Project

ACA is organized into four scientific work packages (WPs) according to the central problem areas and a coordination WP managing the project.

This work package explores concepts to meet the connectivity and communication requirements of neuro-inspired algorithms. Specifically, the underlying multi-scale communication infrastructure has to guarantee bandwidth and minimize latency of data transfers of small packets. Despite the potentially enormous scale, start-up time and dynamic re-configuration of the network has to remain a small fraction of the simulation time. Various hierarchical network topologies are considered and benchmarked with detailed cost functions that are based on physical design data.

In biological brains, neurons and synapses are diligent workers continuously reading in input and, depending on their intrinsic state, delivering spikes to the network. This functionality, together with the transfer of spikes across the network, is most likely critical for the superior processing speed of the brain. Flexible neuromorphic hardware implementations of neuronal systems need to accommodate a variety of neuron and synapse models with different dynamics that require dedicated high-performance differential-equation solvers. This work package explores CMOS hardware architectural concepts to assess different implementations allowing an accelerated simulation of neuronal and synaptic dynamics that is faster than biological time. Additionally, it explores possibilities of integrating new technologies such as memristive devices, dynamics of memristive synapses as well as plasticity and learning with novel materials.

This work package develops the top-level architecture of the neuromorphic computer system in a participatory design process involving all other work packages. The requirements specification is iteratively refined in a co-design process with the users. The architecture-level design aims either at a solely standard hardware realization or a hybrid system combining traditional and novel neuromorphic hardware. In recent years, the computational neuroscience community has made advances in the organization of its software infrastructure and tool chain. This work package ensures that the neuromorphic system developed in this project can be well integrated into existing tool chains. It also coordinates teaching, training and user support to make advances known to the community to enable users to rapidly profit from the results and collect requirements for future designs.

Neuromorphic system development is guided by the principles determining the structure and the dynamics of biological brains. This work package establishes the link to experimental and theoretical neuroscience to identify relevant biological constraints, computational principles and optimization schemes, and to define target performances. In addition, it investigates how hardware-specific limitations (e.g., limited resolution or bandwidth) affect the dynamics and functionality of neural network models. The work package further develops tools and workflows for a systematic validation and benchmarking of neuromorphic systems, including a set of science and test cases designed to evaluate to what extent the neuromorphic system can cope with the key challenges specifically targeted in this project: natural-density and plastic connectivity, speed and flexibility.

This overarching work package coordinates the steering committe meetings (twice a year) as well as the technical boad meetings (monthly), organises the annual meetings and the biweekly Neuromorphic Tea sessions, which allow members of the project to meet frequently and informally.