Welcome to The Neuromorphic Engineer
Books

BOOK REVIEW
Analog VLSI Circuits for the Perception of Visual Motion

PDF version | Permalink

Ralph Etienne-Cummings

1 March 2007

This volume is highly recommended for anyone interested in neuromorphic computation generally, or analog VLSI visual motion circuits more specifically.

Alan A. Stocker, Wiley, March 2006ISBN: 978-0-470-85491-4Hardcover: 242 pagesUS $130.00 / £70.00/ 荤105.00

Ever wondered why progress seems slow in building visually guided autonomous agents that perceive and intelligently interact with their environment? Well, one reason may be that our understanding of perception and the underlining computations involved is incomplete or just plain wrong. This new book by Alan Stocker provides n unconventional and fresh perspectives on how to understand perception and build simple artificial perceptual systems using analog VLSI (very large silicon integration) circuits. Focusing on the example of visual motion perception, it demonstrates how brain-style computation combined with CMOS (complimentary metal-oxide semiconductor technology can lead to efficient and robust ‘neuromorphic’ circuits to solve the hard optimization problems encountered in perception.

One key factor underlying the success of human visual perception lies in its use of constraint satisfaction. That is, the brain presumably applies mechanisms that combine the aspects of its visual input that cohere and segments out those aspects that do not. These mechanisms bootstrap globally coherent (optimal) solutions by rapidly satisfying local consistency constraints. Consistency depends on relative computations such as non-linear comparison, interpolation and error feedback, rather than absolute precision. And this style of computation is very suitable for implementation in analog VLSI circuits, as Dr. Stocker demonstrates.

What makes this book special is that it not only presents practical implementations of constraint-satisfaction networks for visual motion perception, but it also demonstrates a series of useful and impressive aVLSI circuits for solving visual motion problems such as estimating 2D optical flow, motion segmentation, and motion selection. And these chips are useful for robotic applications. Their true strength lies, however, in their broad and principled theoretical foundations.

The book begins with some ecological considerations about why and how visual motion is perceived from changes in the visual input. It then goes on to illustrate the basic computational challenges, discusses possible solutions, and finally concludes in proposing a general computational architecture for visual motion perception. The key concept is that the perceptual process is an optimization problem of finding the visual motion estimate that is maximally consistent with the visual information and the system's expectations. Chapter three makes the connection to associative memory and Hopfield networks as examples of network architectures that compute optimal solutions. It demonstrates how simple problems (e.g. the winner-take-all operation) can be formulated as local constraints that together define the optimal solution. The chapter also shows how to derive appropriate network architectures that find it.

Chapter four then formulates optical-flow estimation as a constraint satisfaction problem, deriving the basic network architecture that is the basis for all further networks discussed in the book. It draws the connection between the formulated constraint solving problem and statistically optimal motion estimation as described with Bayesian frameworks, showing that prior information is essential in achieving a robust design. Furthermore, extensions of the basic network allow even more sophisticated processing such as motion segmentation or motion selection for which the network selects regions in its visual field that match a particular motion and size.

Chapters five to seven extensively deal with aVLSI implementations of the proposed network architectures, providing detailed schematics and measurements of the fabricated chips. The effects of the inevitable non-linearities and mismatch are discussed in detail, showing that clever analog designs can take advantage of nonlinearities to improve robustness and performance.

The book concludes with an interesting final chapter with a comparison to primate visual motion perception systems. It also presents data of head-to-head comparison between humans and the aVLSI chips performing the same perceptual tasks. The dynamics and steady-state behavior similarities are quite surprising, leading the author to conclude that both systems must optimize a similar set of constraints. The future will tell if this is true or not.

The broad approach of this book certainly reflects the background and the interests of the author. He is an expert aVLSI circuit designer, a computational modeler of the visual system, and a psychophysical experimentalist working on human motion perception. I highly recommend this book not only to those who are particularly interested in aVLSI visual motion circuits, but to anyone interested in the novel, neuromorphic, style of computation. The philosophy and methodology of the approach seem general enough and applicable to other perceptual tasks, such as depth perception and texture segmentation. Furthermore, the analog VLSI implementation of the presented computational networks becomes particularly attractive in light of recent technological developments in three-dimensional integrated circuits. Three-dimensional integration permits local vertical connections between different chips, physically stacked as a ‘layer cake’. Recurrent analog networks can naturally be implemented as multi-layered parallel computational blocks of tremendous capabilities, without the need for sophisticated chip-to-chip protocols.




Author

Ralph Etienne-Cummings

 
DOI:  10.2417/1200703.0049

@NeuroEng



Tell us what to cover!

If you'd like to write an article or know of someone else who is doing relevant and interesting stuff, let us know. E-mail the editor and suggest the subject for the article and, if you're suggesting someone else's work, tell us their name, affiliation, and e-mail.