Advanced Processor Technologies Home
APT Advanced Processor Technologies Research Group

SpiNNaker Project - Spiking Neural Applications.


Modelling neural networks with PyNN

Modelling spiking neural networks is a computationally expensive task; dedicated hardware solutions to address the problem have been proposed. However specific hardware can be difficult to use for non hardware experts. In this context we show how to model neural networks using PyNN, a simulator-independent description language for building networks of neurons, on the SpiNNaker system.

The PyNN/SpiNNaker duo exploits the reconfigurability of the ARM cores and of the router constituing the computational and communication heart of the dedicated architecture, letting the user rapidly test networks with different topologies or constituted by different neural models without knowledge of the hardware, and get results back in a standard (NeuroTools) format.

The figure on the right shows an example taken from the PyNN website, comparing simulations on SpiNNaker and on Brian using PyNN
IF_cond_exp.py example

Porting the Neural Engineering Framework on SpiNNaker

The video shows a simulation ran on the SpiNNaker System using the Neural Engineering Framework (Computational Neuroscience Research Group, University of Waterloo, Canada), a mathematical framework which can be used to implement arbitrary dynamical functions into biologically plausible network of spiking neurons. It exploits the configurability of the SpiNNaker system by introducing special encoding and decoding neural models able to represent real value variables into spike trains.

Information is encoded as neural activity on SpiNNaker using the NEF principles; weights are set to compute the square of the encoded input. Information is then decoded and returned to the user interface. More information available at the Telluride workshop website

Robotics interaction

Realized in collaboration with the Cognition for Technical Systems group of the Technische Universitat Munchen. The holonomic robot (TUM) is controlled by a neural network on the SpiNNaker System. The task is to follow a line.

Visual input is fed in the system through a silicon retina (Institute for Neuroinformatics, Zurich) to a retinotopic map of neurons modelled on a single SpiNNaker Chip, while motor control is obtained by the firing rate of simulated motor neurons.

This project is the following up of http://www.youtube.com/watch?v=ZQ7FdQ_VJNg