Emergence of brain-like mirror-symmetric viewpoint tuning in convolutional neural networks
Abstract
Primates can recognize objects despite 3D geometric variations such as in-depth rotations. The computational mechanisms that give rise to such invariances are yet to be fully understood. A curious case of partial invariance occurs in the macaque face-patch AL and in fully connected layers of deep convolutional networks in which neurons respond similarly to mirror-symmetric view (e.g., left and right profiles). Why does this tuning develop? Here, we propose a simple learning-driven explanation for mirror-symmetric viewpoint tuning. We show that mirror-symmetric viewpoint tuning for faces emerges in the fully connected layers of convolutional deep neural networks trained on object recognition tasks, even when the training dataset does not include faces. First, using 3D objects rendered from multiple views as test stimuli, we demonstrate that mirror-symmetric viewpoint tuning in convolutional neural network models is not unique to faces: it emerges for multiple object categories with bilateral symmetry. Second, we show why this invariance emerges in the models. Learning to discriminate among bilaterally symmetric object categories induces reflection-equivariant intermediate representations. AL-like mirror-symmetric tuning is achieved when such equivariant responses are spatially pooled by downstream units with sufficiently large receptive fields. These results explain how mirror-symmetric viewpoint tuning can emerge in neural networks, providing a theory of how they might emerge in the primate brain. Our theory predicts that mirror-symmetric viewpoint tuning can emerge as a consequence of exposure to bilaterally symmetric objects beyond the category of faces, and that it can generalize beyond previously experienced object categories.
Data availability
The stimulus set and the source code required for reproducing our results are available at https://gitfront.io/r/afarzmahdi/p666tmWy7YuY/AL-symmetry-manuscript-codes/.
-
Object categories across viewshttps://github.com/amirfarzmahdi/AL-Symmetry.
Article and author information
Author details
Funding
National Eye Institute (R01EY021594)
- Winrich A Freiwald
National Eye Institute (R01EY029998)
- Winrich A Freiwald
National Institute of Neurological Disorders and Stroke (RF1NS128897)
- Nikolaus Kriegeskorte
Naval Research Laboratory (N00014-20-1-2292)
- Winrich A Freiwald
Charles H. Revson Foundation
- Tal Golan
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Copyright
© 2024, Farzmahdi et al.
This article is distributed under the terms of the Creative Commons Attribution License permitting unrestricted use and redistribution provided that the original author and source are credited.
Metrics
-
- 894
- views
-
- 149
- downloads
-
- 0
- citations
Views, downloads and citations are aggregated across all versions of this paper published by eLife.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading
-
- Neuroscience
In albino mice and EphB1 knockout mice, mistargeted retinal ganglion cell axons form dense islands of axon terminals in the dorsal lateral geniculate nuclei (dLGN). The formation of these islands of retinal input depends on developmental patterns of spontaneous retinal activity. We reconstructed the microcircuitry of the activity-dependent islands and found that the boundaries of the island represent a remarkably strong segregation within retinogeniculate connectivity. We conclude that when sets of retinal input are established in the wrong part of the dLGN, the developing circuitry responds by forming a synaptically isolated subcircuit within the otherwise fully connected network. The fact that there is a developmental starting condition that can induce a synaptically segregated microcircuit has important implications for our understanding of the organization of visual circuits and our understanding of the implementation of activity-dependent development.
-
- Neuroscience
The neuropeptides Substance P and CGRPα have long been thought important for pain sensation. Both peptides and their receptors are expressed at high levels in pain-responsive neurons from the periphery to the brain making them attractive therapeutic targets. However, drugs targeting these pathways individually did not relieve pain in clinical trials. Since Substance P and CGRPα are extensively co-expressed, we hypothesized that their simultaneous inhibition would be required for effective analgesia. We therefore generated Tac1 and Calca double knockout (DKO) mice and assessed their behavior using a wide range of pain-relevant assays. As expected, Substance P and CGRPα peptides were undetectable throughout the nervous system of DKO mice. To our surprise, these animals displayed largely intact responses to mechanical, thermal, chemical, and visceral pain stimuli, as well as itch. Moreover, chronic inflammatory pain and neurogenic inflammation were unaffected by loss of the two peptides. Finally, neuropathic pain evoked by nerve injury or chemotherapy treatment was also preserved in peptide-deficient mice. Thus, our results demonstrate that even in combination, Substance P and CGRPα are not required for the transmission of acute and chronic pain.