Hands-on working group “BrainScaleS-2”
As the next generation BrainScaleS-2 system is now available in prototype form we want to offer a hands-on working group to explore the new features of the next generation chip.
Programmable Plasticity
Besides its analog neuromorphic core the system features a digital plasticity processing unit (PPU) programmable in C and C++. It has read and write access to the synaptic weights and correlation sensors of the synapse array as well as all other digital and analog parameters of the chip. This allows one to implement plasticity rules well known in biology such as STDP and reward-based STDP. By extending the update scheme to modify a network’s connectivity, one can implement structural plasticity rules. The full programmability also would allow to explore implementations of backpropagation, feedback alignment and equilibrium propagation.
Structured Neurons
The next generation BrainScaleS system will feature a highly configurable substrate for emulating a subset of the rich spatial structures found in biology. It features 32 neuron compartments that can be connected into dendritic tree configurations and can additionally emulate NMDA plateau potentials and Calcium spikes. Together with the possibility of implementing complex plasticity algorithms on-chip, this might enable the exploration of biologically inspired backpropagation rules that use multi compartment neurons.
Accelerated Robotics
The high speed-up of the BrainScaleS-2 system (1000x) requires an equally accelerated environment to become part of full closed-loop interaction scenarios. We recently constructed a prototypic versatile accelerated robotic platform that is connected to the BrainScaleS-2 hardware by a patchable analog interface board. In the course of this workshop, we will explore the whole range of possible experiments and applications of the setup.
We would like to invite you to participate in designing suitable experiments.
Timetable
Day | Time | Location |
---|---|---|
Wed, 25.04.2018 | 19:00 - 20:00 | Disco/Lab |