Project themes in computational brain modelling and brain-like computing

Transcription

1 Project themes in computational brain modelling and brain-like computing Supervisor, contact person: Pawel Herman Department of Computational Biology (CB) Please bear in mind that project details and specific research questions within the proposed themes are discussed individually with students depending on their interests. There is also a lot of flexibility in defining the scope and size of these projects. Some project ideas at the cross-sections of the following themes can be proposed/found. Students will have an opportunity to learn to use dedicated simulation software (with a possibility to rely on Python interface) or exploit their programming competence to build their own computational tools for theoretical or applied research. Projects are organized in three main themes, each of which describes a set of proposed topics. The lists of topics and some project ideas are not meant to be limiting in any sense and can therefore be easily expanded by students own ideas.

2 Theme 1: Computational brain, neural network simulations and brain-like computing algorithms General focus is here on developing, studying and/or applying connectionist (network based) brain models. The proposed topics range from simulating detailed spiking neural networks to investigating and validating more abstract brain-like computing architectures. Projects can be formulated to either address theoretical questions or test the networks functionality in applications. 1.1 Simulations and analysis of neural network models with emphasis on attractor memory networks There have been a range of theoretical concepts of brain computations proposed in computational neuroscience. Among the connectionist (network-based) approaches to modelling brain function, an attractor theory of neural computations has recently received particular attention. The functionality of attractor networks has been found helpful in explaining various perceptual and memory phenomena. Consequently, these models can be considered as fundamental components of systems level approach to modelling brain function within the framework of network-of-networks architecture. An implementation of attractor memory models can range from a more biologically plausible networks of spiking neurons to more abstract networks of units with continuous rate-based input/output. More biophysically detailed models with spiking neurons and synapses provide an opportunity to study rich neural dynamics in close relation to biological data, and specifically, recordings from the brain tissue. This way both dynamical and functional aspects of fascinating cortical phenomena can be studied. Such spiking neural network models are usually developed using dedicated simulation software, e.g. Nest, Neuron, Genesis etc. More abstract networks relying on rate-based units (i.e. with non-spiking real-valued input/output) on the other hand allow for constructing larger systems with the aim of exploring functional aspects of the simulated attractor memory system. In this context, both generic theoretical investigations into computational capabilities of memory (learning, memory capacity etc.) as well as specific applications in pattern recognition, whether in a biological or non-biological data mining context, can be pursued. Within this theme other computational theories of the brain, e.g. liquid state machines, can also be studied. In this regard, computational or dynamical aspects as well as application-oriented questions may be explored. Students can make use of existing software simulators or developed their own implementations of network models.

3 a) Studying the effect of different connectivity patterns, network architectures and their dimensionality on the dynamics and function of the attractor model. b) Investigating the sensitivity of the model to the level of biological detail being accounted for (discussion on the required level of complexity and the relevance of biological constraints). c) Exploring population-level (e.g. simple mean-field approximation) approaches to describing the neural dynamics exhibited by a modular attractor network. 1.2 Brain-inspired or brain-like computing algorithms theoretical developments and applications Development of brain models to study neural phenomena, as broadly discussed in topic 1.1 above, often leads to better understanding of the nature and purpose of neural computations. From a broader perspective, these computations can be seen as an inspiring model for novel approaches to generic information processing. Good reputation of neural network architectures in this regard is largely due to the impressive capabilities of information processing in the brain, which robustly handles large volumes of noisy multi-modal data received in continuous streams. Consequently, brain-like computing has long been considered as a particularly appealing concept in a broad field of information science. With the increasing availability of powerful computing platforms and intensive development of brain models as well as a growing body of knowledge about computational mechanisms underlying brain function, there is a surge of interest in adapting these functional aspects to devise algorithms for more generic applications in the field of data mining, pattern recognition etc. These efforts are urgently needed and particularly relevant to real-world problems involving so-called big data, for example in exploratory analysis of large volumes of high-dimensional neuroimaging data for research or clinical purposes. a) Adapting selected brain-like computing paradigms for large-scale data mining, e.g. to perform exploratory search for patterns in brain imaging data (medical diagnostics, see also Theme 3). b) Devising new brain network inspired approaches to generically process temporal or sequential data and/or comparing to the existing state-of-the art attempts. c) General evaluation and validation of brain-like computing algorithms on speech recognition, computer vision or other challenging real-world problems. d) Testing robustness (sensitivity analysis, noise handling capabilities, computational speed) and benchmarking brain-like computing methods against more conventional machine (/statistical) learning techniques on a selected set of benchmark problems. e) Devising network hierarchical architectures to model behavioral phenomena like prediction, expectation and filtering (at a reasonable level of abstraction).

