Dr. Shih-Chii Liu

 

Main Goals, Keywords

To explore computational principles of nervous systems through implementations of sensory and cortical processing models; for example, spike-based cochlea VLSI chips and hybrid VLSI networks of integrate-and-fire neurons and synapses; to develop technology for long-term adaptation and learning in analog VLSI, and to develop sensori-motor control for robotics using neuromorphic analog VLSI preprocessors (neuromorphic sensors, spike-based processing, hybrid VLSI networks of integrate-and-fire neurons and synapses, address-event representation, neural learning mechanisms, adaptation, visuo-motor models of behavior).

Group Members

4 PhD students

Previous and Current Research

I am interested in understanding the computational principles in nervous systems that allow robust computation even in the presence of noise and element differences; and using these principles in the development of hardware VLSI models of visual, auditory, and cortical processing. Explored cortical models consist usually of multi-chip systems that receive input from front-end sensors like the retinae and the cochlea and whose outputs are processed by spike-based neuronal networks. These neuronal networks are reconfigurable and rewireable using the AER (asynchronous address-event representation) protocol and consist of arrays of neurons with a variety of dynamical synapses. The components in the system communicate their events (spikes) through the AER protocol. I am also working on visuo-motor models of insect behavior that can be tested by the use of various types of monolithic neuromorphic sensors on a robotic platform.

 

This figure shows the chip layout for some simple cortical circuits. There are two one-dimensional array of 64 neurons: one layer simulates L4 and the second layer simulates L6 of the visual cortex. The green neuron is representative of one of the excitatory neurons in L4. There is one global inhibitory neuron (in yellow) on this layer. This neuron inhibits all the excitatory neurons on the same layer. The neurons can be driven externally through the red synapses by using the address-event representation protocol. The synapses consist of an excitatory synapse, an inhibitory synapse, and a short-term depressing excitatory synapse. The inputs to the synapses can come from the output of a simulated model of the LGN or from a real cell. The blue neuron is representative of the excitatory neurons in the second layer. There is again one global inhibitory neuron which inhibits all the remaining excitatory neurons in this layer. The neurons in the first and second layers are mutually connected through excitatory or inhibitory synapses. The membrane potential of the neurons can be monitored through a scanning circuit and the output spikes of the neurons can be monitored through the address-event representation protocol.

Future Projects

We are developing an address-event representation infrastructure for an asynchronous event-based multi-chip vision system that can be mounted on a robot. The components are developed by different groups in an EU-funded project (CAVIAR) and the components communicate using their spikes using the address-event representation protocol while all processing within the chips is analog. We are also investigating neural learning mechanisms involved in the sensory fusion area (midbrain) of owls and using these ideas in a sensory-fusion platform on a robot. This platform will hold event-based (or spike-based) AER retinas and cochleas, multi-neuron chips all operating within the AER infrastructure. These ideas will be tested on a robot in a pilotage task.

Techniques and Equipment

Techniques involve the use of software simulation tools (C and Matlab) for simulating neural models of computation including vision, auditory, and cortical processing; and circuit VLSI design simulation tools for designing the aVLSI analog/digital circuits for neuromophic systems. Equipment include the use of test equipment (like scopes) for testing the chips; and robots for visuo-motor control models.

Selected Publications

• Indiveri, G. and Linares-Barranco, B. and Hamilton, T.J. and Van Schaik, A. and Etienne-Cummings, R and Delbruck, T. and Liu, S.C. and Dudek, P. and Häfliger, P. and Renaud, S. and Schemmel, S. and Cauwenberghs, G. and Arthur, J. and Hynna, K. and Folowosele, F, and Saighi, S. and Serrano-Gottarredona, T. and Wijekoon, J. and Wang, Y., and Boahen, K. Neuromorphic silicon neuron circuits, Frontiers in Neuroscience, 5:(73), 2011 
• Finger, H. and Liu, S.C. Estimating the location of a sound source with a spike-timing localization algorithm. Circuits and Systems (ISCAS), 2011 IEEE International Symposium on 2461-2464, 2011
  
• Wang, Y. and Liu, S.C. A Two-Dimensional Configurable Active Silicon Dendritic Neuron Array. Circuits and Systems I: Regular Papers, IEEE Transactions on, 58:(9) 2159-2171, 2011

Selected Lectures, Seminars or Colloquia

• Computation in Neuromorphic analog VLSI Systems
• Design in Neuromorphic analog VLSI Systems
• Kolloquium in Neuroinformatik

Funding

Swiss National Science Foundation, EU Emerging Technologies Program, US Office of Naval Research, Zurich Neuroscience Center, ETH Zurich Research Funding

URL

www.ini.ethz.ch/~shih/