Large scale spatio-temporal point processes: novel machine learning methodologies and application to neural multi-electrode arrays.

点击翻译按钮获取中文摘要

Large scale spatio-temporal data sets are becoming increasingly available due to progress in data gathering technology. In this proposal we are concerned with event-based data: data points are spatial and temporal coordinates of an event, as opposed to analogue measurements of a variable. Such data is pervasive in a number of applications, ranging from epidemiology to social sciences, and poses considerable computational issues: the data is intrinsically high dimensional (indeed infinite dimensional if working in a continuous time framework) and nonlinear, and it is often a noisy observation of complex dynamical processes. Scalable data modelling solutions for this data type are urgently needed, and will require novel research in computational statistics and machine learning. In this proposal, we will address fundamental methodological problems motivated by an application of great relevance in a biomedical scenario: recordings of neural electrical activity by High Density Multi Electrode Arrays (HD-MEA). These are electronic chips with many (>1000) recording channels, which are used to measure electrical activity in a range of in vitro preparations, and enable simultaneous measurement of the spiking activity of thousands of neurons. This novel technology (commercially developed within the last five years) has the potential to enable scientists to answer fundamental questions on how neurons communicate between each other, as well as having direct translational potential as an effective tool to test in vitro the impact of drug treatment over neuronal function. Providing data modelling tools for such data is challenging: HD-MEA recordings are a prime example of big data (data rates of ~3GB/minute) where complex behaviours rule out simple scalable models.In this exciting multi-disciplinary project, we propose to treat an HD-MEA data set as a realisation of a spatio-temporal point-process (a random set of points, i.e. neural spikes), and use and develop techniques from Bayesian statistics and machine learning to infer salient dynamical properties of the biophysical process underlying the data. The major challenges which will be addressed are concerned with devising statistical machine learning methods which can accommodate non-linearities and that can scale to the large size of HD-MEA data, while still giving biologically meaningful insights. In particular, we will focus on determining from data the connectivity of the network of neurons, neuron-intrinsic dynamics, and how chemical (e.g. drugs administered to the culture) and other stimuli influence the electrical response and network properties of the culture. Addressing these challenges will entail considerable work on approximate Bayesian inference for large-scale spatio-temporal point processes, generating methodologies which will be general and applicable to many other domains of science and engineering.The project brings together the machine learning and systems biology expertise of the PI and named RA, the neuroscience expertise of the coI as well as strong collaborative ties with international experimental groups and industrial players, making the team of researchers ideally suited to tackle this challenging project.