Title: "Machine Learning and Neuroscience*
Speakers: Concha Bielza and Pedro Larrañaga (Universidad Politécnica de Madrid)
Abstract
Reverse-engineer the human brain has been identified by the National Academy of Engineering in 2009 as one of the 14 challenges that will influence science and technology for the next decade. This calls engineers and neuroscientists to work together. As participants of the Blue Brain international project, in this talk we will describe a number of interesting problems where Machine Learning and Neuroscience can collaborate. They are briefly explained below.
*A.* Define distinct populations of neurons in the brain cortex according to electrophysiological, molecular and morphological features.
This involves a supervised classification to distinguish between the common pyramidal neurons and interneurons or an unsupervised classification that discovers new groups of neurons inside these two classes, whose identity remains unclear.
*B.* Construct models that can simulate the huge variety of dendritic morphology, the shape of brain cells. Dendritic branching is poorly understood due to be driven by a complex process but understanding how these trees are formed in the brain, we can learn more about their normal function and why they are often malformed in neurological diseases. Virtual trees can be created by measuring from real cells different parameters controlling branching behaviour. These measures can be modelled as statistical distributions and a global simulation model accounts for all the parameters together.
*C.* Accurately reconstruct (semi-)automatically the 3D structure and connectivity of neurological tissue and label dendrites, axons and synapses unambiguously from electron microscope images that can provide a level of detail never reached before. It requires a huge amount of data to be processed (hundreds of terabytes) where machine learning and image processing techniques have to be combined.
*D.* The technique of functional magnetic resonance imaging (fMRI) provides high-contrast images of neuronal activity indicating variations in oxygen consumption. The source data are 3D pixels (voxels) representing different intensities. With brain images, we can identify regions of the ventral-occipital cortex involved in visual processing of objects or faces and predict neural activation while identifying pictures and words. A visionary approach would be to use fMRI to diagnose neurodegenerative diseases like Alzheimer's. Difficulties arise due to the joint use of clinical data and spatial, temporal and function information.
*E.* In the same spirit, Alzheimer's disease data and results from different -omics (genomics, proteomics, metabolomics) may be integrated into a single research to study the possible influences and interactions among them. These include DNA microarrays, SNP chips, mass spectrometry runs... that when combined, can open a major degree of knowledge discovery for the Alzheimer's disease.
*F.* Identify and solve some classes of optimization problems in neuronal modelling. The difficulty of the problems may lead to use evolutionary computation based on machine learning models. We identify three classes: (1) neuronal parameter optimization, where we search for the parameter values that yield a desired electrical activity pattern in a neuron (or neuronal network); (2) neuronal model selection, where besides the parameters we also optimize some of the structural components of the model (useful when an accurate formulation of the neuronal model is unavailable); and (3) structural optimization problems in brain networks, where a network topology satisfying a number of constraints has to be found. Brain activity can be modelled as a dynamic process acting on a network, with brain areas, groups of neurons or individual cells as vertices. Its topology may be a key element in understanding the behaviour of the process.