Representational and Robustness Principles of the Retinal Population Code

Abstract

A comprehensive theory of neural computation requires an understanding of the fundamental principles underlying the neural population codes that are employed throughout the brain. In this thesis, we focus on principles underlying the ganglion cell population code employed by the retina, which is the earliest circuit involved in visual processing. We investigate these principles using data-driven approaches within frameworks motivated by concepts from other quantitative disciplines, which is a methodological unifying theme. In Chapter 1 we place our work in the broader scientific context with a brief overview of the relevant conceptual background. Then in Chapter 2, we provide an overview of the relevant methodological background that is common to the specific methods and approaches discussed in the subsequent chapters.

As hinted at by the title, there are two key conceptual themes of focus in this thesis: the nature of what the retinal population code represents (e.g. which features of the external visual stimulus space are encoded), and how the retinal system functions in spite of noisy components. While the three main projects discussed in Chapters 3 through 5 are related to one or both of these themes, each is aimed at a different level of analysis. Our work in Chapter 3 focuses on the computational goals of the retinal system, which is investigated via a latent variable model approach. Building on this work, in Chapter 4 we investigate the nature of the representation of the noise-robust population codewords identified in Chapter 3, and discuss preliminary work that focuses on biologically-plausible algorithms for decoding in Chapter 5. We conclude with a discussion of open questions and future directions related to this thesis work.