Ferenc Huszár

I am a master's student in information engineering at the University of Technology in Budapest. The expected year of my graduation is 2009. My interests include topics of NIPSian machine learning, specifically nonparametric Bayesian methods. Beyond classical applications of machine learning I am particularly interested in application of ideas and methods from machine learning to biologically, cognitively relevant learning problems. I am currently working on model-based analysis of cognitive scientific experiments under the supervision of Máté Lengyel.
My Erdős number is 5 ( Paul Erdős -> Aviezri S. Fraenkel -> Alberto Apostolico -> Jotun Hein -> Péter Ittzés -> Ferenc Huszár ) My name is pronounced as [ˈfɛrɛnts ˈhusaːr].

Teaching

Current courses:

Past courses:

Current research projects

Ideal observer-based analysis of cognitive scientific experiments
with Máté Lengyel, Uta Noppeney and Peter Dayan

A central challenge in cognitive science is to measure and quantify experimentally the mental representations humans (and other animals) develop -- in other words, to "read" subjects' minds. In order to eliminate potential biases in reporting mental contents due to verbal elaboration, subjects' responses in experiments are often limited to simple binary decisions or discrete choices that do not require conscious reflection upon their mental contents. However, it is unclear what such impoverished data can tell us about the potential richness and dynamics of subjects' mental representations. To address this problem, we used ideal observer models that formalise choice behaviour as a quasi-optimal (stochastic) function of subjects' representations in long-term memory, acquired through prior learning, and the information currently available to them. Bayesian inversion of such models allowed us to infer subjects' mental representation from their choice behaviour in a task as simple as the standard one-back task -- in which successively presented items have to be judged as being the same or different. In comparison with earlier methods developed along similar lines (eg. Sanborn & Griffiths, NIPS 2008), our method does not require the introduction of several trials of a special-purpose psychophysics task and thus has sufficient temporal resolution to track as mental representations develop through learning.

Gibbs sampling for the infinite sites model

Motivation: Finding the probability of a set of genomes being related to a known or unknown common ancestor or more generally evaluating a posteriori expectation of some function under a given probabilistic model of sequence evolution has become a central challenge in computational population genetics. Even for models as limited as the infinite sites model (ISM) exact calculation of such quantities is intractable, thus we have to resort to approximation methods, typically Monte Carlo integration. Previous studies investigated different forms of importance sampling and concentrated on refining the proposal distribution. Although Markov chain Monte Carlo (MCMC) techniques have been observed to be superior to importance sampling in many problems, we are not aware of any previous work investigating the application of these techniques in this context.
Results: In this project, we have developed a Gibbs sampling algorithm for simulation of ancestral histories and estimation of expectations under the infinite sites model. We proved the irreducibility of teh resulting chain and are about to test the method's performance empirically on simulated and real-world datasets.

Publications

Technical report

Ferenc Huszár and Stephen G. O'Keeffe

Corner cutting approaches to Ethier-Griffiths-Tavaré recursions.
Technical report, Genome Analysis and Bioinformatics Group, Department of Statistics, University of Oxford, 2008.
[ www | .pdf ]

Given a sample of sequences (e.g. nucleotide or amino acid sequences) from a population, it is a standard problem to calculate its probability under probabilistic models of evolution in order to make estimates of quantities such as mutation rates, and produce accurate models of population structure, selection and recombination. A commonly used simple model of sequence evolution is the infinite sites model. In the period 1987-1995 Ethier, Griffiths and Tavaré published a series of recursions for calculating lieklihoods in this model. Hovewer, exact solution of these recursions is computationally hard for large datasets and approximations must be made in odred to accelerate calculations, whilst maintaining reliability of results. In this project, we implemented the exact recursions and applied corner cutting methods to accelerate the computations. Corner cutting is a deterministic approximation method, that has been applied successfully for estimating likelihood in models for statistical alignment and sequence evolution. In this report we give a summary of our results, observations and prospects for further research.

Book chapter

Eörs Szathmáry, Zoltán Szatmáry, Péter Ittzes, Szabolcs Számadó, István Zachar, Ferenc Huszár, Anna Fedor, and Máté Varga

In silico evolutionary developmental neurobiology and the origin of natural language.
In C. Lyon, C. Nehaniv, and A. Cangelosi, editors, Emergence of Communication and Language, chapter 8, pages 151-188. Springer Verlag, 2007.
[ www | .pdf ]

It is justified to assume that part of our genetic endowment contributes to our language skills, yet it is impossible to tell at this moment exactly how genes affect the language faculty. We complement experimental biological studies by an in silico approach in that we simulate the evolution of neuronal networks under selection for language-related skills. At the heart of this project is the Evolutionary Neurogenetic Algorithm (ENGA) that is deliberately biomimetic. The design of the system was inspired by important biological phenomena such as brain ontogenesis, neuron morphologies, and indirect genetic encoding. Neuronal networks were selected and were allowed to reproduce as a function of their performance in the given task. The selected neuronal networks in all scenarios were able to solve the communication problem they had to face. The most striking feature of the model is that it works with highly indirect genetic encoding—just as brains do.

Posters

Ferenc Huszár, Uta Noppeney, and Máté Lengyel

Machine learning in the service of understanding human learning: an ideal observer-based analysis of mental representations.
Workshop on Machine Learning Meets Human Learning, held at NIPS 2008, Whistler, Canada, December 12th, 2008
[ .pdf ]

A central challenge in cognitive science is to measure and quantify experimentally the mental representations humans (and other animals) develop - in other words, to "read" subject's minds. In order to eliminate potential biases in reporting mental contents due to verbal elaboration, subjects' responses in experiments are often limited to simple binary decisions or discrete choices that do not require conscious reflection upon their mental contents. However, it is unclear what such impoverished data can tell us about the potential richness and dynamics of subjects' mental representations. To address this problem, we used ideal observer models that formalize choice behavior as a quasi-optimal (stochastic) function of subjects' representations in long-term memory, acquired through prior learning, and the information currently available to them. Bayesian inversion of such models allowed us to infer subjects' mental representation from their choice behavior in a task as simple as the standard one-back task - in which successively presented items have to be judged as being the same or different. In comparison with earlier methods developed along similar lines1, our method does not require the introduction of several trials of a special-purpose psychophysics task and thus has sufficient temporal resolution to track as mental representations develop through learning.

Péter Ittzés and Ferenc Huszár

Network properties of evolved neural networks.
Poster presentated at ECAgents Review Meeting, held at Roma, Italy, June 2006.
[ .pdf ]

The neural networks evolved by the ENGA system are recurrent networks. The neuron and synapse number vary from agent to agent. The exact comparison of networks of this type is a difficult theoretical problem, therefore we applied simple graph theoretical methods to perform a preliminary analysis.