Why psychologists should embrace rather than abandon DNNs.

Deep neural networks (DNNs) are powerful computational models, which generate complex, high-level representations that were missing in previous models of human cognition. By studying these high-level representations, psychologists can now gain new insights into the nature and origin of human high-level vision, which was not possible with traditional handcrafted models. Abandoning DNNs would be a huge oversight for psychological sciences.

Deep learning models challenge the prevailing assumption that face-like effects for objects of expertise support domain-general mechanisms.

The question of whether task performance is best achieved by domain-specific, or domain-general processing mechanisms is fundemental for both artificial and biological systems. This question has generated a fierce debate in the study of expert object recognition. Because humans are experts in face recognition, face-like neural and cognitive effects for objects of expertise were considered support for domain-general mechanisms. However, effects of domain, experience and level of categorization, are confounded in human studies, which may lead to erroneous inferences. To overcome these limitations, we trained deep learning algorithms on different domains (objects, faces, birds) and levels of categorization (basic, sub-ordinate, individual), matched for amount of experience. Like humans, the models generated a larger inversion effect for faces than for objects. Importantly, a face-like inversion effect was found for individual-based categorization of non-faces (birds) but only in a network specialized for that domain. Thus, contrary to prevalent assumptions, face-like effects for objects of expertise do not support domain-general mechanisms but may originate from domain-specific mechanisms. More generally, we show how deep learning algorithms can be used to dissociate factors that are inherently confounded in the natural environment of biological organisms to test hypotheses about their isolated contributions to cognition and behaviour.

Face Recognition Depends on Specialized Mechanisms Tuned to View-Invariant Facial Features: Insights from Deep Neural Networks Optimized for Face or Object Recognition.

Face recognition is a computationally challenging classification task. Deep convolutional neural networks (DCNNs) are brain-inspired algorithms that have recently reached human-level performance in face and object recognition. However, it is not clear to what extent DCNNs generate a human-like representation of face identity. We have recently revealed a subset of facial features that are used by humans for face recognition. This enables us now to ask whether DCNNs rely on the same facial information and whether this human-like representation depends on a system that is optimized for face identification. In the current study, we examined the representation of DCNNs of faces that differ in features that are critical or non-critical for human face recognition. Our findings show that DCNNs optimized for face identification are tuned to the same facial features used by humans for face recognition. Sensitivity to these features was highly correlated with performance of the DCNN on a benchmark face recognition task. Moreover, sensitivity to these features and a view-invariant face representation emerged at higher layers of a DCNN optimized for face recognition but not for object recognition. This finding parallels the division to a face and an object system in high-level visual cortex. Taken together, these findings validate human perceptual models of face recognition, enable us to use DCNNs to test predictions about human face and object recognition as well as contribute to the interpretability of DCNNs.

Developmental prosopagnosics and super recognizers rely on the same facial features used by individuals with normal face recognition abilities for face identification.

Face recognition depends on the ability of the face processing system to extract facial features that define the identity of a face. In a recent study we discovered that altering a subset of facial features changed the identity of the face, indicating that they are critical for face identification. Changing another set of features did not change the identity of a face, indicating that they are not critical for face identification. In the current study, we assessed whether developmental prosopagnosics (DPs) and super recognizers (SRs) also rely more heavily on these critical features than non-critical features for face identification. To that end, we presented to DPs and SRs faces in which either the critical or the non-critical features were manipulated. In Study 1, we presented SRs with a famous face recognition task. We found that overall SRs recognized famous faces that differ in either critical or non-critical features better than controls. Similar to controls, changes in critical features had a larger effect on SRs' face recognition than changes in non-critical features. In Study 2, we presented an identity matching task to DPs and SRs. Similar to controls, DPs and SRs perceived faces that differed in critical features as more different than faces that differed in non-critical features. Taken together, our results indicate that SRs and DPs use the same critical features for face identification as normal individuals. These findings emphasize the fundamental role of this subset of features for face identification.

You won't believe what this guy is doing with the potato: The ObjAct stimulus-set depicting human actions on congruent and incongruent objects.

Perception famously involves both bottom-up and top-down processes. The latter are influenced by our previous knowledge and expectations about the world. In recent years, many studies have focused on the role of expectations in perception in general, and in object processing in particular. Yet studying this question is not an easy feat, requiring-among other things-the creation and validation of appropriate stimuli. Here, we introduce the ObjAct stimulus-set of free-to-use, highly controlled real-life scenes, on which critical objects are pasted. All scenes depict human agents performing an action with an object that is either congruent or incongruent with the action. The focus on human actions yields highly constraining contexts, strengthening congruency effects. The stimuli were analyzed for low-level properties, using the SHINE toolbox to control for luminance and contrast, and using a deep convolutional neural network to mimic V1 processing and potentially discover other low-level factors that might differ between congruent and incongruent scenes. Two online validation studies (N = 500) were also conducted to assess the congruency manipulation and collect additional ratings of our images (e.g., arousal, likeability, visual complexity). We also provide full descriptions of the online sources from which all images were taken, as well as verbal descriptions of their content. Taken together, this extensive validation and characterization procedure makes the ObjAct stimulus-set highly informative and easy to use for future researchers in multiple fields, from object and scene processing, through top-down contextual effects, to the study of actions.

