RANDOMNESS IS A valuable commodity. Computer models of complex systems ranging from the weather to the stockmarket are voracious consumers of random numbers. Cryptography, too, relies heavily on random numbers for the generation of unbreakable keys. Better, cheaper ways of generating and handling such numbers are therefore always welcome. And doing just that is the goal of a project with the slightly tongue-in-cheek name of COINFLIPS, which allegedly stands for Co-designed Improved Neural Foundations Leveraging Inherent Physics Stochasticity.
COINFLIPS operates under the aegis of Brad Aimone, a theoretical neuroscientist at Sandia National Laboratories (originally one of America’s nuclear-weapons laboratories, but which has now branched out into other areas, too). Dr Aimone’s starting-point is the observation that, unlike the circuits of digital computers, which will, if fed a given input, respond with a precise and predictable output, the link between input to and output from a nerve cell is more haphazard—or, in the jargon, “stochastic”. He wants to imitate this stochastic behaviour in something less squishy than a nerve cell. By doing so, he thinks he might be able to tune the distribution of digits that a random-number generator spits out, without affecting their underlying randomness.
That would be useful. Existing random-number generators produce uniform distributions. (A “3”, say, is exactly as likely to appear as a “7”.) But, as Dr Aimone’s colleague Darby Smith notes, the real world that computer modellers are trying to model does not work like this. For example, the temperature in London in December may vary between -7°C and 17°C, but is most likely to be in the range 3°C to 8°C. Similarly, vessels are more likely to be in trouble close to a busy shipping route than in a remote backwater. Distorting uniform distributions of random numbers to take account of these realities is tedious and unsatisfactory. As Dr Smith observes, it would be more efficient if the random numbers used corresponded to the natural distribution in the first place.
There is also an abundance problem. Finding random phenomena in nature that can be transformed into computer bits is not easy. Often the source is computing itself—for example, by gathering the last digits in the numbers of milliseconds between keystrokes made by zillions of users. Otherwise, specialist, expensive hardware needs to be used to do things such as measuring heat flux through a silicon chip.
To eke out these scarce supplies, such truly random numbers are often then employed to seed programs called pseudo-random-number generators. The algorithms behind those generate sequences of numbers that have the statistical properties of randomness. But this is not the same as the real thing. As John von Neumann, one of computing’s pioneers, observed: “Anyone who considers arithmetical methods of producing random digits is, of course, in a state of sin.” Moreover, if the purpose is cryptography, this method is particularly risky. The opposition might be able to work out the algorithm involved.
Another member of the project, Shashank Misra, says COINFLIPS’ researchers have identified two hardware-based approaches for the production of tuneable, abundant random numbers. One relies on the patterns magnetic films make when disturbed, the other on how electrons travel through the barrier of a quantum-tunnelling diode. Both of these things are truly random. And both can be tuned to provide the sort of random-number distributions COINFLIPS requires.
In a quantum-tunnelling diode, electrons randomly leap (or fail to leap) across a gap, and the distribution of success or failure can be changed (and therefore tuned) by altering the voltage. A magnetic film, meanwhile, is composed of many tiny magnets jostling with each other. Flip the polarity of one of these and others around it flip in response, creating a random pattern. In this case the plan is to build tuneability into the magnetic medium itself, by tinkering with its composition, and also applying strain to it.
Random-number-generating hardware based on either of these approaches could be built directly into chips. Alternatively, a random-number-generating unit might be added to existing devices to yield an endless supply of high-speed randomness in the way that graphics-processing units are added to machines that have to handle a lot of graphical material. Crucially, unlike the heat-flux method, which requires intermediate hardware involving thousands of transistors to translate the signal into digital bits, the results of the diode and film methods can be read off as bits directly, using only one or two transistors to do so.
The efficiency offered by COINFLIPS means some tasks which currently require a supercomputer might be carried out with desktop hardware, depending on how much of the load involves generating and manipulating random numbers. Dr Aimone says the current approach tends to be simply to build bigger computers where needed. But even then some large-scale tasks may be too costly in money and time lost to conduct. You can, for example, run a model of a hurricane’s path only so many times before the real thing makes landfall.
The heart of the matter
Despite its peregrinations elsewhere, though, Sandia is ultimately in the nuclear business, and one early application of whatever COINFLIPS comes up with is likely to involve interpreting the results of collisions in particle accelerators—something the team have been exploring in collaboration with Temple University, in Philadelphia. The idea is to build a device which incorporates COINFLIPS hardware into the sensor itself. This will allow results from collisions, which will be randomly distributed, but in particular ways, to be compared with artificial random distributions, to see if they match. To be able to do this in real time is useful, because it allows an immediate decision to be made about whether or not to store a particular result. Modern colliders generate so many collisions that such immediacy in decision-making is important.
That is one, rather specific, application. But in the longer run, says Dr Aimone, COINFLIPS should enable many types of calculation that are currently impossible because of the volume of random numbers needed—for example, artificial-intelligence systems that capture the uncertainty of the world. This might be done in the form of neural networks which, like the human brain, have randomness available at each synapse. And that, in turn, may lead to COINFLIPS, a project inspired by biology, returning the compliment by providing a better way of understanding of how brains themselves work. ■