Neuromorphic technology from IBM and Intel may prove useful for analyzing X-rays and stock markets
Brain-inspired "neuromorphic" microchips from IBM and Intel may be useful for more than artificial intelligence. A new study says they may also be ideal for a class of computing useful in a wide range of applications, including analyzing medical X-rays and financial economics.
Scientists have long tried to mimic the brain's operation with programs known as neural networks and hardware known as neuromorphic chips. Until now, neuromorphic computing has mostly focused on implementing neural networks. It was unclear whether this hardware could be useful beyond AI applications.
Neuromorphic chips typically mimic the work of neurons in various ways, such as performing many calculations in parallel. Also, just as biological neurons both compute and store data, neuromorphic hardware often aims to combine processors and memory, potentially reducing the energy and time that conventional computers lose when moving data between these components.
While conventional microchips use signals triggered at regular time intervals to coordinate circuit activity, activity in neuromorphic architecture often appears as spikes that fire only when an electrical charge reaches a certain value, much like what happens in a brain like ours.
So far, the main advantage of neuromorphic computing has been energy efficiency: features such as spikes and the combination of memory and processing led to IBM's TrueNorth chip, whose power density was four orders of magnitude lower than that of conventional microprocessors of the time.
“From many studies we know that neuromorphic computing will have energy efficiency advantages, but in practice people don't care about saving energy if it means you work much more slowly”
Now scientists believe neuromorphic computers may be well suited to so-called Monte Carlo methods, where problems are essentially treated as games and solutions are reached through many random simulations or "walks" through these games.
“We arrived at random walks by looking at problems that do not scale very well on conventional hardware. Typically, Monte Carlo solutions require a large number of random walks to provide a good solution, and although what each walker does over time is individually easy to compute, in practice the need to perform many such walks becomes prohibitively high”
In contrast, instead of simulating in parallel a group of random walkers, each performing its own calculations, scientists can program a single neuromorphic grid of circuits to represent all the computations a random walk can perform, and then, thinking of each random walk as a spike moving across the grid, solve the entire problem at once.
In particular, just as previous studies showed that quantum computing can demonstrate a "quantum advantage" over classical computing on a large set of tasks, researchers found that a "neuromorphic advantage" may exist when it comes to random walks through discrete-time Markov chains.
“If you imagine the problem as a board game, the "chain" here means playing through a sequence of states or spaces. Markov chains mean that the next space you can move to depends only on your current space, not on your previous history, as in board games such as Monopoly or Candy Land. "Discrete time" simply means that there is a fixed interval of time between the changing spaces — a turn”
In experiments with IBM's TrueNorth and Intel's Loihi neuromorphic chips, a server-class Intel Xeon E5-2662 processor and an Nvidia Titan Xp graphics processor, the scientists found that when it comes to solving this class of problems at large scales, neuromorphic chips proved more energy efficient than conventional semiconductors and competitive, if not better, in terms of time.
“I strongly believe that neuromorphic computing for AI is very interesting, and that brain-based hardware will lead to smarter and more powerful AI. But at the same time, by showing that neuromorphic computing can be effective in conventional computing applications, the technology could have a much broader impact on society”
One way neuromorphic chips achieve their advantage in performance and energy efficiency has been a high degree of parallelism. This was compounded by the ability to represent each random walk as a single burst of activity instead of a more complex set of actions.
“A major limitation of these Monte Carlo methods is that we need to simulate many walkers. Since spikes are such a simple way to represent a random walk, adding an extra random walker is simply adding one extra spike to the system that will run in parallel with all the others. As a result, the time cost of an additional walker is very low”
The ability to efficiently solve this class of problems has a wide range of potential applications, such as modeling stocks and options in financial markets, better understanding how X-rays interact with bones and soft tissues, tracking how information moves through social networks, and modeling how diseases spread through populations. Aimone believes that "applications of this approach are everywhere".
But the fact that a neuromorphic advantage may exist for some tasks "does not mean that brain-inspired computers can do everything better than conventional computers," Aimone warns. In his view, the future of computing is likely to involve a combination of different technologies. Scientists are not trying to say that neuromorphic computing will replace everything.
Currently, scientists are studying ways to manage interactions between multiple "walkers" or participants in these scenarios, which would enable applications such as molecular dynamics simulation. They are also developing software tools to help other developers work on this research.
