Representation of Categories in Filters of Deep Neural Networks

Katerina Malakhova; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018, pp. 1973-1975

Abstract


Transparency in decision-making is an essential aspect of the secure and unbiased application of deep learning for classification problems. Neural networks pre-trained on one dataset can serve as feature extractors to solve various tasks. In this work, I study how categories are represented in latent space of neural networks using an example of face recognition by a network trained without an explicit category for the human person. I propose a semantic-based approach to determine if a model has pre-trained filters for a given set of classes of interest and which layer is better suited for feature extraction. The method is similar to category-selectivity measures used in neuroscience to estimate tuning curves of neurons in high-level areas of the visual cortex.

Related Material


[pdf]
[bibtex]
@InProceedings{Malakhova_2018_CVPR_Workshops,
author = {Malakhova, Katerina},
title = {Representation of Categories in Filters of Deep Neural Networks},
booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2018}
}