Image analysis and image synthesis are the goals of computer vision and computer graphics, respectively. These research areas represent the domains into which the work presented in this dissertation fall. Specifically, we present work on three problems: segmentation and estimation of large disparity motion, simulating the reflectance for rough surfaces using microfacet models, and the perception of material reflectance. We present a novel framework for motion segmentation that combines the concepts of layer-based methods and feature-based motion estimation. We demonstrate our approach on image pairs containing large inter-frame motion and partial occlusion. The approach is efficient, and it successfully segments scenes with inter-frame disparities beyond the scope of previous methods. We also present an extension that accounts for the case of non-planar motion, and applications of our method to automatic object removal and to structure from motion. The Bidirectional Reflectance Distribution Function (BRDF) describes the way a surface reflects light. Microfacet reflectance models have been shown to work well for simulating the interaction of light with a rough surface. We give an overview of the existing techniques for reflection modeling and show how these techniques can be extended to handle refraction in a unified framework. To this end, two new derivations are presented for computing quantities required for refraction as well as a result that is (to our knowledge) previously unpublished. While BRDFs allows for a complete radiometric description of light reflecting from a surface, they are complex mathematical objects that can be difficult to use in practice. Our aim is to construct a low-dimensional, perceptual space for BRDFs that can be easily navigated. To this end, we design and carry out a comprehensive psychophysical study of the perception of measured reflectance. This is the largest study of its kind to date, and the first to use real material measurements. In addition, we introduce a new multidimensional scaling (MDS) algorithm for analyzing ordinal data that unlike existing methods is both efficient and optimal. We use the results of our study to construct a perceptual space of these BRDFs and introduce a new method for perceptual construction of novel BRDFs

The best-performing artificial intelligence systems for object recognition are deep neural networks (DNNs). For several years now, engineers and neuroscientists have claimed that DNNs show human-level performance in object recognition. However, Barbu et al. (2019) reported accuracies of around 30% for state-of-the-art object recognition systems when testing on their better-controlled image set – ObjectNet. How do humans perform on ObjectNet? In our experiment, we tested 25 undergraduates' ability to classify the ten categories of objects in ObjectNet that the Deep Convolutional Neural Networks (DCNNs) found easiest, moderate and hardest. Although humans and DCNNs had similar overall accuracy levels, there were some everyday, basic-level categories for which machine performance was much lower than humans. The pattern of errors generated by the DCNNs was about as similar to human error patterns as the individual human error patterns were to each other. Implications of these results and plans for future work are discussed.

The inverse base-rate effect is a robust irrational bias that arises when people face ambiguity. The most prominent theories of this irrational bias depend on prediction error. In this study, we gradually removed elements of a predictive learning design to test the extent to which error-driven processes underlie this bias. In our first experiment, we removed explicit feedback by implementing the inverse base-rate effect in an observational learning procedure. In our second study, we further removed any causal relationship between stimulus features and category labels by moving towards an unsupervised learning procedure. This removed any information participants could use to identify category labels. In both experiments, the inverse base-rate effect persisted and remained robust. This outcome suggests that this irrational bias is independent of supervised learning procedures. We propose that any theories and models of the inverse base-rate effect must manage information encoding and connection updates without explicit prediction error.

This paper examines the ability of a dual-system, formal modelof categorization COVIS (Ashby, Paul & Maddox, 2011) topredict the learning performance of participants on the six cat-egory structures described in Shepard, Hovland and Jenkin’s(1961) seminal study. COVIS assumes that category learningis mediated by two dissociable neural systems that compete tocontrol responding. The verbal system explicitly tests verbal-izable rules, whereas the implicit system gradually associateseach stimulus with the appropriate response. Although COVISis highly influential, there are no published evaluations of theformal model against classic category learning data (COVIS ismost typically applied heuristically to the design of new exper-iments). In the current paper, we begin to address this gap inthe literature. Specifically, we demonstrate that COVIS is ableto accommodate the ordinal pattern found by Shepard et al.,provided that adjustments consistent with the model’s theoret-ical framework are made.

There is a growing interest in alternative explanations to the dual-system account of how people learn category structures varying in their optimal decision bounds (unidimensional and information-integration structures). Recognition memory performance and hippocampal activation patterns in these tasks are two interesting findings, which have not been formally explained. Here, we carry out a formal simulation with SUSTAIN (Love, Medin, & Gureckis, 2004), an adaptive model of category learning, which had great success in accounting for recognition memory performance and fMRI activity patterns. We show, for the first time, that a formal single-system model of category learning can accommodate recognition performance after learning and is consistent with fMRI data obtained while participants learned these structures.

In this paper, we introduce CALM, a process model that is de-signed to abstract solutions in simple and complex categorylearning tasks. The model includes strong assumptions aboutthe interaction of processes driving learning behavior, typicallyaddressed in terms of feature attention, stimulus generaliza-tion, rule abstraction and knowledge partitioning. We presentsimulations of CALM, showing that the model can accountboth for systematic variations in Type II category difficulty,and for individual differences in extrapolation of an XOR cat-egory structure.

Some argue that category learning is mediated by two com-peting learning systems: one explicit, one implicit (Ashby etal., 1998). These systems are hypothesised to be responsi-ble for learning rule-based and information-integration cate-gory structures respectively. However, little experimental workhas directly investigated whether people are conscious of cat-egory knowledge supposedly learned by the implicit system.Here we report one experiment that directly compared explicitrecognition memory for exemplars between these two categorystructures. Contrary to the predictions of the dual-systems ap-proach, we found preliminary evidence of superior exemplarmemory after information-integration category learning com-pared to rule-based learning. This result is consistent with thehypothesis that participants learn information-integration cate-gory structures by using complex rules.

Considerable behavioral evidence has been cited in support ofthe COVIS dual-system model of category learning (Ashby &Valentin, 2016). The validity of the inferences drawn fromthese data critically depend on the accurate identification ofparticipants’ categorization strategies. In the COVIS literature,participants’ strategies are identified using a model-based anal-ysis inspired by General Recognition Theory (Maddox, 1999).Here, we examine the accuracy of this analysis in a model-recovery simulation. We find that participants can appear tobe using implicit, procedural strategies when their responseswere actually generated by explicit rule-based strategies. Theimplications of this for the COVIS literature are discussed.