By training I am a theoretical computer scientist with the focus on artificial intelligence and machine learning. Towards the end of my studies, I have increasingly often tapped neuroscience and psychology for inspirations on how to build intelligent systems. This dual interest lead me to join a program in Neuroinformatics at the Unviersity of Edinburgh, where my interest in the brain grew further, eventually leading to my PhD thesis on the topic of Development of functional properties in primary visual cortex under the supervision of James Bednar. My first post-doctoral post was in the lab of Tom-Mrsic Flogel at UCL, where I had the chance to brush up my machine learning skills while investigating novel ways to estimate receptive fields in local populations of mouse primary visual cortex neurons. Next I have switched my attention to studying cat early vision using large-scale spiking neural modelling approach under the co-supervision of Andrew Davison and Yves Frégnac at the UNIC department of CNRS. I then continued to pursue the study of early vision at Institute de la Vision in Paris in the group of Pr. Ryad Benosman, where I applied the large scale models of V1 to the problem of optogenetic based cortical visual prosthesis, as well as utilized my experience in artificial intelligence and cortical modeling to pursue projects in the field of Neuromorphic Vision. Currently I continue to purse all these lines of research at my alma mater, the Faculty of Mathematics and Physics of Charles University in Prague.
Current research interests
- visual system
- computational neuroscience
- system identification
- machine learning
- neuromorphic computation
Available student projects
Would you like to contribute to our research? This is a list of projects available for interested students. Most are designed to be completed within 3 to 6 month, but some can be expanded into longer projects even full PhD scope. If you are interested in working on any of the projects please contact me.
Computational Neuroscience Projects
Unifying retinal mozaik model with activity based development.
During post-natal development, primary visual cortex undergoes remarkable functional organization resulting - among others - in expression of topologically smooth orientation map across it’s surface. The most common type of explenation for this phenomena is activity based development, whereby internally generated or visually driven activity coupled with plasticity in the thalamo-cortical and corico-cortical pathway induces gradual establishment of the orientation maps. LISSOM based familiy of models is an example of such activity + plasticity driven models. An alternative explanation has been proposed by Ringach (see also this) , in which the initial orientation maps are directly established by the very specific geometric properties of retinal ganglion cells RFs positions in visual space: retinal mozaiks. However, this explanation can account only for initial very weak orientation maps, and low orienation selectivities of individual neurons in particular, and it is clear that the system has to undergo major further refinement in order to match the experimentally observed adult state. The goal of this project is to combine the two hypothesis of orientation map development and investigate their possible interactions. Specifically retinal mozaiks will be introduced into a LISSOM model, thus inducing the initial orientation maps based on Ringach et al. theory. This will be followed by simulation of the activity and plasticity driven development, which should lead to refinement of the intial maps. The correspondance between the initial retinal mozaik induced map with the final developed map will be assesed, and possible advantages of such dual orientation map development mechanism will be investigated.
Embedding of detailed compartmental neuron models into large-scale model of V1.
One of the ongoing projects in our group is development of large-scale integrative model of cat primary visual cortex (V1). This model is based on the Adaptive-Exponential Leaky Integrate and Fire neuron model, which reduces biological neurons to a point process, ignoring its geometrical properties. In this project student will embed single compartmental model of V1 pyramidal neuron into the large scale point process simulation available in the group, and investigate the behavior of the added detailed neuron under the influence of the input coming from the large scale V1 simulation, focusing on properties influenced by the neuron’s geometry.
Local-field potentials (LFP) in a large-scale model of cat primary visual cortex.
One of the ongoing projects in our group is development of large-scale integrative model of cat primary visual cortex (V1). LFP is an electrophysiological signal generated by the summed electric current flowing from multiple nearby neurons within a small volume of nervous tissue. The goal of this project is to investigate the LFP signals that would be generated in our simulations of V1. The V1 model under investigation does not explicitly contain LFP signals, only the sub-threshold and spiking responses of individual neurons are available. Therefore one of previously proposed models of LFP signals such as this one will be used to generate artifical LFP signals based on the outputs of the V1 simulation. This will be followed by thourough analysis of the resulting LFPs and results compared to previous findings, including recent data recorded at UNIC by the Yves Frégnac group.
Processing of higher order correlations in early visual cortex: a computational modelling investigation
Recent experimental studies have revealed differences in how neurons in primary (V1) and secondary (V2) visual cortices (the first two stages of visual cortical processing) process high-order statistics in visual scenes, indicating emergence of sensitivity to 2nd order correlations in V2 but not V1 neurons [1,2,3]. However, the neural mechanisms of such sensitivity, and their implementation in biological neural substrate remain unknown. To address this question, we will use detailed large-scale spiking neural network modelling paradigm to formulate hypothesis of neural circuits that can explain such neural functional properties. The student will be responsible for the first stage of longer-term project, in which he will implement a set of specialized visual stimuli as in . Subsequently, student will test an existing model of V1 using the same experimental paradigm as in  for sensitivity to 2nd order correlations.
Stimulation protocols for cortical visual prosthesis.
