Face detection in untrained deep neural networks

Baek, Seungjun; Song, Min; Jang, Jaeson; Kim, GwangSu; Paik, Se-Bum

doi:10.1038/s41467-021-27606-9

Cited by 41 publications

(54 citation statements)

References 91 publications

(116 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This result is consistent with human fMRI results that the object-selective LOC shows similar sensitivity to face identities for upright and inverted faces (Yovel and Kanwisher, 2005 ). Notably, although the AlexNet did not show a behavioral FIE in our study, it is reported that responses of face-selective units in untrained AlexNet responded more highly to upright faces than inverted faces (Baek et al, 2021 ). The discrepancy may be caused by different analysis levels (behavioral level vs. single unit response level) or different layers (FC layers vs. convolution layers).…”

Section: Discussioncontrasting

confidence: 73%

The Face Inversion Effect in Deep Convolutional Neural Networks

Fang

Xie

Song

et al. 2022

Front. Comput. Neurosci.

View full text Add to dashboard Cite

The face inversion effect (FIE) is a behavioral marker of face-specific processing that the recognition of inverted faces is disproportionately disrupted than that of inverted non-face objects. One hypothesis is that while upright faces are represented by face-specific mechanism, inverted faces are processed as objects. However, evidence from neuroimaging studies is inconclusive, possibly because the face system, such as the fusiform face area, is interacted with the object system, and therefore the observation from the face system may indirectly reflect influences from the object system. Here we examined the FIE in an artificial face system, visual geometry group network-face (VGG-Face), a deep convolutional neural network (DCNN) specialized for identifying faces. In line with neuroimaging studies on humans, a stronger FIE was found in VGG-Face than that in DCNN pretrained for processing objects. Critically, further classification error analysis revealed that in VGG-Face, inverted faces were miscategorized as objects behaviorally, and the analysis on internal representations revealed that VGG-Face represented inverted faces in a similar fashion as objects. In short, our study supported the hypothesis that inverted faces are represented as objects in a pure face system.

show abstract

Section: Discussioncontrasting

confidence: 73%

The Face Inversion Effect in Deep Convolutional Neural Networks

Fang

Xie

Song

et al. 2022

Front. Comput. Neurosci.

View full text Add to dashboard Cite

show abstract

“…It is worth noting that our present results were not about face selectivity (i.e., contrasting response between faces vs. objects) but identity selectivity (i.e., contrasting response between face identities, which does not require face selectivity 20 ). However, we found that the response of face identity-selective units could well generalize within faces but barely generalize to non-face objects, consistent with a dedicated and specialized face perception system 1 , 38 . It is also worth noting that most monkey MUA channels showed strong face responsiveness (i.e., modulation by face onset; e.g., Fig.…”

Section: Discussionsupporting

confidence: 65%

“…5 ) and found that our findings could generalize to other DNNs (although as expected face-recognition performance was reduced in some DNNs that were not trained for face recognition), consistent with a previous report surveying a large class of DNN models for face representation 37 . Interestingly, a recent study has even shown that face-selective units can emerge from an untrained DNN 38 . A future study will need to investigate whether identity-selective units can emerge from an untrained DNN.…”

Section: Discussionmentioning

confidence: 99%

“… 54 , 55 for more general natural image reconstruction). The latter includes a flurry of literature on the convergent evolution of face spaces across DNN layers and human face-selective brain areas 19 , the neurally plausible efficient inverse graphics model for face processing 56 , and spontaneous generation of face recognition in untrained DNNs 38 . Our recent study showing feature-based encoding of face identities in the human MTL using DNN-extracted visual features 20 also employed an encoding model.…”

Section: Discussionmentioning

confidence: 99%

See 1 more Smart Citation

Face identity coding in the deep neural network and primate brain

et al. 2022

View full text Add to dashboard Cite

A central challenge in face perception research is to understand how neurons encode face identities. This challenge has not been met largely due to the lack of simultaneous access to the entire face processing neural network and the lack of a comprehensive multifaceted model capable of characterizing a large number of facial features. Here, we addressed this challenge by conducting in silico experiments using a pre-trained face recognition deep neural network (DNN) with a diverse array of stimuli. We identified a subset of DNN units selective to face identities, and these identity-selective units demonstrated generalized discriminability to novel faces. Visualization and manipulation of the network revealed the importance of identity-selective units in face recognition. Importantly, using our monkey and human single-neuron recordings, we directly compared the response of artificial units with real primate neurons to the same stimuli and found that artificial units shared a similar representation of facial features as primate neurons. We also observed a region-based feature coding mechanism in DNN units as in human neurons. Together, by directly linking between artificial and primate neural systems, our results shed light on how the primate brain performs face recognition tasks.

show abstract

“…Recent computational investigations have revealed that ANNs with randomly initialized weights nevertheless displayed rudimentary object recognition abilities, and their internal neuron-like units exhibited sensitivities to categories such as faces (Baek, Song, Jang, Kim, & Paik, 2021;Xu, Zhang, Zhen, & Liu, 2020). The authors show that with a large Gaussian distribution of random weights, model units will, by chance, exhibit sensitivity to simple features like curvilinearity and topheaviness.…”

Section: Innate and Experiential Constraints On Object Recognitionmentioning

confidence: 99%

Development of object recognition

Ayzenberg¹,

Behrmann²

2022

Preprint

View full text Add to dashboard Cite

Object recognition is a fundamental cognitive ability that helps humans organize the visual world into meaningful perceptual units. To understand the nature of object recognition in humans fully, it is important to understand its developmental origins in infancy and the processes by which it reaches maturity. At birth, infants demonstrate a wide range of visual competencies including the ability to discriminate and categorize simple shapes. By 6-months, infants readily form holistic and three-dimensional shape representations which allow them to recognize objects in a viewpoint-invariant manner and categorize objects after exposure to only a few exemplars. Remarkably, infants acquire these skills with experience with just a few objects and, on average, only three faces. Several infant-specific biological and experiential constraints support the rapid development of object recognition. For instance, infants are born with innate perceptual biases for features like top-heaviness and symmetry that provide a scaffold for learning objects. Moreover, rather than being a hindrance, low visual acuity in infancy supports the development of holistic shape representations, allowing even newborns to perceptually group elements into a complete shape. Finally, infants visual experience with objects is seemingly ‘self-curated’ to support object learning, such that they display a bias for viewpoints that maximally display objects’ shape structure. Indeed, incorporating infant-like constraints into artificial neural network models improves the model’s performance on object recognition tasks. The neural mechanisms that support these abilities in infancy, however, may be different than adults. Although early visual cortex is relatively mature in infants, category-selective regions in higher-level areas of the ventral visual pathway do not mature fully until at least adolescence. Instead, neuroimaging and computational modelling research suggests that object recognition in infancy is supported by a more distributed neural code that represents lower-level features. Altogether, these studies suggest that the development of object recognition is bootstrapped by biological and experiential factors specific to infants’ developmental niche. By understanding the origins and development of object recognition, we gain a deeper understanding of the mechanisms that support object recognition in adulthood.

show abstract

Face detection in untrained deep neural networks

Cited by 41 publications

References 91 publications

The Face Inversion Effect in Deep Convolutional Neural Networks

The Face Inversion Effect in Deep Convolutional Neural Networks

Face identity coding in the deep neural network and primate brain

Development of object recognition

Contact Info

Product

Resources

About