Research With Robots
In short, they build it and see what it takes to make it work. The challenges involved in this process all illuminate something about the model in question... In the Computational Sensorimotor Systems Lab, Dr. Timothy “Timmer” Horiuchi and his students design hardware models of neural systems. Their media of choice are the same silicon chips found in cell-phones and computers, often tethered by a nest of wires to a remote-control car, plane, or other robotic platform. The philosophy behind their work is that getting model systems to function in the real world will reveal issues which may be overlooked in theoretical discussions or simplified simulation environments. In short, they build it and see what it takes to make it work. The challenges involved in this process all reveal something about the model in question, indicating where it works well, where it doesn't, and where changes or revisions may be in order.
Project Microchipoptera: Robot Bats
The largest of the lab's research projects is the development of a flying robot capable of navigating autonomously via echolocation, a robot bat in common parlance. This lends the project it's name, Microchipoptera (taken from Microchiroptera, the suborder of echolocating bats). A recent portion of the Microchipoptera project has dealt with the development of a neuromorphic VLSI chip capable of performing echolocation in two dimensions. The chip localizes targets by their range and their angle to the right or left of center (azimuth). Neurophysiologists researching bats and other mammals have hypothesized models of how localization in each of these dimensions is performed, based on neurophysiological data. However, it is unclear how the range and azimuth cues from a single target are combined into a single representation. The 2D-chip seeks to provide insight into this problem by combining range and azimuth data via an ostensibly simple spike-timing mechanism.
From Cells to Silicon
The chip is populated with silicon neurons, which are electronic circuits designed to have the basic components and functions of biological neurons. Across the top of the chip is an array of neurons which receive inputs from the left and right 'ears' of the robot. These circuits are modeled after neurons in the brainstem of the bat, which recieve excitatory input from one ear and inhibitory input from the other, making them responsive to interaural level difference (ILD): the difference in loudness between the ears. This cue can be translated to target azimuth. Along one side of the the chip is another array of silicon neurons. These neurons respond selectively to the delay between the robot-bat's outgoing vocalization and echoes returning from targets: the cue which indicates target range. They function by implementing a dynamic membrane property known as post-inhibitory-rebound to open a window of sensitivity to returning echoes with certain time delays. In the center of the chip is a grid of neurons which respond to spikes in the delay-tuned cells and the ILD-tuned cells which occur nearly simultaneously. These 2D-tuned cells respond to the combination of an edge of activity in the ILD-tuned array of cells and simple spikes in the delay-tuned array. As long as the computation times required for the ILD and delay cues are consistent, a mechanism such as this should be a valid way to combine these cues.
.
Making it work
As designed, the 2D-chip is indeed capable of combining ILD and delay cues into a single representation of 2D space. The design and setup decisions that went into the creation of the chip have yielded some interesting insights into how such a system would work in bats. For one, it has reiterated the importance of spike-timing in the bat's echolocation system, even for computations such as edge-detection, which are considered relatively simple in other systems. This is due to the highly transient nature of the stimuli involved. The vocalizations produced by bats last a few milliseconds at most, and the typical neuron processing such sounds produce only one or two spikes in response. Thus, features which affect the timing of computations, such as synaptic and membrane time constants play a major role throughout the entire system. Development of the 2D-chip also revealed that it is possible for the delay-tuned neurons to function in a regime where their spike latencies are fixed relative to the bat's vocalization and not the returning echoes, a rather unusual property.
From Silicon Back to Cells
Hopefully the information gleaned from the 2D-chip will help to guide the future work of neurophysiologists, inspiring further experiments and allowing for new interpretations of past results
Hopefully the information gleaned from the 2D-chip will help to guide the future work of neurophysiologists, inspiring further experiments and allowing for new interpretations of past results. Further, the 2D-chip and it's successors will function as platforms for the study of higher level echolocation-based computations in the bat, helping to answer questions such as, “What are the challenges of navigating using echolocation?” and “How are echolocation targets represented in the silent period between vocalizations?”
