A Computational Model of the Development of Separate Representations of Facial Identity and Expression in the Primate Visual System

Abstract

Experimental studies have provided evidence that the visual processing areas of the primate brain represent facial identity and facial expression within different subpopulations of neurons. For example, in non-human primates there is evidence that cells within the inferior temporal gyrus (TE) respond primarily to facial identity, while cells within the superior temporal sulcus (STS) respond to facial expression. More recently, it has been found that the orbitofrontal cortex (OFC) of non-human primates contains some cells that respond exclusively to changes in facial identity, while other cells respond exclusively to facial expression. How might the primate visual system develop physically separate representations of facial identity and expression given that the visual system is always exposed to simultaneous combinations of facial identity and expression during learning? In this paper, a biologically plausible neural network model, VisNet, of the ventral visual pathway is trained on a set of carefully-designed cartoon faces with different identities and expressions. The VisNet model architecture is composed of a hierarchical series of four Self-Organising Maps (SOMs), with associative learning in the feedforward synaptic connections between successive layers. During learning, the network develops separate clusters of cells that respond exclusively to either facial identity or facial expression. We interpret the performance of the network in terms of the learning properties of SOMs, which are able to exploit the statistical indendependence between facial identity and expression.

Deep Neural Networks: A New Framework for Modeling Biological Vision and Brain Information Processing

“…The best performing models are deep neural nets and they are also best at explaining the IT representational geometry (Khaligh-Razavi et al 2014;Cadieu et al 2014). Khaligh-Razavi et al (2014) tested a wide range of classical computer vision features, several neuroscientifically motivated vision models, including VisNet (Wallis & Rolls 1997;Tromans et al 2011) and HMAX (Riesenhuber & Poggio 1999), and the deep neural network of Krizhesvsky et al (2012; Fig. 3).…”

The visually guided development of facial representations in the primate ventral visual pathway: A computer modeling study.

“…However, these previous theoretical and experimental studies do not explain the precise learning mechanisms by which these neuronal representations of global attributes, such as identity and expression, may become mapped onto separate processing areas in the later stages of the visual system. Recently, Tromans et al (2011) developed the first neural network model demonstrating how physically separated representations of facial identity and expression may develop through a biologically plausible process Figure 2. Physiological evidence from the single unit recording study carried out by Freiwald et al (2009) showing neuronal selectivity for the spatial relationships between facial features with monotonic tuning curves.…”

“…The important question is how such selective cell response properties could develop given that the visual system is always exposed to both facial identity and expression simultaneously during early visually guided learning and self-organization in the visual system. In earlier work carried out by Tromans et al (2011), it was shown that when VisNet was trained on a large number of faces, the higher layers of the model developed neurons that either responded to the identity of a face regardless of its emotional expression, or responded to the facial expression irrespective of facial identity. The hypothesized learning mechanism was as follows.…”

“…Second, in the study carried out by Tromans et al (2011), VisNet was trained on every possible combination of facial identity and expression to ensure the strongest possible statistical decoupling between these two facial dimensions. This helped to force individual neurons in the higher competitive layers to learn to respond to either a particular identity or expression.…”

Interactions between facial emotion and identity in face processing: Evidence based on redundancy gains

“…These authors argued that the functional dissociation between facial expression and identity related directly to the fact that these two facial characteristics load on different dimensions of the stimulus. These arguments for the fractionated processing of structural information about face identity and emotion have recently been bolstered by computational work that has suggested that independent processing of these two types of information is a natural consequence of the statistical independence between the image features for structural identity and emotion (Tromans, Harris, & Stringer, 2011).…”