The BrainScaleS Project
Results 2011

This text reports about the results from the first project year (2011).

Project Context and Objectives

The BrainScaleS project aims at understanding function and interaction of multiple spatial and temporal scales in brain information processing. The fundamentally new approach of BrainScaleS lies in the in-vivo biological experimentation and computational analysis. Spatial scales range from individual neurons over larger neuron populations to entire functional brain areas. Temporal scales range from milliseconds relevant for event based plasticity mechanisms to hours or days relevant for learning and development. In the project generic theoretical principles will be extracted to enable an artificial synthesis of cortical-like cognitive skills. Both, numerical simulations on petaflop supercomputers and a fundamentally different non-von Neumann hardware architecture will be employed for this purpose. The planning and eventual execution of demonstration experiments are crucial aspects of BrainScaleS.

In-vivo multiscale analysis across species and modalities

As planned, biological in-vivo data have been acquired in the primary visual and somato-sensory cortex during the anaesthetized and behaving states. Acquisition took place for 4 different species (rodent (mouse and rat), cat and monkey). Neural Network state dynamics have been measured at multiple spatial and temporal scales for parameterized sets of low-level sensory tasks, organized as a function of stimulus complexity. Explored integration scales range from intracellular level to large populations of neurons (multiple electrode arrays MEA and population two-photon network imaging) and the mesoscopic averaging of electrical (local field potential LFP, EEG) or evoked depolarizing (voltage sensitive dye fluorescence VSD) cortical activity. Mesoscopic measures have been correlated with perceptual performance, in both humans and non-human primates, based on high-resolution oculomotor and classical psychophysical measurements.

Bridging and understanding multiple spatial and temporal scales

The project has started to combine methods from mathematical modelling, high-performance computing, and analysis to study principles of computation in large neuronal systems across several spatial and temporal scales. One example from the very broad area of activity the project has investigated models of in vivo-like states based on sparsely-connected networks of integrate and fire neurons with conductance-based synapses. The effect of the connectivity on the network state exhibited by the network was the main focus of this work. The main finding was that ``macroscopic properties of the spiking activity (averaged at large scale), such as ensemble correlations and mean firing rates, were similar for different intracortical connectivity profiles ranging from randomly connected networks to networks with Gaussian-distributed local connectivity.

Novel non-von Neumann Computing

The work on a novel type of non-von Neumann computing system was focused on the construction of a first complete wafer-module for the Hybrid Multiscale facility (HMF). The following components had to be manufactured and tested : (1) an analog network ASIC containing the neurons and synapses; (2) a digital network ASIC for the host and wafer-to-wafer communication; (3) an 8 inch wafer in 180 nm CMOS technology; (4) the horizontal interconnections on the wafer providing communication channels for the neurons; (5) the communication modules containing Field-Programmable-Gate-Arrays and (6) the mechanical and electrical components to assemble the system. After successful test and assembly commissioning of the first accelerated, wafer scale integrated neuromorphic system has started in the last quarter of the first BrainScaleS project year. Figure 1 shows the neuromorphic structures on the wafer at different geometrical scales.

Figure: Post-processed wafer containing about 20 million plastic synapses.

During the second project year experiments will be performed on the wafer system according to the original plan.

Novel methods for neuroscience, modeling, computing and technology

The work in this activity has progressed along the following themes:

1. synthesis of stimuli with controllable geometric complexity;
2. progress in different simulation software packages (PyNN, NEST);
3. overview and further plans for database software and management;
4. methods for improving HPC performance supporting large scale simulation efforts;
5. methods for analyzing experimental data at different scales, local field potentials and multi-electrode recordings;
6. methods to enable the development and programming of neuromorphic hardware.

Towards demonstration experiments

The conception, planning and eventual execution of demonstration experiments is at the heart of the BrainScaleS project. The demonstration experiments will be based on the theoretical work, executed on the HPC (high performance computing) and HMF (hybrid multiscale facility) platforms and compared to the multiscale recordings from biology. The project has defined 3 categories of demonstration experiments :

1. open loop experiments,
2. closed loop experiments and
3. generic computing experiments.

