Models of vision need some action.

Bowers et al. focus their criticisms on research that compares behavioral and brain data from the ventral stream with a class of deep neural networks for object recognition. While they are right to identify issues with current benchmarking research programs, they overlook a much more fundamental limitation of this literature: Disregarding the importance of action and interaction for perception.

Behavioral signatures of face perception emerge in deep neural networks optimized for face recognition.

Human face recognition is highly accurate and exhibits a number of distinctive and well-documented behavioral "signatures" such as the use of a characteristic representational space, the disproportionate performance cost when stimuli are presented upside down, and the drop in accuracy for faces from races the participant is less familiar with. These and other phenomena have long been taken as evidence that face recognition is "special". But why does human face perception exhibit these properties in the first place? Here, we use deep convolutional neural networks (CNNs) to test the hypothesis that all of these signatures of human face perception result from optimization for the task of face recognition. Indeed, as predicted by this hypothesis, these phenomena are all found in CNNs trained on face recognition, but not in CNNs trained on object recognition, even when additionally trained to detect faces while matching the amount of face experience. To test whether these signatures are in principle specific to faces, we optimized a CNN on car discrimination and tested it on upright and inverted car images. As we found for face perception, the car-trained network showed a drop in performance for inverted vs. upright cars. Similarly, CNNs trained on inverted faces produced an inverted face inversion effect. These findings show that the behavioral signatures of human face perception reflect and are well explained as the result of optimization for the task of face recognition, and that the nature of the computations underlying this task may not be so special after all.

CNNs reveal the computational implausibility of the expertise hypothesis.

Face perception has long served as a classic example of domain specificity of mind and brain. But an alternative "expertise" hypothesis holds that putatively face-specific mechanisms are actually domain-general, and can be recruited for the perception of other objects of expertise (e.g., cars for car experts). Here, we demonstrate the computational implausibility of this hypothesis: Neural network models optimized for generic object categorization provide a better foundation for expert fine-grained discrimination than do models optimized for face recognition.

Using artificial neural networks to ask 'why' questions of minds and brains.

Neuroscientists have long characterized the properties and functions of the nervous system, and are increasingly succeeding in answering how brains perform the tasks they do. But the question 'why' brains work the way they do is asked less often. The new ability to optimize artificial neural networks (ANNs) for performance on human-like tasks now enables us to approach these 'why' questions by asking when the properties of networks optimized for a given task mirror the behavioral and neural characteristics of humans performing the same task. Here we highlight the recent success of this strategy in explaining why the visual and auditory systems work the way they do, at both behavioral and neural levels.

Brain-like functional specialization emerges spontaneously in deep neural networks.

The human brain contains multiple regions with distinct, often highly specialized functions, from recognizing faces to understanding language to thinking about what others are thinking. However, it remains unclear why the cortex exhibits this high degree of functional specialization in the first place. Here, we consider the case of face perception using artificial neural networks to test the hypothesis that functional segregation of face recognition in the brain reflects a computational optimization for the broader problem of visual recognition of faces and other visual categories. We find that networks trained on object recognition perform poorly on face recognition and vice versa and that networks optimized for both tasks spontaneously segregate themselves into separate systems for faces and objects. We then show functional segregation to varying degrees for other visual categories, revealing a widespread tendency for optimization (without built-in task-specific inductive biases) to lead to functional specialization in machines and, we conjecture, also brains.

How face perception unfolds over time.

Within a fraction of a second of viewing a face, we have already determined its gender, age and identity. A full understanding of this remarkable feat will require a characterization of the computational steps it entails, along with the representations extracted at each. Here, we used magnetoencephalography (MEG) to measure the time course of neural responses to faces, thereby addressing two fundamental questions about how face processing unfolds over time. First, using representational similarity analysis, we found that facial gender and age information emerged before identity information, suggesting a coarse-to-fine processing of face dimensions. Second, identity and gender representations of familiar faces were enhanced very early on, suggesting that the behavioral benefit for familiar faces results from tuning of early feed-forward processing mechanisms. These findings start to reveal the time course of face processing in humans, and provide powerful new constraints on computational theories of face perception.

Use and Usefulness of Dynamic Face Stimuli for Face Perception Studies-a Review of Behavioral Findings and Methodology.

