’s response time, accuracy and neural activations were recorded following the presentation of images with the various types of deformations. Furthermore, the fundamental characteristics of the architecture of the network and the backpropagation algorithm used for the training process are discussed in comparison to the structure of the visual stream and to the synaptic update processes that are thought to be employed by the brain for learning. Our investigation highlights that a great issue with current Deep Neural Networks is the limited performance under image distortions as compared to humans’ invariant recognition ability. Furthermore, the present study underlines the differences in the implementation of the learning algorithm in computational models and in the brain as a starting point to improve Deep Learning towards more efficient and more biologically plausible networks.
Faculteit der Sociale Wetenschappen