Neuromorphic Networks


Luca Mazzucato

An interview with Konstantin Likharev. He joined the Stony Brook Physics Department in 1991, after a long and successful career at Moscow State University. He is a pioneer in nano-electronics and the developer of CrossNets, a new kind of chip that could change the way we think about computers.

In recent years, you have been developing a revolutionary hardware. Can you tell us about it in simple terms?

We are developing hybrid integrating circuits which combine a subsystem of highly functional semiconductor transistors with a layer of less functional but much more dense nano-devices. A brief two-page description of our concept may be found here.

What are its possible applications?

First of all, estimates show that digital versions of our hybrids may overcome some integrated circuits (most importantly, nonvolatile memories and field-programmable gate arrays) in density by about 2 orders of magnitude, at similar design rules and dissipated power. This potential advantage may be sufficient to lure semiconductor industry into high-volume manufacturing of these circuits.

Technology to build CPUs has reached a peak in clock speed. Is your system capable of replacing current CPU architecture and speed up performances?

For most applications, the calculated speed of our digital hybrids is not much higher than that of the usual semiconductor-transistor (“CMOS”) circuits; the hybrids’ major advantage for digital applications is their higher density.

Are you going to rescue moore’s law?

As any exponential growth, the Moore’s Law cannot be continued forever; however, our hopes are to extend it for 10 to 15 years.

Can you build artificial neural networks out of your new circuitry?

Mixed-signal versions of such circuits may be used as neuromorphic networks (“CrossNets”) which may perform virtually all cognitive functions which had been demonstrated using software-implemented artificial neural networks (ANN), at much higher speed. Moreover, these networks are apparently the first hardware which may overcome the mammal cerebral cortex in areal density, at similar connectivity, much higher speed, and manageable power (though current functionality of CrossNets as of today is still incomparably lower than that of their biological prototypes).

How does your system compare to the current state of the art of artificial neural networks? Can you do better?

For a few examples we have studied, a single-chip CrossNet would outperform a modern microprocessor (of a similar area and power) by approximately 6 orders of magnitude: a million times faster! The advantage results mostly from replacing digital calculations with analog-signal functions whose lower accuracy is quite acceptable for many machine learning algorithms with their high redundancy.

Your research is financed by the department of defense. Does the grant come with no strings attached?

No it does not: we have got what military call the “6.1 money”, slated for support of fundamental research…

How long is it going to take realistically to have a working demo? And to produce it industrially?

A small hybrid integrated circuit of our type was experimentally demonstrated by Hewlett-Packard Laboratories in Palo Alto, CA in 2009. Our current 5-year project is aimed, in particular, at demonstrations of several much more complex hybrid circuits, including a few neuromorphic CrossNets.

Do you see a possible application of crossnets in building robots?

I hope so, but this will not be very soon.

How long before we’ll have to burn into your CrossNet the three laws of robotics?

I do not think this will be a 21st century issue, but who knows?