direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Page Content

Models of Neural Systems

PhD Theses

On Modeling Data from Visual Psychophysics
Citation key Dold2012
Author Dold, Hannah
Year 2012
School Technische Universität Berlin
Abstract Computational models are an essential tool in the area of visual psychophysics. Experiments are derived from them, they provide a framework for thinking, and they formalize our understanding of psychological processes. Models in visual psychophysics can be either mechanistic, inspired by nature and imitating psychological processes, or they can be statistical, exclusively describing a function that maps input to output. If a model is defined as a Bayesian graphical model, Bayesian inference allows to estimate the parameter posterior distribution. In addition to the classical point estimate, the posterior distribution provides further diagnostics of the model that influence the conclusions drawn from the model and its parameters. In this thesis I show how we can model visual processing at two very different levels using a Bayesian approach. A multi-resolution filter model with a static nonlinearity is the standard model for early spatial vision. Its parameters are assumed to be biologically plausible. Conclusions about a model that are based on specific estimated parameter values are only valid if the parameters are constrained by the data at hand. I show that Bayesian statistics led to the discovery of weaknesses in the vision model formulation and missing experimental data. Visual inspection of the posterior suggested starting points for improvements, and allowed me to tailor the early vision model to its founding psychophysical data. The psychometric function relates stimulus intensity to behavioral performance and the function's parameters are estimated with data from a single experimental condition. I present data that suggests that in some situations a parameter is shared across conditions which would require all conditions to be estimated simultaneously. Instead of building a more complex model, I present how Bayes rule can formulate the inference procedure for psychometric functions to fit many experimental conditions in an effectively simultaneous fashion. This joint procedure allows us to estimate a common parameter across conditions on the basis of established routines. A common issue in modeling, for example, early spatial vision or a number of psychometric functions is that several experimental conditions have to be combined. This requires access to substantial amounts of data. When I started working on this thesis there was no adequate tool for handling psychophysical data. This, combined with vastly differing storage formats and contents between laboratories and experiments, makes a reanalysis or meta-analysis of earlier experiments difficult. I developed a new software package for Python that stores and manages experimental data including annotations in a database. It is open-source, operating system independent, and does neither require knowledge of the structured query language, nor of databases. This thesis demonstrates how close the interaction of modeling, theory, and experiment is, and how important a good experimental basis is for accurate modeling. Well formulated models, elaborate statistical procedures, and extensive data sets — combined they are able to work as engine for further research.
Link to publication Download Bibtex entry

Zusatzinformationen / Extras

Quick Access:

Schnellnavigation zur Seite über Nummerneingabe

Auxiliary Functions