The brain is a distributed computational system. While the brain has been understood to exhibit at least weak modularity for over a century, numerous important questions about the degree and consequences of that modularity remain. My thesis work consisted of three projects, which are each related to distinct questions about the nature and function of modularity in the brain. First, I investigated how the neural code within distinct regions of the brain could be made reliable in the face of the unreliability of individual neurons. In this work, I found that in- creasing the dimensionality of neural representations through conjunctive mixing of multiple stimulus features improves the reliability of those representations by orders of magnitude relative to representations without mixing. This work provides an explanation for a commonly observed phenomenon in experiments: The apparent random conjunctive mixing of stimulus features in single neurons. However, it also intersects with questions about modularity. In particular, the benefits that can be derived from this conjunctive mixing depend strongly on the size of the neural population available to participate in the representation – that is, on the size of a particular brain region. Further work will explore how this pressure for larger regions is tempered by other constraints in the brain. Second, I investigated how the neural code across distinct regions of the brain could be made reliable in the presence of multiple heterogeneous objects that are represented only partially within each region. As an example, two cats in the world have both visual and auditory representations in the brain. To guide behavior, the brain must integrate these different facets of the same animals. We describe the necessary conditions for this integration process to be reliable. Further, we outline a tradeoff, in which the fidelity of each individual representation can be increased at the cost of a greater risk of catastrophic integration errors, in which the auditory features of one cat are integrated with the visual features of the other cat. More generally, this work provides another constraint on the modularity exhibited by the brain. We show that redundancy in the information represented by distinct brain regions is absolutely necessary for reliable integration. Thus, this work illustrates a pressure for the brain to use fewer distinct modules, so that it can satisfy the overall goal of redundancy reduction for producing efficient neural codes. Third, I performed electrophysiological experiments to investigate the role of a particular brain region, the lateral intraparietal area (LIP), in two distinct tasks. These experiments revealed that the putative function of LIP in the representation of both visually guided actions and the behavioral relevance of different parts of the visual field is highly task- dependent. In particular, our results indicate that LIP may serve this role primarily in the context of directed tasks, while it is less engaged in undirected, free-viewing behavior. These results illustrate the extreme context-dependence of many of our inferences about the functional role of different brain regions. Together, my thesis work refines our understanding of the role and reason for modularity in the brain – and points to numerous directions for future work. In particular, my work provides many of the necessary tools for beginning to build a more comprehensive normative theory of neural modularity that will be necessary to a comprehensive understanding of the brain.