New ways of computing are coming to the fore as development in conventional computing slows. A team of engineers is trying to pioneer a form of computation at Penn State that mimics the efficiency of the neural networks of the brain while leveraging the analog nature of the brain.
Digital, consisting of two states, on-off or one and zero, is modern computing. An analog computer, like the brain, has several states that are possible. That is the difference between turning on or off a light switch and turning a dimmer switch to different lighting numbers.
According to Saptarshi Das, the team leader and Penn State assistant professor of engineering science and mechanics, neuromorphic or brain-inspired computing has been studied for over 40 years. What’s new is that the need for high-speed image processing , for example, for self-driving vehicles, has risen as the boundaries of digital computing have been reached. Another catalyst in the pursuit of neuromorphic computing is the rise of big data, which needs forms of pattern recognition for which the brain architecture is especially well suited.
The problem is that you have to store the memory in one location and do the computation somewhere else, “Das said.” We have strong machines, no doubt about that.
It takes a lot of energy to shuttle this data from memory to logic and back again and slows the speed of computation. This computer architecture, however, needs a lot of space. This bottleneck could be removed if computing and memory storage could be housed in the same room.
Thomas Shranghamer, a doctoral student in the Das group and first author of a paper recently published in Nature Communications, explained, “We are developing artificial neural networks that try to mimic the energy and area efficiencies of the brain.” “The brain is so small that it can sit on top of your head, while two or three tennis courts are the size of a modern supercomputer.”
The artificial neural networks the team is developing can be reconfigured by applying a brief electric field to a sheet of graphene, the one-atomic-thick layer of carbon atoms, like synapses linking the neurons in the brain that can be reconfigured. At least 16 possible memory states are shown in this work, as opposed to the two in most oxide-based memristors, or memory resistors.
“What we have shown is that, using simple graphene field effect transistors, we can control a large number of memory states with precision,” Das said.
The team thinks it is feasible to ramp up this technology to a commercial scale. With many of the biggest semiconductor businesses aggressively exploring neuromorphic computing, Das believes this work would be of interest to them.
This thesis was supported by the Army Research Office. A patent for this invention has been filed by the team.
Source of Story: Penn State Supplied Materials. Initial Walt Mills Writing. Note: For style and length, material can be edited.