To generalization and beyond: an integrated approach to perceptual and expert learning

Visual Perceptual Learning takes up a great deal of our lives, where for some of us, it leads to becoming an expert (e.g., radiology). The tradition of training specific outcomes is challenged by generalization to other features, retinal locations or other general domains (e.g., multi-tasking). Yet, the mechanisms underlying transfer lead to inconsistent findings, such as whether playing video games results in transfer. The first study aimed to integrate as many factors known to contribute to transfer, which resulted in a gamified perceptual learning paradigm. A simple game was given to the control group, whilst the experimental group was trained with the paradigm for 10 hours. We observed effects of transfer to other objects, locations, and domains (e.g., executive functioning), which were stronger in the experimental group for object transfer. These positive effects led to an investigation of the neural underpinnings of these findings in a second fMRI study. Before and after a 15 hour training on the same paradigm/game, subjects participated in a fMRI session. Results showed enhanced neural sensitivity for low-level stimuli in the brain's object region with training. An unfortunate lack of signal in the scanning paradigm caused problems picking up expected and more subtle sensitivity differences.

The surge of Convolutional Neural Networks (CNNs) in VPL resulted in two more studies with a computational approach on visual object expertise. Study three aimed to investigate intra-class specificity in different instances of a CNN and correlate these representations with the behavioral and neural representations in novices and experts. The same CNN architecture is differently trained to obtain a similar state of mind as to that of a newborn (i.e., random network), novice (i.e., pretrained) and expert (i.e., fine-tuned) in birds. Results showed increased bird expertise from random, over pretrained, to fine-tuned. The effect of expertise is best associated with neural representations in frontal cortex of real-life bird experts. A fourth study focused upon objects that received animal characteristics (i.e., zoomorphic objects). More specifically, this study aimed to investigate the discrepancy in how the human brain represents these objects versus what has been previously reported in CNNs: closer to real animals (i.e., animal bias) versus more like real objects (i.e., object bias) in the latter. Several hypotheses based on differences between the human brain and the CNNwere tested: extensive animal bias, animacy, face-body object sensitivity, natural image categories, and shape-texture bias. Only an explicit animal bias training led to an animal bias in zoomorphic objects.

All these studies contributed to the knowledge on VPL by means of state-of-the-art methods.