Spontaneous generation of face recognition in untrained deep neural networks

Abstract

11Face-selective neurons are observed in the primate visual pathway and are considered the basis of facial 12 recognition in the brain. However, it is debated whether this neuronal selectivity can arise spontaneously, 13 or requires training from visual experience. Here, we show that face-selective neurons arise 14 spontaneously in random feedforward networks in the absence of learning. Using biologically inspired 15 deep neural networks, we found that face-selective neurons arise under three different network conditions: 16 one trained using non-face natural images, one randomized after being trained, and one never trained. 17 We confirmed that spontaneously emerged face-selective neurons show the biological view-point-18 invariant characteristics observed in monkeys. Such neurons suddenly vanished when feedforward 19 weight variation declined to a certain level. Our results suggest that innate face-selectivity originates from 20 statistical variation of the feedforward projections in hierarchical neural networks. 22The ability to identify and recognize faces is a crucial function in visual-priority social animals such as 23 humans and other primates, and is thought to originate from neuronal tuning at a single or multi-neuronal 24 level. Neurons that selectively respond to faces (face-selective neurons) are observed to occur in the 25 inferior temporal cortex (IT) 1-6 , superior temporal sulcus (STS) 7-10 , and fusiform face area (FFA) [11][12][13][14][15] in 26 the primate brain ( Fig. 1A). Several contradictory observations on the origin of face-selective neurons in 27 infant animals have been reported, raising two different scenarios for the development of this intriguing 28 functional tuning. 29The first scenario is that visual experience develops face-selective neurons. A study using 30 functional magnetic resonance imaging (fMRI) to examine FFA in monkeys reported that the category of 31 selective neuronal activity observed, depends greatly on what a subject had experienced in its lifetime 16 . 32Another fMRI study of IT in monkeys reported that robust tuning of face-selective neurons is not observed 33 until one year after birth 5 and that face-selectivity relies on experience during the early infant years. 34Furthermore, it was reported that monkeys raised without face exposure did not develop normal face-35 selective domains 17 . These results suggest that face-selective neurons are developed from training with 36 visual experience.37However, the other view suggests that face-selectivity can innately arise without visual 38 experience 18-21 . It was observed that human infants behaviorally prefer to look face-like objects rather 39 than non-face ones [22][23][24] , implying that face-encoding units may already exist in infants. It was also 40 reported that adult humans with no visual experience have category-selective domains including face, in 41 the ventral visual cortex 19 . In addition, a recent fMRI study of infant animals reported that face-selective 42 neurons are observed with movie st...

Face identity selectivity in the deep neural network and human brain

“…The copyright holder for this preprint (which this version posted May 14, 2021. ; https://doi.org/10.1101/2021.05.12.443829 doi: bioRxiv preprint system [1,23]. Lastly, consistent with our DNN lesion results (Fig.…”

The Face Module Emerged in a Deep Convolutional Neural Network Selectively Deprived of Face Experience

“…Was a face module formed in the d-AlexNet to support these functions? To answer this question, we searched all the channels in Conv5 of the d-AlexNet, where face-selective channels have been previously identified in the AlexNet (Baek et al, 2019 ). To do this, we calculated the activation of each channel in Conv5 after ReLU in response to each category of the classification dataset, and then identified channels that showed significantly higher response to faces than non-face images with Mann-Whitney U test ( p s < 0.05, Bonferroni corrected).…”

“…This study presented a DCNN model of selective visual deprivation of faces. Specifically, we chose the AlexNet as a test platform because of the functional correspondence along the hierarchy between the AlexNet and primates' ventral visual pathway (e.g., Krizhevsky et al, 2012 ; Cadieu et al, 2014 ; Wen et al, 2017 ; Pospisil et al, 2018 ; Baek et al, 2019 ). We found that without genetic predisposition and face-specific visual experiences, DCNNs were still capable of face perception.…”

The face module emerged in a deep convolutional neural network selectively deprived of face experience

“…Was a face module formed in the d-AlexNet to support these functions? To answer this question, we searched all the channels in Conv5 of the d-AlexNet, where face-selective channels have been previously identified in the AlexNet (Baek, Song, Jang, Kim, & Paik, 2019). To do this, we calculated the activation of each channel in Conv5 after ReLU in response to each category of the classification dataset, and then identified channels that showed significantly higher response to faces than non-face images with Mann-Whitney U test (ps < .05, Bonferroni corrected).…”