Computational principles

We develop computational principles and learning schemes that support the use of currently available large neuromorphic systems, and guide the design of the final version of neuromorphic systems. The development of these computational principles and learning schemes are based on insight from other subprojects, from new experimental data in Neuroscience, Cognitive Science, and Machine Learning,as well as from theory and exploratory software simulations.

Fast large scale simulations on neuromorphic hardware will enable us to test the viability of different hypotheses about the organization of computation and learning in the brain, and will thereby support other subprojects of the HBP and the use of neuromorphic hardware by the scientific community at large. In addition, we aim at the discovery of new principles for non-van-Neumann computation in new physical realizations and technologies.

Main objectives of our work

First, development of principles of stochastic computation and probabilistic inference that support computational uses of operational HBP neuromorphic systems, and which can guide the design of updated systems. A special focus will be on principles that enable larger neuromorphic systems to carry out a variety of demanding computational and cognitive functions.

Second, the development of learning schemes and learning architectures that enable large neuromorphic systems to learn a variety of computational and cognitive functions. We want to enable neuromorphic systems to learn not only with the help of a teacher, but also from its own observations, play, and transfer of learnt knowledge in order to learn new tasks very fast. Hence we will focus on methods that enable large and complex neuromorphic systems to create --with little or no supervision-- hierarchical internal models for streams of data to which they are exposed, and to use these internal representations --guided by rewards-- to carry out decisions, to plan, and to initiate purposeful actions. In particular methods for long-term learning of neuromorphic systems will be developed and tested on the neuromorphic hardware developed, where years of learning in complex brain networks can be simulated within hours.

Third, development of blueprints that large neuromorphic systems to act as artificial brains, encompassing higher cognitive functions such as abstract reasoning, planning on several levels of abstraction, and solving problems on the basis of suitably structured learnt knowledge.

Fourth, creation of highly innovative new non-von-Neumann computing paradigms that have a potential to shape future computing in a variety of new physical realizations and materials. This research will be supported by open brainstorming workshops that will bring together researchers from many disciplines.