4 1.3. Bayesian learning in spiking neural network models The theoretical framework of Bayesian statistics is commonly considered as an intuitively attractive model for representing and processing uncertain information in the brain. It has received a lot of attention in computational studies of learning and inference mechanisms underlying brain function. Since the Bayesian machinery for capturing probabilistic information in distributed neural networks corresponds to a commonly accepted and biologically inspired Hebbian idea of synaptic processes taking place in the connections between cells, there have been numerous attempts to adapt Bayesian inference as an unsupervised learning principle. In this context it is particularly challenging to translate Bayesian algorithms from abstract theoretical formulations to biologically plausible computations in spiking neural network models. In the following projects a student will support the ongoing research efforts in the lab, where our own Bayesian learning scheme (Bayesian Confidence Propagating Neural Network, commonly referred to as BCPNN) has been developed. a) Benchmarking synaptic Bayesian-Hebbian learning rules in a spiking sparse activity cortical associative memory and/or in popular pattern recognition/machine learning tasks. b) Simulation and analysis of spiking neural network models pre-trained with a Bayesian learning algorithm; studying implications of Bayesian learning on the network dynamics and function. 1.4 Visualisation in large-scale neural modelling Visualisation is one of the most neglected aspect of a rapidly developing field of computational biology. Only recently can we observe an emerging trend for combining neural simulation frameworks with visualisation software. Still there are a plethora of challenging problems that need to be urgently addressed (high-dimensional data, pre-processing, integration with a simulation software, 3-D visualization of ongoing brain model activity, demands for purely visual aspects, interactive environment) to render visualisation a practical tool in computational studies. This is envisaged to facilitate computational modelling and assist in demonstrating scientific findings. a) Visualisation of existing data produced by models (different types of high-dimensional spatiotemporal data are available). b) Conceptual integration with simulating environment to help with data pre-processing (or postprocessing) and facilitate interactive mode with the user. c) Review of the state-of-the-art methodology and a motivated choice of a tool for the computational problem at hand.

5 1.5. Investigations into parallel implementations and simulations of brain network models on graphics processing unit (GPU) clusters Simulations of large-scale brain models have gained growing importance in neuroscience mostly due to the better availability of comprehensive sets of relevant experimental data and, certainly, due to continuously increasing computational power. In the broad field of scientific computing the latter factor is particularly appreciated as it allows researchers to expand the complexity and size of their models. The majority of brain models are nowadays deployed on supercomputers. However, their availability is rather limited and they are commonly dedicated to large-scale simulations. Recently, graphics processing units (GPUs) have attracted attention as cheaper and more widely accessible simulation platforms, particularly for prototyping and evaluating models at lower-scales. Developments of GPU environments for neural simulations are still at early stages, especially when compared to supercomputer platforms. This opens up a lot of interesting research opportunities and the proposed projects within this theme could serve as suitable starting points. Students choosing this set of projects should have some prior experience with CUDA or OpenGL. a) Testing of existing GPU software for neural simulation. Potential development and a comparative analysis of parallel implementations of simple neural models (spiking or more abstract neural networks). b) Investigations into parallel simulations of simple brain models (distributed spiking or rate-based models, basic mean-field models etc.) at different scales deployed on GPU clusters using different programming interfaces - OPEN CL and CUDA.

6 Theme 2: Machine learning aspects of brain-inspired learning systems The projects proposed under this theme are concerned with Machine Learning (ML) and computational tools that benefit from brain inspirations but at the same time are not necessarily even considered biomimetic. Unlike for the projects proposed under Theme 1, the focus here is rather on the relevance of separate ideas borrowed from brain research (or more loosely related to computational neuroscience), such as various architectures in artificial neural networks including deep hierarchies, self-organization, local learning in distributed systems, a wide range of unsupervised and reinforcement learning approaches etc., to ML problems. Potential projects can explore the usefulness of ML algorithms inspired by brain computations in specific applications (spatio-temporal pattern recognition, time series prediction, inference in noisy environments under uncertainty, novelty detection, control of agents in computer games etc.) or study the nature and robustness of these biomimetic contributions to ML. 2.1 Exploring computational capabilities and properties of deep neural networks. 2.2 Exploiting self-organization and/or competitive learning principles in network computations. 2.3 Studying the potential of unsupervised or semi-supervised learning methods in problems with limited, unbalanced, noisy and uncertain data. 2.4 Designing/testing algorithms to adaptively control behaviour of virtual agents, e.g. in computer games, possibly with reinforcement learning methodology (or relating to the state-of-the-art methods). More specific project ideas within these broadly formulated and highly exploratory research themes will be added at a later time. For now, please contact Pawel Herman for details.

