How learning might strengthen existing visual object representations in human object-selective cortex

Highlights

• Object categorization and discrimination change cortical object representations.

• This change includes an increase in multi-voxel pattern information.

• This increase is related to a strengthening of the pre-training representations.

Abstract

Visual object perception is an important function in primates which can be fine-tuned by experience, even in adults. Which factors determine the regions and the neurons that are modified by learning is still unclear. Recently, it was proposed that the exact cortical focus and distribution of learning effects might depend upon the pre-learning mapping of relevant functional properties and how this mapping determines the informativeness of neural units for the stimuli and the task to be learned. From this hypothesis we would expect that visual experience would strengthen the pre-learning distributed functional map of the relevant distinctive object properties. Here we present a first test of this prediction in twelve human subjects who were trained in object categorization and differentiation, preceded and followed by a functional magnetic resonance imaging session. Specifically, training increased the distributed multi-voxel pattern information for trained object distinctions in object-selective cortex, resulting in a generalization from pre-training multi-voxel activity patterns to after-training activity patterns. Simulations show that the increased selectivity combined with the inter-session generalization is consistent with a training-induced strengthening of a pre-existing selectivity map. No training-related neural changes were detected in other regions. In sum, training to categorize or individuate objects strengthened pre-existing representations in human object-selective cortex, providing a first indication that the neuroanatomical distribution of learning effects depends upon the pre-learning mapping of visual object properties.

Introduction

Visual object representations in high-level visual cortex are dynamically updated through learning (e.g., Gauthier et al., 1999b, Grill-Spector et al., 2000, Kourtzi et al., 2005, Li et al., 2009, Op de Beeck et al., 2006). Specifically, learning is associated with an increased selectivity for those object properties that are relevant during training (e.g., Folstein et al., 2013, Gillebert et al., 2009, Jiang et al., 2007, van der Linden et al., 2008, Zhang et al., 2010). However, it is still unclear how learning effects are distributed across visual cortex (Bukach et al., 2006, Harel et al., 2010, Op de Beeck and Baker, 2010).

The neural representation of objects and concepts is very complex, with large-scale distributed maps (Haxby et al., 2001, Huth et al., 2012, Konkle and Oliva, 2012, Op de Beeck et al., 2008a), as well as strong focal specificity for faces (e.g., in the FFA) (Grill-Spector, 2003, Kanwisher et al., 1997, Kanwisher and Yovel, 2006), body parts (Downing et al., 2001, Schwarzlose et al., 2005), houses (Epstein et al., 1999) and words (Cohen et al., 2000). Previous studies of learning, focusing either upon laboratory training or upon visual expertise obtained in a more natural setting, have regularly observed learning effects in face-selective regions, in particular in the FFA (Gauthier et al., 1999b, McGugin et al., 2012). In these studies, effects have sometimes also been noted in other functional regions (older study of Gauthier et al.) or in non-face-selective voxels nearby the FFA (McGugin et al., 2012, Wong et al., 2009). There are also studies that did not observe learning effects in face-selective regions, but more in other functional regions such as general object-selective cortex (e.g., the lateral occipital complex or LOC, see Op de Beeck et al., 2006) or even very distributed across visual cortex (Harel et al., 2010).

This heterogeneity illustrates an important point made before (Bukach et al., 2006), namely that the development of visual expertise through experience is a complex phenomenon, characterized by interactions between experience, task demands, and neural architecture. Here, we highlight the potential role of just one of these factors, namely the neural architecture. Recently, it was proposed how the anatomical distribution of learning-induced changes might depend upon the functional neuroanatomy in high-level visual cortex as it exists prior to the learning experience (Op de Beeck, 2012, Op de Beeck and Baker, 2010). Based upon the multidimensional functional properties of brain regions, it is expected that the degree to which each region participates in a particular visual task, and therefore in the learning of this task, varies. More specifically, learning might mostly be associated with changes in the tuning of neural units (neurons or clusters of neurons such as voxels) which are most informative for the learning situation at hand. This informativeness hypothesis predicts a link between the pre-learning functional properties of a neural unit and the effect of learning in this unit (Fig. 1). This prediction has been supported by single-unit recordings in monkeys in the specific context of a simple orientation discrimination task (Raiguel et al., 2006), but the current literature does not deliver direct evidence in the context of object recognition (Op de Beeck and Baker, 2010). Several previous studies, using single-unit recordings (e.g., Sigala and Logothetis, 2002) and human fMRI (e.g., Folstein et al., 2013, Gillebert et al., 2009, Lerner et al., 2008) have already shown that object processing and learning mostly enhances the selectivity for behaviorally relevant stimulus distinctions, but these studies did not relate learning effects to the pre-training selectivity of individual neurons or voxels and they did not directly compare pre- and post-training representations. Here, we present human functional magnetic resonance imaging (fMRI) evidence which is consistent with this informativeness hypothesis in a design which involves training in complex object tasks such as categorization and object differentiation and which allows for an explicit comparison between the pre-learning functional properties in multi-voxel patterns and the effect of learning.

We adapted the design of a previous study (Op de Beeck et al., 2006) in which we investigated how learning affects the general pattern of response for a trained set of objects. Translated to everyday objects, this previous study showed that training to discriminate among cars can change the pattern of response to the category of cars as a whole. However, this previous study could not reveal to what extent effects of training would be related to informativeness, because the study did not include a pre-training measurement (nor post-training) of the selectivity for those stimulus features which were relevant for training, that is, the stimulus properties which differentiate among individual cars. To remediate this problem, the current study targets such selectivity for the fine differences between members of one of the same set of objects. The study includes three sets or ‘types’ of objects, and for each a further distinction between two sub-classes of objects. Subjects were scanned twice, once before and once after subjects were trained to distinguish exemplars within two of the three object types. We included two training regimes, categorization and differentiation, because previous studies have suggested that the training task can influence the type of changes at the neural level (Wong et al., 2009). The third object type was not trained between the two scan sessions, so that the sub-class selectivity for that object type served as a control. Importantly, the stimulus features defining the sub-classes were relevant during training, allowing us to relate training effects to the informativeness of neural units prior to training.