From concepts to percepts in human and machine face recognition: A reply to Blauch, Behrmann & Plaut.

Intact recognition of familiar faces is critical for appropriate social interactions. Thus, the human face processing system should be optimized for familiar face recognition. Blauch et al. (2020) used face recognition deep convolutional neural networks (DCNNs) that are trained to maximize recognition of the trained (familiar) identities, to model human unfamiliar and familiar face recognition. In line with this model, we discuss behavioral, neuroimaging and computational findings that indicate that human face recognition develops from the generation of identity-specific concepts of familiar faces that are learned in a supervised manner, to the generation of view-invariant identity-general perceptual representations. Face-trained DCNNs seem to share some fundamental similarities with this framework.

Same critical features are used for identification of familiarized and unfamiliar faces.

Many studies have shown better recognition for faces we have greater experience with, relative to unfamiliar faces. However, it is still not clear if and how the representation of faces changes during the process of familiarization. In a previous study, we discovered a subset of facial features, for which we have high perceptual sensitivity (PS), that were critical for determining the identity of unfamiliar faces. This was done by assigning values to 20 different facial features based on perceptual rating, converting faces into feature-vectors, and measuring the correlations between face similarity ratings and distances between feature-vectors. In the current study, we examined the contribution of high and low-PS features to face identity after familiarization. To familiarize participants with unfamiliar faces, we used an individuation training protocol that was found to be effective in previous studies, in which different names are assigned to different faces and participants are asked to learn the face-name association. Our findings show that even after repeated exposure to the same image of each identity, which allows close examination of all facial features, only high-PS features contributed to face identity, while low-PS features did not. This subset of high-PS features includes both internal and external features and part and configuration features. We therefore conclude that identification of familiarized and unfamiliar faces may rely on the same subset of critical features. These findings further support a new categorization of facial features according to their perceptual sensitivity.

Critical features for face recognition.

Face recognition is a computationally challenging task that humans perform effortlessly. Nonetheless, this remarkable ability is better for familiar faces than unfamiliar faces. To account for humans' superior ability to recognize familiar faces, current theories suggest that different features are used for the representation of familiar and unfamiliar faces. In the current study, we applied a reverse engineering approach to reveal which facial features are critical for familiar face recognition. In contrast to current views, we discovered that the same subset of features that are used for matching unfamiliar faces, are also used for matching as well as recognition of familiar faces. We further show that these features are also used by a deep neural network face recognition algorithm. We therefore propose a new framework that assumes similar perceptual representation for all faces and integrates cognition and perception to account for humans' superior recognition of familiar faces.

Reverse engineering the face space: Discovering the critical features for face identification.

How do we identify people? What are the critical facial features that define an identity and determine whether two faces belong to the same person or different people? To answer these questions, we applied the face space framework, according to which faces are represented as points in a multidimensional feature space, such that face space distances are correlated with perceptual similarities between faces. In particular, we developed a novel method that allowed us to reveal the critical dimensions (i.e., critical features) of the face space. To that end, we constructed a concrete face space, which included 20 facial features of natural face images, and asked human observers to evaluate feature values (e.g., how thick are the lips). Next, we systematically and quantitatively changed facial features, and measured the perceptual effects of these manipulations. We found that critical features were those for which participants have high perceptual sensitivity (PS) for detecting differences across identities (e.g., which of two faces has thicker lips). Furthermore, these high PS features vary minimally across different views of the same identity, suggesting high PS features support face recognition across different images of the same face. The methods described here set an infrastructure for discovering the critical features of other face categories not studied here (e.g., Asians, familiar) as well as other aspects of face processing, such as attractiveness or trait inferences.

Amelioration of operation-induced suppression of marginating pulmonary NK activity using poly IC: a potential approach to reduce postoperative metastasis.