7 Theme 3: Computational approaches to real-world problems in large-scale data mining, pattern recognition, operational research etc. Projects listed below describe a wide range of applied tasks within broad areas of data mining, pattern recognition, time series prediction and operations research. They all require methods capable of learning from noisy data (sometimes nonstationary) or heuristic approaches (to NPcomplete problems), hence offering an opportunity to identify interesting research questions beyond serving only a specific application. The list of application areas can be easily extended at students initiative and projects will certainly have to eventually be made more concrete. The ambition is to propose, evaluate and examine approaches to specific computationally demanding problems with available data sets for prototyping and benchmarking. In some application areas there is an opportunity to work on real-world data sets exploited in previous data science research projects. 3.1 Brain signal pattern recognition for future applications Pattern recognition and machine learning have significantly advanced the field of biological data analysis leading in consequence to the development of effective diagnostic tools and supporting research efforts. The contribution of novel pattern recognition methods has been particularly appreciated in brain data mining as this new approach allows for exploratory search for spatio-temporal patterns in large quantities of high-dimensional nonstationary recordings of brain activity. The emerging trend is to combine machine learning techniques with brain-inspired computing algorithms to address increasingly demanding objectives of brain signal analysis in novel applications. a) Develop your own approach or build upon the existing approaches to a specific brain signal pattern recognition problem, e.g. electroencephalographic (EEG) signal classification for a braincomputer interface (BCI), automated sleep scoring based on physiological signals including EEG, drowsiness or cognitive load detection, EEG-based epileptic seizure prediction (identifying precursors in high-dimensional brain signal recordings) b) Alternatively, select and compare a few existing state-of-the-art methods. Focus on selected aspects of a brain signal pattern recognition problem of your choice (handling signal, extracting patterns, classifying and interpreting brain signal correlates).

8 3.2 Medical diagnostics Computer-aided diagnosis has been extensively validated in various medical domains, ranging from biomedical image or signal analysis to expert systems facilitating the process of decision making in clinical settings. Although the usefulness of computational approaches to medical diagnostics is beyond any doubt, there is still a lot of room for improvement to enhance the sensitivity and specificity of algorithms. The diagnostic problems are particularly challenging given the complexity as well as diversity of disease symptoms and pathological manifestations. In the computational domain, a diagnostic problem can often be formulated as a classification or inference task in the presence of multiple sources of uncertain or noisy information. This pattern recognition framework lies at the heart of medical diagnostics projects proposed here. a) Define a diagnostic problem within the medical domain and examine the suitability of machine learning, connectionist (artificial network-based), statistical or soft computing methods to your problem. b) Survey the state-of-the-art in computational tools supporting classification of disease symptoms and comparatively examine the diagnostic performance of some of them on a wide range of available benchmark data sets. Define a measure for diagnostic performance. Discuss most recent trends in the field and address some of the urgent challenges for computer-assisted diagnostics in medicine. 3.3 University timetabling problem or other challenging problems in scheduling Planning is one of the key aspects of our private and professional life. Whereas planning our own daily activities is manageable, scheduling in large multi-agent systems with considerable amounts of resources to be allocated in time and space subject to multitude of constraints is a truly daunting task. In consequence, scheduling or timetabling as prime representatives of hard combinatorial problems have increasingly become addressed algorithmically with the use of computational power of today's computers. This computer-assisted practice in setting up timetables for courses, students and lecturers has also gained a lot of interest at universities around the world and constitutes an active research field. Within this topic, students can address a scheduling problem of their own choice or they can use available university timetabling benchmark data and tailor it to the project's needs. An important aspect of such project would be to select or compare different algorithms for combinatorial optimisation, and define a multi-criterion optimisation objective. It could be an opportunity to test computational intelligence and machine learning methodology.

9 3.4 Intelligent control There is a clear trend for smarter machines that are able to collect data, learn, recognize objects, draw conclusions and perform behaviors to emerge in our daily life. Advanced intelligent control systems affect many aspects of human activities and can be found in a wide range of industries, e.g. healthcare, automotive, rail, energy, finance, urbanization and consumer electronics among others. By adapting and emulating certain aspects of biological intelligence this new generation of control approaches makes it possible for us to address newly emerging challenges and needs, build large-scale applications and integrate systems, implement complex solutions and meet growing demand for safety, security and energy efficiency. a) Select a real-world control problem (traffic control, energy management, helicopter or ship steering, industrial plant control, financial decision support and many others) and propose a new approach using machine learning and soft computing methodology (computational intelligence) that enhances functionality, automatisation and robustness when compared to classical solutions. b) Demonstrate functional (and other) benefits of computationally intelligent control approaches in relation to the classical methodology in a range of low-scale control problems (benchmarks). Discuss a suitable framework of comparison and potential criteria. c) Consider a robotic application with all constraints associated with autonomous agents and realworld environments (which can be emulated in software). Propose computationally intelligent methods to enable your agent to robustly perform complex tasks (learn from the environment, evolve over time, find solutions to new emerging problems, adapt to new conditions etc.). 3.5 Optimisation and parameter search in computational modelling Model's parameters have a decisive effect on its behaviour and dynamics. Search for parameters is at the same time the most tedious component of computational modelling. Neural simulations are no exception. On the contrary since they account for nonlinear and stochastic effects in brain data, parameters need to be carefully tuned to obtain a desirable functional and/or dynamical outcome. This optimisation procedure is commonly carried out manually on a trial-and-error basis. It is thus desirable to automatise this tedious process by providing an effective parameter search and optimisation scheme. One of key challenges to address is computational efficiency of the implemented method and the definition of a cost function based on the existing "manual" evaluation criteria. Tests in the project will be performed with the use of existing neural models or a low-scale simulation demo will be developed.