The open loop demonstrations will be based on early cortical sensory processing in the visual and somatosensory system. Input will be processed without feedback from the environment. The first year work has mainly concerned to set up and evaluate the different component networks and to study specific properties of those. Important progress has been made e.g. in terms of modeling the development of the primary visual cortex by a self-organizing process driven by visual input. The dynamics of visual perception of moving objects has been studied and modeled and the emerging dynamics in large-scale modular and layered cortical network models have been studied, in some cases using supercomputers.

In preparation of the closed loop experiments monkey data have been analyzed and used to model the neural mechanisms by which a decision can be reported after a delay period. The two derived models, one using synaptic facilitation and the other one graded neural activity, effectively reproduce the experimental data. Also, a probabilistic model for the integration of low-level motion processing and higher level predictive signals in the context of a simple closed-loop behavior has been built. The model, implementing both the recursive loop for visual motion computation and a simple horizontal and vertical oculomotor plant, reproduces many observations. A local multi-layer spiking microcircuit model of primary visual cortex has been extended to a multi-area neural network model comprising all macaque visual cortical areas. The model uses supercomputing facilities and is a first step in the construction of a model with closed functional circuits at microscopic resolution. Finally, a model of the basal ganglia, which is believed to play a critical role in action selection and reinforcement learning has been developed.

For the generic computing demonstrators a foundation for the implementation of novel non-Turing computing methods in the HMF was established during the first year. They enable an implementation of probabilistic inference in Bayesian networks and general graphical models in the HMF. This new method for emulating probabilistic inference in networks of spiking neurons is commonly referred to as “neural sampling”.

A salient property of neural sampling is that it requires stochastic computation, i.e. it employs stochastic, rather than deterministic, neurons as computing units for its operation. Therefore, in analytical studies and computer simulations a method to operate neuron model of the HMF in an appropriate stochastic firing regime that matches the theoretical requirements has been devised.

In another line of work blueprints of neural implementations of probabilistic inference in general Bayesian networks and other graphical models that include higher-order dependencies have been studied. This extension of the neural sampling framework was important in order to expand its applicability to a wide variety of practical probabilistic inference tasks that use Bayesian networks as probabilistic models.

Expected Results and Impact of the Project

The expected overall scientific outcome of the BrainScaleS project is an improved understanding of the role of multiple spatial and temporal scales in large scale neural systems. The improved understanding should lead to the construction of novel non-von Neumann computing platforms that exploit the observed multiscale nature of the systems observed in neurobiology.

BrainScaleS has identified 3 impact domains, in which the project might be able deliver transformative changes beyond the scientific goals described before,

The first impact domain is the microscopic interpretation of macroscopic imaging. BrainScaleS will develop tools to emulate and simulate structured brain-scale networks with as many neurons and synapses as the brain of small mammals. Such models for the first time create truly in-vivo like activity states and can transform the field of computational neuroscience and its applications, since they allow a description of macroscopic phenomena and processing at the level of the entire brain in terms of microscopic models and without dramatic sub-sampling of the system under study. The second impact domain is novel computing. The theory work develops mathematical tools to bridge spatial scales from single neurons to large populations of neurons, to understand learning and brain plasticity, and to bridge spatio-temporal scales to explain the interaction of learning and macroscopic brain-like activity states. It is also expected to bring forward a new method for probabilistic inference in large Bayesian networks via Gibbs sampling in networks of spiking neurons and the possibility of using spiking neurons to solve partial differential equations. Those methods can be applied to a variety of computational tasks even outside the realm of brain science. The final objective is to understand generic principles of computation in parallel systems with local-global interaction. The methods and tools developed in this project represent very promising directions for future ICT research and are ideally suited for implementation on non-von Neumann hardware platforms. If this type of theoretical investigation is successful, it has the potential to found a new area of research and technology with apparent transformational impact.

Finally, the BrainScaleS project will contribute significantly to the standardization of software tools and databases to be used in neuroscientific research. One of the significant innovative experimental principles of the BrainScaleS project is to quantitatively compare observations between biological multiscale recordings, simulations and hardware emulations. Given the need to compare results between different approaches and different partners, the development of shared tools and methods becomes an imperative. It is likely that the establishment of this workflow together with the definition and implementation of the required databases will have high impact to the neuroscience community. The BrainScaleS consortium has the necessary experience and tools to ensure proper dissemination of the new techniques through conferences, publications, interactive webservices and dedicated hands-on workshops.