150 Psychologie
Refine
Year of publication
- 2013 (1) (remove)
Document Type
- Doctoral Thesis (1) (remove)
Language
- English (1)
Has Fulltext
- yes (1)
Is part of the Bibliography
- no (1)
Institute
- Informatik (1) (remove)
The human brain is an unparalleled system: Through millions of years of evolution and during a lifespan of learning, our brains have developed remarkable abilities for dealing with incoming sensory data, extracting structure and useful information, and finally drawing the conclusions that result in the actions we take. Understanding the principles behind this machinery and building artificial systems that mimic at least some of these capabilities is a long standing goal in both the scientific and the engineering communities. While this goal still seems unreachable, we have seen tremendous progress when it comes to training data-driven algorithms on vast amounts of training data, e.g. to learn an optimal data model and its parameters in order to accomplish some task. Such algorithms are now omnipresent: they are part of recommender systems, they perform speech recognition and generally build the foundation for many semi-autonomous systems. They start to be integral part of many technical systems modern technical societies rely on for their everyday functioning. Many of these algorithms were originally inspired by biological systems or act as models for sensory data processing in mammalian brains. The response properties of a certain population of neurons in the first stages of the mammalian visual pathway, for example, can be modeled by algorithms such as Sparse Coding (SC), Independent Component Analysis (ICA) or Factor Analysis (FA). These well established learning algorithms typically assume linear interactions between the variables of the model. Most often these relationships are expressed in the form of a matrix-vector products between a matrix with learned dictionary-elements (basis vectors as column vectors) and the latent variables of these models. While on the one hand this linear interaction can sometimes be justified by the physical process for which the machine learning model is proposed, it is on the other hand often chosen just because of its mathematical and practical convenience. From an optimal coding point of view though, one would generally expect that the ideal model closely reflect the core interactions of the system it is modeling. In vision for example, one of the dominant processes giving rise to our sensory percepts are occlusions. Occluding objects are omnipresent in visual scenes and it would not be surprising if the mammalian visual system would be optimized to process occluding structures in the visual data stream. Yet, the established mathematical models of the first stages of the visual processing path (like, e.g., SC, ICA or FA) all assume linear interactions between the active image components. In this thesis we will discuss new models that aim to approximate the effects of occluding components by assuming nonlinear interactions between their activated dictionary elements. We will present learning algorithms that infer optimal parameters for these models given data. In the experiments, we will validate the algorithms on artificial ground truth data and demonstrate their ability to recover the correct model parameters. We will show that the predictions made by these nonlinear models correspond better to the experimental data measured in-vivo than the predictions made by the established linear models. Furthermore, we systematically explore and compare a large space of plausible combinations of hyperparameters and preprocessing schemes in order to eliminate any effects of artefacts on the observed results. Training nonlinear sparse coding models is computationally more demanding than training linear models. In order to perform the numerical experiments described in this thesis we developed a software framework that facilitates the implementation of massive parallel expectation maximization (EM) based learning algorithms. This infrastructure was used for all experiments described in here, as well as by collaborators in projects we will not discuss. Some of the experiments required more than 1017 floating point operations and were run on a computer cluster running on up to 5000 CPU Cores in parallel. Our parallel framework enabled these experiments to be performed.