CNS*2020 Online has ended
Welcome to the Sched instance for CNS*2020 Online! Please read the instruction document on detailed information on CNS*2020.
Back To Schedule
Sunday, July 19 • 9:00pm - 10:00pm
P152: Face-selective neurons arise spontaneously in untrained deep neural networks

Log in to save this to your schedule, view media, leave feedback and see who's attending!

Feedback form is now closed.
Please attend to following Google Meeting URL:

Min Song
, Seungdae Baek, Jaeson Jang, Gwangsu Kim, Se-Bum Paik

In the primate brain, the neurons that selectively respond to faces are observed and considered as the basis of face recognition [1]. Although such face-selective neurons are observed in infant animals [2], the origin of face- selectivity is still under debate, because conflicting findings have raised questions whether this neuronal selectivity can arise spontaneously [3], or requires training from visual experience [2]. Here, we show that face- selective neurons can spontaneously arise in untrained deep neural networks (DNN), together with the previous notion that DNN could be considered as a visual cortex model that can perform human-level visual function and predict neuronal responses. Using biologically-inspired neural networks, AlexNet, we measured responses of the last convolutional layer to the image sets of face and 15 non-face classes. We found that face-selective neurons arise in untrained AlexNet with randomly permuted weights, where the face-selective neuron was defined as a neuron that showed a significantly higher response to face images compared to non-face images. To qualitatively examine the feature- selective response of these face-selective neurons, we reconstructed the preferred feature images of individual neurons using the reverse correlation method. We found face-components, such as eyes, nose, and mouth, in preferred feature images of face-selective neurons whereas no noticeable shape was found in neurons with no selectivity. Next, to test whether the selective response of these neurons could provide sufficient information to classify a face from other objects, we trained a support vector machine (SVM) to classify whether the given image was a face using neural responses of the untrained network. As a result, the SVM trained with only face-selective neurons shows significantly better performance than that trained with neurons with no selectivity. Next, to examine whether the face-selective neurons show view-point invariant characteristics observed in monkeys, we measured the responses of the permuted AlexNet while face images from five different angles were provided to the network. Surprisingly, the face-selective neurons in the network show viewpoint invariant responses and their level of invariance increased along the network hierarchy in the permuted AlexNet, similar to that in monkey IT. Lastly, to examine the origin of face-selectivity in untrained neural networks, we implemented a randomly initialized network where values in each weight kernel were randomly drawn from a weight distribution of the pre- trained AlexNet. We found that the number of face-selective neurons abruptly decreases when the weight variation is reduced to 52% of that in the pre- trained network. These results suggest that statistical variation present in the random feedforward projections could solely drive the emergence of innate face-selective neurons in the visual system. Overall, our findings provide insight into the origin of cognitive functions in both artificial and biological neural networks.

[1] Tsao, D. Y., Freiwald, W. A., Tootell, R. B. H. & Livingstone, M. S. A cortical region consisting entirely of face-selective cells. Science (2006) 311, 670–674

[2] Livingstone, M. S. et al. Development of the macaque face-patch system. Nat. Commun. (2017) 8

[3] Deen, B. et al. Organization of high-level visual cortex in human infants. Nat. Commun. (2017) 8

This work was supported by Grant Number: (2019M3E5D2A01058328, 2019R1A2C4069863)

avatar for Min Song

Min Song

PhD Student, Bio and Brain engineering, KAIST

Sunday July 19, 2020 9:00pm - 10:00pm CEST
Slot 07