BACKGROUND AND OBJECTIVES: Pulmonary metastasis is a major cause of death in cases of operable cancer, and evidence suggests that postoperative immunosuppression contributes to this complication. In this study, we aimed to circumvent this risk and identify immunocytes critical in preventing pulmonary metastases. METHODS: F344 rats were treated with either vehicle or repeated low doses of poly I-C (0.2 mg/kg i.p., days 5, 3, and 1 preoperatively), a Th1-cytokine-inducing agent, then subjected or not to laparotomy. Using a non-immunogenic syngeneic mammary adenocarcinoma line (MADB106) we studied: (a) NK cytotoxicity (NKC) in marginating-pulmonary (MP) and in circulating leukocytes; (b) resistance to experimental lung metastasis; and (c) in vitro susceptibility of NKC to corticosterone and prostaglandin-E(2), substances thought to mediate postoperative immunosuppression. RESULTS: MP but not circulating leukocytes showed significant NKC against MADB106 cells. Surgery suppressed this MP-NKC per NK cell and promoted MADB106 metastasis, and poly I-C treatment completely abolished both effects. Poly I-C quadrupled the numbers of MP-NK cells without causing apparent side effects, and protected MP-NKC from in vitro suppression by corticosterone and prostaglandin-E(2). CONCLUSIONS: MP-NK cells are unique in their ability to kill this apparently immunoresistant tumor. Low doses of synthetic ds-RNA (poly I-C), and potentially Th1 cytokines, can expand this MP-NK population and protect it from immunosuppression. The novelty of such a prophylactic approach is targeting the immediate postoperative period, which is characterized by high vulnerability to residual disease, and protecting critical anti-metastatic immunity against postoperative suppression. Testing such a potentially innocuous intervention in oncology patients preparing for surgery may reduce metastatic recurrence.

Marginating pulmonary-NK activity and resistance to experimental tumor metastasis: suppression by surgery and the prophylactic use of a beta-adrenergic antagonist and a prostaglandin synthesis inhibitor.

Surgery is imperative for cancer treatment, but was suggested to suppress immunity and facilitate metastasis. Here we study the involvement of catecholamines and prostaglandins (PG) in such outcomes, and the role played by marginating-pulmonary (MP)-NK cells in controlling MADB106 metastasis. Non-operated and laparotomized F344 rats were injected postoperatively with a PG synthesis inhibitor (indomethacin, 4 mg/kg i.p.), a beta-blocker (nadolol, 0.6 mg/kg s.c.), both drugs, or vehicle. Rats were then inoculated intravenously with non-immunogenic syngeneic MADB106 cells, and 24 h later lung tumor retention was assessed, or 3 weeks later lung metastases were counted. Additionally, 12 h after surgery we harvested MP-NK cells and circulating-NK cells and compared their numbers and cytotoxicity against MADB106 cells and standard YAC-1 target cells. Surgery significantly increased MADB106 metastasis. Nadolol and indomethacin reduced this effect by approximately 50% when used alone, and significantly more (75%) when used together. Only MP-leukocytes exhibited NK cytotoxicity against MADB106 cells. Surgery markedly suppressed it, and nadolol and indomethacin additively restored it. Similar effects were observed assessing MP-NK and circulating-NK cytotoxicity against YAC-1 target cells. Alterations in the numbers of NK cells were partly associated with alterations in total MP-NK activity, but not with circulating-NK activity. Last, administrating nai ve rats with physiologically relevant doses of a beta-adrenergic agonist (metaproterenol), and/or with PGE2, additively and independently of each other promoted MADB106 metastasis, simulating the effects of surgery. These findings point at potential prophylactic measures in cancer patients undergoing surgery, and suggest a role for MP-NK cells in resisting metastasis of apparently insensitive tumors.

Prostaglandin e(2) suppresses NK activity in vivo and promotes postoperative tumor metastasis in rats.

BACKGROUND: Prostaglandins (PGs) were shown in vitro to suppress several functions of cellular immunity. It is unclear, however, whether physiological levels of PGs can suppress cellular immunity in vivo and whether such suppression would compromise postoperative host resistance to metastasis. METHODS: Fischer 344 rats were administered PGE(2) in doses (18 to 300 micro g/kg subcutaneously) that increased the serum levels approximately 2- to 4-fold. We then assessed the number and activity of circulating natural killer (NK) cells, as well as rats' resistance to experimental metastasis of a syngeneic NK-sensitive tumor (MADB106). To study whether endogenously released PGs after surgery compromise these indices, we tested whether laparotomy adversely affects them and whether a cyclooxygenase-synthesis inhibitor, indomethacin (4 mg/kg), attenuates these effects. RESULTS: PGE(2) dose-dependently suppressed NK activity per NK cell and dose-dependently increased 4- and 24-hour MADB106 lung tumor retention (LTR); 240 micro g/kg of PGE(2) quadrupled the number of lung metastases counted 3 weeks later. Selective depletion of NK cells abrogated the promotion of LTR by PGE(2). Surgery significantly suppressed NK activity and increased MADB106 LTR, and indomethacin halved these effects without affecting nonoperated rats. CONCLUSIONS: PGE(2) is a potent in vivo suppressor of NK activity, and its postoperative release may promote tumor recurrence.