Abstract
Various techniques focus on understanding the way an image or a category concept is represented within convolutional neural networks. It is common to assume that one latent neuron can be a detector for a category or its dominant features. The analysis of the collective activity of neurons in hidden layers shows that representation of most of the categories is complex and distributed across as many as 93% of units participating in the encoding. To account for the complexity, this study suggests an approach that represents category by constructing a prototype formed as a covariance matrix of layer neurons’ activation. The approach makes it possible to take into account the coordinated response of the population of artificial neurons, as well as to preserve the complexity of the distributed activation pattern at different stages of processing, from low-level image statistics to the level of abstract and semantic features.
© 2022 Optica Publishing Group
PDF Article
More Like This
Cited By
You do not have subscription access to this journal. Cited by links are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
Contact your librarian or system administrator
or
Login to access Optica Member Subscription