Technical Program

Paper Detail

Paper: PS-1A.47
Session: Poster Session 1A
Location: H Lichthof
Session Time: Saturday, September 14, 16:30 - 19:30
Presentation Time:Saturday, September 14, 16:30 - 19:30
Presentation: Poster
Publication: 2019 Conference on Cognitive Computational Neuroscience, 13-16 September 2019, Berlin, Germany
Paper Title: A human-like view-invariant representation of faces in deep neural networks trained with faces but not with objects
Manuscript:  Click here to view manuscript
License: Creative Commons License
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Authors: Naphtali Abudarham, Galit Yovel, Tel Aviv University, Israel
Abstract: Face recognition depends on the generation of a view-invariant representation. Faces are known to engage specialized mechanisms and it is therefore of interest to reveal to what extent is specific experience with faces necessary for the development of this representation. This question is hard to study in humans, but can be studied with Deep Convolutional Neural Networks (DNNs) trained with faces or with objects. To examine whether a face-trained and an object-trained networks generate a human-like, view-invariant representation, we first examined the similarity between the representations of faces across different head views. We then examined whether the networks use the same view-invariant facial features that are used by humans for face recognition. Our findings show that a human-like view-invariant representation of faces emerges at higher layers of a face-trained DNN, but not the object-trained DNN. The representations of faces were similar at lower layers of the face-trained and object-trained networks. These findings may resemble the face and object pathways in the human brain that are similar in low-level areas and diverge at higher levels of the visual cortex. They further imply that invariant face recognition depends on experience with faces, during which the system learns to extract face-specific, invariant features.