Artificial intelligence (AI) has been at the cutting edge of computer research for years but has been held back by one thing. The chip architecture, that is the underlying hardware on which the artificial intelligence is based upon, is acting as a bottleneck. The idea of AI has been to apply human brain learning techniques and patterns to machines powered by computer chips. So far the speed at which the human brain can learn has never been matched. What our brains can do in seconds takes a neural network, a server room, megawatts of power and weeks. Researchers from MIT have been focusing their efforts on this and have developed components that mimic the human brain more than traditional computer components.
The hope is that these “neuromorphic” processors will be able to power the next generation of AI. The technology promises to lower the amount of energy required. The chips are made from layers of programmable resistors. By repeating these arrays of resistors, researchers can create a network of artificial “neurons” and “synapses”. This new material is also compatible with existing silicon fabricating devices so it could pave the way for deeper integration with commercial computing. “With that key insight, and the very powerful nanofabrication techniques we have at MIT.nano, we have been able to put these pieces together and demonstrate that these devices are intrinsically very fast and operate with reasonable voltages” explains senior author Jesús A. del Alamo, a professor in MIT’s Department of Electrical Engineering and Computer Science (EECS). In essence, computations can be carried out in parallel and if the size of the matrix of resistors expands, more operations can be completed simultaneously.
The MIT researchers argued that the current components, built for long term memory storage, are ill suited to the requirements of AI. Previous research has focused on designing electronic components that control the flow of current based on how much charge previously flowed through the device. Mimicking the basic idea of how biological synapses work, these components could in principle be used to create neural networks. The problem with this approach is that the technology the components rely on does not allow for high-speed switching and transfer of data. The MIT team have designed a new component to help solve this problem. The device, made of phosphosilicate glass (PSG), has protons removed or added to regulate the strength of the transfer. The switch to this new technology has dramatically sped up the data transfer.
Some scientists have voiced the scalability of the technology. The components are mere nano-metres across and are capable of transferring pulses of up to 10 volts 10,000 times faster than the human brain. Whether or not the technology can be used at scale this is an interesting concept and a promising direction in the search for new hardware that can match the human brain.
