Neuromorphics: Compiling code by configuring connections

With a power budget of 12 watts, the brain perceives, decides and acts. KAIST's battery-powered HUBO 2 robot uses over a quarter of its energy just to control walking even in predictable environments. BrainGate's implantable brain machine interface (BMI) uses half of its energy to wirelessly transmit recorded neural signals because it cannot decode the intended movement. How can engineers build processors as energy-efficient as the brain? Our approach is to mimic its graded dentritic potentials using analog circuits and its all-or-none axonal spikes using digital circuits. To compute with our energy-efficient silicon neurons, we use use the Neural Engineering Framework (NEF).

Content on this page requires a newer version of Adobe Flash Player.

Get Adobe Flash player

Neuromorphic robotic controller
The GUI on the left side of the computer monitor shows the spiking activity of neurons on Neurogrid as they control the robot. Robot joint angles are transmitted to the computer and used to render the model in the GUI on the right side of the computer monitor. A user commands the position of the robot's end effector. The input is translated into spikes and transmitted to the silicon neuron neural network. The network computes the energy optimal joint torques for the robot to use to reach the commanded position. By controlling torques, the robot is compliant to external perturbations as demonstrated by bringing a board into contact with the pen attached to the robot.

NEF enables us to compile perceptual, cognitive and motor algorithms onto networks of spiking neurons.

Developed by Eliasmith and Anderson, NEF defines three principles. First, a population of neurons collectively represents a time-varying vector through nonlinear encoding and linear decoding. Second, alternative linear decodings that transform the vector (linearly or nonlinearly) are used to compute weighted connections from one neural population to the next. Third, recurrent connections—from a neural population back to itself—realize a transformation that governs the vector's dynamics.

Our ultimate goal is to build an autonomous robot that perceives, decides and acts using large-scale networks of silicon neurons.

Our robot will perceive the world through a silicon retina, decide what to do using a silicon cortex, and act on the world with a mechanical arm. Currently, we are using NEF to configure silicon neurons to decode intended movements energy-efficiently for a next-generation implantable BMI and to control a robotic arm energy effiently for a next-generation humanoid robot. We are also architecting an NEF chip capable of implementing networks large enough to realize a fully cognitive system.

Alexander Neckar is architecting an NEF chip.
Sam Fok is prototyping robot controllers and BMIs on Neurogrid.
Samir Menon is building robots and applying operational space control.
Paul Nuyujukian (Shenoy lab) is testing spiking-neural-network BMIs.

Chris Eliasmith
Oussama Khatib
Rajit Manohar
Krishna Shenoy

National Institutes of Health
Office of Naval Research


ID Article Full Text
S Choudhary, S Sloan, S Fok, A Necker, E Trautmann, P Gao, T Stewart, C Eliasmith, and K Boahen, Silicon Neurons that Compute, International Conference on Artificial Neural Networks, LNCS vol VV, pp XX-YY, Springer, Heidelberg, 2012. In Press

Full Text
C39 J Dethier, P Nuyujukian, C Eliasmith, T Stewart, S A Elassaad, K V Shenoy, and K Boahen, A Brain-Machine Interface Operating with a Real-Time Spiking Neural Network Control Algorithm, Advances in Neural Information Processing Systems 24, Curran Associates, Inc., pp 2213-21, 2011.
Full Text
C38 J Dethier, V Gilja, P Nuyujukian, S A Elassaad, K V Shenoy, and K Boahen, Spiking Neural Network Decoder for Brain-Machine Interfaces, IEEE EMBS Conference on Neural Engineering, IEEE Press, pp 369-399, 2011.
Full Text
J38 J Dethier, P Nuyujukian, S I Ryu, K V Shenoy, and K A Boahen, Design and validation of a real-time spiking-neural-network decoder for brain machine interfaces, Journal of Neural Engineering. Vol 10, p. 036008, Apr 2012
Full Text