Neurocomputational Models of Face Processing

Abstract

Until the day we can record from multiple neurons in undergraduates, understanding how humans process faces requires an interdisciplinary approach, including building computational models that mimic how the brain processes faces.Using machine learning techniques, we can often build models that perform the same tasks people do, in neurophysiologically plausible ways. These models can then be manipulated and analyzed in ways that people cannot, providing insights that are unavailable from behavioral experiments. For example, as we will see below, our model of perceptual expertise can be "raised" in an environment where its "parents" are cups or cans instead of faces, and the same kind of processing ensues. This demonstrates, at least from our point of view, that there is nothing special about faces as an object class per se; rather, it is what we have to do with them -fine level discrimination of a homogeneous class -that is special.In this chapter, we will delineate two dimensions along which computational models of face (and object) processing may vary, and briefly review three such models (Dailey and Cottrell, 1999;O'Reilly and Munakata, 2000;Riesenhuber and Poggio, 1999). Subsequently, we will focus primarily on the model we are most familiar with (our own!) and how this model has been used to reveal potential mechanisms underlying the neural processing of faces and objects -the development of a specialized face processor, how it could be recruited for other domains, hemispheric lateralization of face processing, facial expression processing, and the development of face discrimination. At the end, we return to the Riesenhuber and Poggio model to describe the elegant way it has been used to predict fMRI data on face processing. The overall strategy of these modeling efforts is

Are Face and Object Recognition Independent? A Neurocomputational Modeling Exploration

“…These results are consistent with a neurocomputational model of face processing ("The Model" (TM); Dailey & Cottrell (1999) ; Cottrell & Hsiao (2011)). TM has been used to explain how and why an area of visual expertise for faces (the Fusiform Face Area) could be recruited for other nonface object categories: The resources in the face network can be shared with other object processing, provided that this processing is at the subordinate (expertise) level task (Joyce & Cottrell, 2004; Tong et al, 2008.…”

Invariant visual object recognition: biologically plausible approaches

“…However, we do emphasize that our aim in the present research is to investigate how the cerebral cortex operates in vision, not how computer vision attempts to solve similar problems. Within computer vision, we note that many approaches start with using independent component analysis (ICA) (Kanan 2013 ), principal component analysis (PCA) (Cottrell and Hsaio 2011 ), sparse coding (Kanan and Cottrell 2010 ), and other mathematical approaches (Larochelle and Hinton 2010 ) to derive what may be suitable ‘feature analysers,’ which are frequently compared to the responses of V1 neurons. Computer vision approaches to object identification then may take combinations of these feature analysers and perform statistical analyses using computer-based algorithms that are not biologically plausible such as Restricted Boltzmann Machines (RBMs) on these primitives to statistically discriminate different objects (Larochelle and Hinton 2010 ).…”

Toward a unified model of face and object recognition in the human visual system

“…Rather than just being interested in the presence of a nose, or the distance between the eyes, or hair color, or any other (more abstract) single feature of a face, I am arguing that under certain circumstances neurons will seek out multiple dimensions of the input, making the neuron highly selective for a few stimuli and highly sensitive to changes in any one of a number of features. This approach differs from arguments based on strictly holistic features such as those generated by principal component analysis (O'Toole et al, 1991; Cottrell and Hsiao, 2011), since neurons can be tuned to a single nameable feature and anything in between.…”

“…Their work then demonstrates how such a feature-based representation, “warped” or “molded” by the input, naturally generates an other-race effect (O'Toole et al, 1991; Furl et al, 2002). The precise mechanisms by which such a model might be implemented in cortex are left largely unexplored (Cottrell and Hsiao, 2011), but in many ways that is not the goal of the work. The authors generally use the PCA-features as a front-end to a classification network trained using supervised methods to prove the in-principle relation between the other-race effect and a feature-based representation.…”