Object classi cation under the predictive processing approach

Keywords
Loading...
Thumbnail Image
Issue Date
2021-06-18
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Current state-of-the-art algorithms for visual object classi cation are based on training large convolutional arti cial neural networks on labeled data sets. This means that rst, the features are directly learned for object classi cation and the computation is performed using the forward pass of the network. Conversely, evidence in computational neuroscience reveals recurrency, top-down modulation and unsupervised learning in the visual cortex. Here, I study the potential of the predictive processing approach in classi cation tasks. First, I compared the classi cation performance of unsupervised learnt features (using a variational autoencoder) against standard supervised approaches (VGG). Second, I studied the advantages of modeling the classi cation task as an inference process, with bottom-up and top-down modulation. While using predictive processing for classi cation on the CIFAR-10 data set, was shown to be inferior to that of the VGG model that is trained directly on classi cation error in a supervised manner with 56.71% against 84.23% classi cation accuracy, the classi cation accuracy of 98.63% on the MNIST data set showed there is de nitely potential. By modeling the classi cation task as an inference process, with bottom-up and top-down modulation, I showed that this method can help with the reduction of uncertainty that is present in the real world.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen