The face inversion effect through the lens of deep neural networks

What is the functional organization of the human mind and brain?This question has fuelled a longstanding debate in philosophy and science.One hypothesis proposes that the brain is organized according to mechanisms specialized for processing distinct stimulus domains in our sensory environments, such as birds or human faces.This hypothesis has been referred to as the domain-specific processing account.In contrast, another hypothesis proposes that the brain is organized around processing mechanisms that can be applied across domains, such as discriminating between different individuals of the same domain.This hypothesis has been referred to as the domain-general processing account.These two hypotheses feature prominently in the area of face perception [1].Because of their biological and social importance and their distinct visual appearance, faces can be considered a prototypical domain.This makes face perception a good candidate for a function supported by domain-specific processing mechanisms.However, studies testing this hypothesis have yielded mixed evidence.One set of results suggests that face perception is indeed supported by domain-specific processing.A prominent example of such results is the face inversion effect, which is an impairment in face recognition after faces are turned upside-down [2].Inversion effects are weaker or not observed for other objects, which led researchers to conclude that faces are processed differently than other objects, with a larger emphasis on holistic processing [3].Other results, however, favour the domain-general processing hypothesis.Behavioural indices of holistic processing have been found for objects other than faces, especially if they require individuation and share a stereotypical configuration of parts [4].This line of evidence motivated the expertise hypothesis: behavioural effects that appear face-specific are driven by expertise in individuating homogeneous objects [4].It remains challenging to conclusively adjudicate between these hypotheses because domain and process tend to be correlated in the real world: we are all experts at face recognition.This leaves open if the face inversion effect is specific to the domain of faces or driven by the process of individuation.Yovel et al. [5] set out to address this question using artificial deep neural networks.Over the last decade, deep neural networks have emerged as promising computational models of the human visual system [6].Deep neural networks are loosely inspired by the brain's visual system: they consist of artificial neurons grouped in processing layers and connected by weights, which are iteratively adjusted as the network learns to perform real-life tasks on visual input.Deep neural networks can learn to categorize images, and predict brain activity measured in humans during image viewing [6].Importantly, domain and process can be independently manipulated in deep neural network training.This makes deep neural networks attractive as a framework for addressing questions about visual learning that are challenging, or practically impossible, to address in humans [7].To examine if the inversion effect is face-specific or domain-general, Yovel et al. used deep neural networks as computational models of perceptual expertise.The authors used a feedforward neural network architecture, VGG-16, pretrained to categorize images of inanimate objects.Pretraining served to build general experience with the visual world.