Recently we have applied the large-scale models developed in our team to the problem of cortical visual prosthesis. New approach to sensory prosthetics is being developed, wherby the the cortex is stimulated via opto-genetic tools, which are being translated from mice to higher-order mammals including primates. While all the technological components of the visual prosthesis are still under development, an important question remains open: how to stimulate the cortex to elicit percepts that are close to those due to the perception of the given stimulus under normal vision. This is where our large-scale modelling approach comes in. Using our V1 model simulations to test potenial stimulation strategies, we are making progress at answering this question. Currently, we have gained insights on how to eleicit simple canonical visual stimuli, specifically sinusoidal gratings. In the next step the student will be responsible for expanding the design and analysis to generic stimulation protocol capable of eliciting arbitrary visual stimuli. The current protocol can be straightforwardly expanded to this general case . The student’t responsibility will be to implement this new stimulation protocol in our simulation framework, test the protocol in our model of V1, and implement and perform all the required analysis. Strong programming and analytical skills required. Knowledge of Python, computation neuroscience or neurobiology of visual system a plus.
Deep biologically structured system identification approaches for studying primary visual cortex function.
A common approach for studying the function of early sensory systems is to determine the relationship between sensory inputs and associated (experimentally recorded) neural responses. In the past, mostly linear , or shallow non-linear techniques were utilized, leading to limited predictive and consequently explanatory power of models fitted in this way. More recently, the popular deep convolutional architectures were successfully tested on the neural data [2,3]. These general, machine-learning motivated models ,however, ignore the known anatomical and functional architecture of visual system. Recently, we have presented a multi-stage model of V1 which reflected some of the most prominent features of the retino-cortical pathway , and demonstrated that such incorporation of V1 biology can improve performance in comparison to state-of-the-art models. In this project we will built upon these early results, and develop novel deep-architectures inspired by the deep convolutional networks, but enriched by biologically inspired elements. The student will be responsible for designing, implementing and subsequently testing the new models on neural population recordings from cat primary visual cortex. This project will be undertaken in collaboration with experimental lab of Yves Fregnac, CNRS, France, and computational lab of Dan Butts, University of Maryland. Prior experience in machine learning is a plus.
Software engineering projects
Mozaik data-store module based on object oriented database.
Mozaik is a an automated workflow for large-scale neural simulations, with a highly modular architecture. One of the core Mozaik modules is a data-store, in which recordings from simulations richly annotated with metadata regarding experimental context are stored. Currently the data-store module is implemented as a database-like system based on Neo library for internal representation of recorded data. The goal of this project is to develop an alternative data-store module based around dedicated key-value database such as BerkelyDB or CodernityDB.
A model inspection/visualization tool for Mozaik.
Mozaik is a an automated workflow for large-scale neural simulations. The model of primary visual cortex developed in our lab, and implemented in Mozaik, has a complex connectivity structure. Although there are various tests that the connectivity has been realized as expected, currently, there is no easy way to visualize the network spatial structure and connectivity in Mozaik. The aim of this project is to develop a model inspection and visualization tool, for Mozaik, possibly building on existing tools such as ConnPlotter, Moogli, and NeurAnim.
Sumatra integration with Mozaik.
Mozaik is a an automated workflow for large-scale neural simulations. Sumatra is a tool for provenance tracking. Sumatra shares several features with Mozaik, but it also posses features that would enhance the Mozaik workflow. The goal of this project is to integrate Sumatra with Mozaik, and remove overlapping features from Mozaik and delegating them to Sumatra, in line with long term goal of outsourcing as much functionality from Mozaik to dedicated tools. This project is suitable for students with interest in Neuroinformatics and moderate skills in Python and versioning systems.
Web development projects
Open Vision project.
Mozaik is a an automated workflow for large-scale neural simulations. Inspired by the OpenWorm initiative, this project strives to bring neural based modelling of vision to the public. It will seek to engage the cognitive sciences enthusiast community into coordinate effort to build a comprehensive model of early and higher vision. We envision multiple phases of the project:
(1) Build a server running mozaik based V1 model and serve it on the new Open Vision website. The website will allow any member of public to submit a video and receive back the responses of selected model cells.
(2) Develop a web frontend to the Mozaik toolkit and use it to expand the Open Vision website to allow full configuration of the served model. Publish more models and experimental protocols already develop in our group.
(3) Expand upon 1 and 2 to build full open science platform similar to OpenWorm project, and build striving community around it.
Graphical user interface for Mozaik datastore and analysis.
Mozaik is a an automated workflow for large-scale neural simulations. Mozaik automatically records data from simulations, annotates it with metadata regarding experimental context, and stores them in an internal data-store. An query based interface allows analysis and visualization modules to efficiently navigate through the stored data based on the attached metadata. Currently, Mozaik offers only programatic API to perform these interactions with data-store. The goal of this project would be to write a HTML based graphical user interface frontend, to the Mozaik data-store, that will allow users to conveniently and interactively navigate and select data from the data-store and subsequently execute on them anaysis and visualization routines from Mozaik libraries.
Deployment of Mozaik on a HPC platrform.
Mozaik is a an automated workflow for large-scale neural simulations. Mozaik depends on a moderate software stack including PyNN as a simulator independent model specification language, and Nest as the simulator of choice in our projects. Currently we deploy Mozaik (together with the software stack) on a local cluster, however already at this relatively small scale we are aware of number of inefficiencies in terms of its performance in the parallel environment. Furthermore, in future we would like to deploy Mozaik on a large-scale High Performance Computing (HPC) platform such as ADA. The goal of this project is to test and optimize Mozaik and it’s underlying software stack to run efficiently on the local cluster, and subsequently scale it up to a large-scale HPC platform. This project is suitable for students with experience and interest in parallel programming and HPC.