Faces that move contain rich information about facial form, such as facial features and their configuration, alongside the motion of those features. During social interactions, humans constantly decode and integrate these cues. To fully understand human face perception, it is important to investigate what information dynamic faces convey and how the human visual system extracts and processes information from this visual input. However, partly due to the difficulty of designing well-controlled dynamic face stimuli, many face perception studies still rely on static faces as stimuli. Here, we focus on evidence demonstrating the usefulness of dynamic faces as stimuli, and evaluate different types of dynamic face stimuli to study face perception. Studies based on dynamic face stimuli revealed a high sensitivity of the human visual system to natural facial motion and consistently reported dynamic advantages when static face information is insufficient for the task. These findings support the hypothesis that the human perceptual system integrates sensory cues for robust perception. In the present paper, we review the different types of dynamic face stimuli used in these studies, and assess their usefulness for several research questions. Natural videos of faces are ecological stimuli but provide limited control of facial form and motion. Point-light faces allow for good control of facial motion but are highly unnatural. Image-based morphing is a way to achieve control over facial motion while preserving the natural facial form. Synthetic facial animations allow separation of facial form and motion to study aspects such as identity-from-motion. While synthetic faces are less natural than videos of faces, recent advances in photo-realistic rendering may close this gap and provide naturalistic stimuli with full control over facial motion. We believe that many open questions, such as what dynamic advantages exist beyond emotion and identity recognition and which dynamic aspects drive these advantages, can be addressed adequately with different types of stimuli and will improve our understanding of face perception in more ecological settings.

Task-dependent enhancement of facial expression and identity representations in human cortex.

What cortical mechanisms allow humans to easily discern the expression or identity of a face? Subjects detected changes in expression or identity of a stream of dynamic faces while we measured BOLD responses from topographically and functionally defined areas throughout the visual hierarchy. Responses in dorsal areas increased during the expression task, whereas responses in ventral areas increased during the identity task, consistent with previous studies. Similar to ventral areas, early visual areas showed increased activity during the identity task. If visual responses are weighted by perceptual mechanisms according to their magnitude, these increased responses would lead to improved attentional selection of the task-appropriate facial aspect. Alternatively, increased responses could be a signature of a sensitivity enhancement mechanism that improves representations of the attended facial aspect. Consistent with the latter sensitivity enhancement mechanism, attending to expression led to enhanced decoding of exemplars of expression both in early visual and dorsal areas relative to attending identity. Similarly, decoding identity exemplars when attending to identity was improved in dorsal and ventral areas. We conclude that attending to expression or identity of dynamic faces is associated with increased selectivity in representations consistent with sensitivity enhancement.

Near-optimal integration of facial form and motion.

Human perception consists of the continuous integration of sensory cues pertaining to the same object. While it has been fairly well shown that humans use an optimal strategy when integrating low-level cues proportional to their relative reliability, the integration processes underlying high-level perception are much less understood. Here we investigate cue integration in a complex high-level perceptual system, the human face processing system. We tested cue integration of facial form and motion in an identity categorization task and found that an optimal model could successfully predict subjects' identity choices. Our results suggest that optimal cue integration may be implemented across different levels of the visual processing hierarchy.

Identity information content depends on the type of facial movement.

Facial movements convey information about many social cues, including identity. However, how much information about a person's identity is conveyed by different kinds of facial movements is unknown. We addressed this question using a recent motion capture and animation system, with which we animated one avatar head with facial movements of three types: (1) emotional, (2) emotional in social interaction and (3) conversational, all recorded from several actors. In a delayed match-to-sample task, observers were best at matching actor identity across conversational movements, worse with emotional movements in social interactions, and at chance level with emotional facial expressions. Model observers performing this task showed similar performance profiles, indicating that performance variation was due to differences in information content, rather than processing. Our results suggest that conversational facial movements transmit more dynamic identity information than emotional facial expressions, thus suggesting different functional roles and processing mechanisms for different types of facial motion.

Quantifying human sensitivity to spatio-temporal information in dynamic faces.

A great deal of perceptual and social information is conveyed by facial motion. Here, we investigated observers' sensitivity to the complex spatio-temporal information in facial expressions and what cues they use to judge the similarity of these movements. We motion-captured four facial expressions and decomposed them into time courses of semantically meaningful local facial actions (e.g., eyebrow raise). We then generated approximations of the time courses which differed in the amount of information about the natural facial motion they contained, and used these and the original time courses to animate an avatar head. Observers chose which of two animations based on approximations was more similar to the animation based on the original time course. We found that observers preferred animations containing more information about the natural facial motion dynamics. To explain observers' similarity judgments, we developed and used several measures of objective stimulus similarity. The time course of facial actions (e.g., onset and peak of eyebrow raise) explained observers' behavioral choices better than image-based measures (e.g., optic flow). Our results thus revealed observers' sensitivity to changes of natural facial dynamics. Importantly, our method allows a quantitative explanation of the perceived similarity of dynamic facial expressions, which suggests that sparse but meaningful spatio-temporal cues are used to process facial motion.