Dynamics of population coding for object views following object discrimination training

Highlights

• Responses in a population of monkey inferotemporal cells to prior experienced object views have been investigated.

• High-response similarity to larger separated object views was observed later after object image presentation.

• The results demonstrate the dynamics of population coding for object views following object discrimination.

Abstract

We have previously demonstrated that inferotemporal neurons respond to objects viewed from a range of angles, even without any prior experience in learning the associations among the views. Several models have been proposed to explain object recognition across disparate views. However, direct neuronal evidence is rare. In the present study, we focused on the response similarity of a population of inferotemporal cells to object views, following different prior experiences. Two monkeys were subjected to a task in which object discrimination across views was required. We found significantly higher neural response similarity to 30° separated views, 190 ms after object image presentation, than without any prior discrimination experience across views. The time period over which the similarity was significant began and endured similarly for 60° separated views at 190–850 ms. For 90° separated views, the time period over which the similarity was significant was shorter and started later, at 230–550 ms. The results demonstrate the dynamics of cell population activity and suggest a possible explanation for object recognition across disparate views.

Introduction

Our visual system can discriminate one object from others despite changes in viewing angle. Rotation in depth is one of the most difficult and challenging problems since features can appear drastically different from one view to another (Tarr and Bülthoff, 1995, Logothetis and Sheinberg, 1996, Ullman, 1996, Edelman, 1998, Biederman, 2000, Hayward, 2003, Kersten and Yuille, 2003, Kersten et al., 2004, Palmeri and Gauthier, 2004, Pinto et al., 2008, Poggio and Ullman, 2013). For a 3D object, it is always the case that features on the surface of an object that are closer to the observer are visible but those on farther surfaces are not. Nearby views largely share common visible object features, whereas widely separated views do not. Our previous study (Wang et al., 2005) showed that once monkeys experienced discrimination of objects from several individual views, they were able to recognize the objects over a certain range of viewpoints, but views with large angular differences required additional learning of the association among views. The association of nearby views developed intrinsically owing to their common features, whereas the underlying mechanism was different for the association of widely separated views.

The inferotemporal cortex (IT) is the final stage of the ventral cortical pathway representing information on object shape. It has been demonstrated to be critical for object recognition and discrimination (for a review, see Mishkin et al., 1983). The representation of objects in IT has been studied extensively. IT neurons usually have large receptive fields and show response tolerance to some object attributes (for a review, see Tanaka, 1996). However, reports on view invariance remain controversial. Familiarity with the targeted objects has a significant impact on the formation of stimulus selectivity. IT neurons respond tolerantly to familiar natural objects or faces over almost all viewing angles (Perrett et al., 1991, Booth and Rolls, 1998, Freiwald and Tsao, 2010, Eifuku et al., 2011), whereas for unfamiliar artificial objects, the viewing angles over which the neurons respond tolerantly are much narrower (Yamane et al., 2008, Okamura et al., 2014). For wireframe objects, the response was found to be view-dependent although with extensive prior training (Logothetis et al., 1994, Logothetis et al., 1995). In our previous experiment, even though we extensively trained the association between views of 3D artificial objects across 90° shifts in viewing angle, we failed to find any inferotemporal cell showing perfect view-invariant responses (Okamura et al., 2014). Association learning across views of the same objects broadened the cell tuning across views, but did not lead to full view invariance. All cells had tuning widths less than 60°, even following extensive long-term training. The present study asked how fully-invariant object recognition is completed, if indeed it is, in IT. We clarified the processing of view-invariant object discrimination in IT by using population activity, in addition to analysis at the single-cell level.