Probabilistic neural computation and neural simulation technology

Deciphering the working principles of brain function is of major importance from at least two perspectives. From the clinical viewpoint, a deeper understanding of our brains will lead to better treatments for psychological and neurodegenerative diseases. The technological perspective promises smart machines that rival our ability to perceive, learn and act in the real world. It is generally believed that the relevant physical processes can be understood in terms of large, plastic networks of nerve cells. Over the last decade, probability theory has gained popularity as a normative model of brain function, since it offers a unifying view for many behavioural phenomena. To connect this high-level description to low-level implementations in neural substrates, computational models provide effective means to investigate specific hypotheses. Many neural network models for probabilistic inference employ approximation algorithms, relying on stochasticity to achieve a sample-based representation of probability distributions. In the first part of this thesis, we investigate biophysically plausible sources of stochasticity for models of approximate inference in neural substrates. We consider network models derived from a popular model class for sample-based inference, Boltzmann machines, but equipped with a natural source of stochasticity: synaptic input from other neurons. We demonstrate that stochasticity generated by recurrent neural networks can outperform other approaches based on independent random number streams if resources are limited. We further show that the essential effect required for this approach is also present in networks emulated in a mixed-signal neuromorphic system with strong temporal and spatial heterogeneities. Subsequently we consider network models with alternative sources of stochasticity, including variability in the coupling strength between neurons. Our results show that while these networks are able to represent well-defined probability distributions, probabilistic inference cannot be implemented using straightforward methods familiar from Boltzmann machines. The human brain contains approximately one hundred billion individual nerve cells connected by about one hundred trillion synapses: This complexity poses a significant computational challenge for the numerical simulation of such systems. To support simulations of large-scale network models at cellular resolution, researchers can rely on well-tested high-performance simulation software that makes efficient use of the largest supercomputers available today with tens of thousands of compute nodes. However, it has an inherent scalability bottleneck arising from the memory consumption of the connection infrastructure and the undirected communication algorithms typically employed. In the second part of this thesis, we introduce a new two-tier connection infrastructure employing directed communication across compute nodes. By implementing this new technology in NEST, a widely-used simulator for large-scale neural network models, we demonstrate that it solves previous scalability bottlenecks in terms of memory consumption. In addition, we show that on current supercomputers, the directed communication between processes significantly improves the simulation performance, and at the same time maintains high efficiency in small-scale simulations executed on laptops and small clusters. Simulation software is however only one part required for neuroscientific modeling: for functional networks, providing a realistic environment that interacts with the models is equally important. Researchers should not have to rely on handcrafted solutions as this makes models difficult to evaluate and compare objectively. To tackle these issues, we develop a toolchain that enables closed-loop interactions between neural-network models in established simulators and reinforcement-learning environments from machine-learning toolkits. In conclusion, this thesis brings models for approximate inference in neural circuits further into the biophysically plausible domain, with direct applications for neuromorphic implementations, and presents new simulation technology preparing software tools for the simulation of human cortex at the resolution of individual neurons and synapses on future HPC systems