Hybrid Representation and Visualization Methods Applied to Neuronal Microcircuits

For the past twenty years, scientific visualization has assisted simulation-based research to explore problems of organized complexity. Environmental research, social sciences, astronomy, and modern physics have exploited the increasing power of computer graphics to interactively visualize three-dimensional models. Of late, neuroscience has witnessed the use of large-scale simulation with biologically detailed neuron models, and aspires to benefit from scientific visualization too. This thesis is the result of a close collaboration with the Blue Brain Project, a neuroscience endeavor working on biologically detailed large-scale simulation of the brain, and tries to answer the following question: How can scientific visualization improve neuroscience insights in the context of simulation-based research? Scientific visualization of a model typically preserves its spatial layout. In continuity with this axiom, we identified in more depth the requirements for a scientific visualization solution in neuroscience. Thus, the preservation of the domain specificity, the hybrid visualization, the validation, and the adaptation to the limitations of the human-visual perception enhance the readability of the generated results. We applied these requirements to elaborate an hybrid representation of neuronal structures, and proposed a unified framework of visualization methods which preserves the domain specificities of neuroscience. This framework was finally assessed in case studies, where neuroscientific investigations about structure, composition, and dynamics of neuronal microcircuit challenge the developed visualization solutions to derive insight. We believe that these representations and visualization methods form a solid ground for the elaboration of virtual instruments dedicated to a new approach of neuroscience, where in silico landscapes of the brain are complementary to the classical in vitro and in vivo investigations.

Predictive Engineering the Membrane Composition of Neocortical Neurons

Georges Khazen

The brain is the most mysterious and intricate biological structure known to man. It dominates the way we live, think, reason and behave. It is built of billions of neurons that communicate with each other through trillions of connections using electrical signals. The neocortex is undoubtedly the most complex brain region, hosting all higher brain functions such as sensory perception and cognitive behavior. Scientists have been struggling for more than a century to identify the anatomical blueprint of its cellular organization and to understand its dynamic behavior. The Blue Brain Project aims to develop the first simulation-based research environment for modeling and studying the basic building blocks of the neocortex, the neocortical column. The model column is based on data obtained from the juvenile rat somatosensory cortex. The best way to build such a large-scale model is by following a data-constraint driven approach and integrating biological information at different levels. In this thesis, I present novel informatics approaches to systematically constrain the membrane protein composition of different neocortical neuron types. In the first study, I investigate the single-cell gene expression patterns of ion channels in different neuron types and present a combinatorial rule extractor model that reverse engineers the observed expression rules in ten different neocortical neuron types. In the second study, I present a physical constrain checker that mines different publicly available data resources regarding all known membrane proteins such as channels and receptors, and extracts available information relating to their specific location on different parts of the neuronal morphologies, their 3D morphological structures as well as their genetic and proteomic annotations. Combining the results from both studies allows us to build neuronal models that better reflect the biological reality of neocortical neurons, ensuring the correct localization and distribution of membrane proteins across neuronal morphologies and computing their maximum count number based on their molecular surface areas and the available neuronal membrane surface area. This novel data-constraint driven approach is the first step towards integrating existing biological knowledge and constraints from various data sources into the process of building neuron models, in this way improving the neuron models as they come closer to the biological reality of real neurons.

EPFL2011

Neuroscience thinks big (and collaboratively)

Henry Markram

Despite cash-strapped times for research, several ambitious collaborative neuroscience projects have attracted large amounts of funding and media attention. In Europe, the Human Brain Project aims to develop a large-scale computer simulation of the brain, whereas in the United States, the Brain Activity Map is working towards establishing a functional connectome of the entire brain, and the Allen Institute for Brain Science has embarked upon a 10-year project to understand the mouse visual cortex (the MindScope project). US President Barack Obama's announcement of the BRAIN Initiative (Brain Research through Advancing Innovative Neurotechnologies Initiative) in April 2013 highlights the political commitment to neuroscience and is expected to further foster interdisciplinary collaborations, accelerate the development of new technologies and thus fuel much needed medical advances. In this Viewpoint article, five prominent neuroscientists explain the aims of the projects and how they are addressing some of the questions (and criticisms) that have arisen.

Nature Publishing Group2013

Biologically plausible unsupervised learning in shallow and deep neural networks

Bernd Albert Illing

The way our brain learns to disentangle complex signals into unambiguous concepts is fascinating but remains largely unknown. There is evidence, however, that hierarchical neural representations play a key role in the cortex. This thesis investigates biologically plausible models of unsupervised learning of hierarchical representations as found in the brain and modern computer vision models. We use computational modeling to address three main questions at the intersection of artificial intelligence (AI) and computational neuroscience.The first question is: What are useful neural representations and when are deep hierarchical representations needed? We approach this point with a systematic study of biologically plausible unsupervised feature learning in a shallow 2-layer networks on digit (MNIST) and object (CIFAR10) classification. Surprisingly, random features support high performance, especially for large hidden layers. When combined with localized receptive fields, random feature networks approach the performance of supervised backpropagation on MNIST, but not on CIFAR10. We suggest that future models of biologically plausible learning should outperform such random feature benchmarks on MNIST, or that such models should be evaluated in different ways.The second question is: How can hierarchical representations be learned with mechanisms supported by neuroscientific evidence? We cover this question by proposing a unifying Hebbian model, inspired by common models of V1 simple and complex cells based on unsupervised sparse coding and temporal invariance learning. In shallow 2-layer networks, our model reproduces learning of simple and complex cell receptive fields, as found in V1. In deeper networks, we stack multiple layers of Hebbian learning but find that it does not yield hierarchical representations of increasing usefulness. From this, we hypothesise that standard Hebbian rules are too constrained to build increasingly useful representations, as observed in higher areas of the visual cortex or deep artificial neural networks.The third question is: Can AI inspire learning models that build deep representations and are still biologically plausible? We address this question by proposing a learning rule that takes inspiration from neuroscience and recent advances in self-supervised deep learning. The proposed rule is Hebbian, i.e. only depends on pre- and post-synaptic neuronal activity, but includes additional local factors, namely predictive dendritic input and widely broadcasted modulation factors. Algorithmically, this rule applies self-supervised contrastive predictive learning to a causal, biological setting using saccades. We find that networks trained with this generalised Hebbian rule build deep hierarchical representations of images, speech and video.We see our modeling as a potential starting point for both, new hypotheses, that can be tested experimentally, and novel AI models that could benefit from added biological realism.