Sep, 2020 - By WMR
Ever since artificial intelligence has gained mainstream attention, it has become a major subject of research. AI has shown how it can emulate human brain capacities, even though it’s up to a certain extent. Now, researchers are focused on creating a human brain-inspired electronic system, to augment the AI capabilities as much as by 1,000 times more energy efficient. Researchers from University College London (UCL) are very close to creating an energy-efficient artificial intelligence to improve the accuracy of the brain-inspired computing system. This system, as reported, utilizes an innovative electronic device named memristors to form an artificial neural network, which is extremely energy efficient as compared to traditional transistor-based AI.
The current AI we have is extremely energy-intensive. To put that in perspective, training one simple AI can generate around 284 tons of carbon dioxide, which is equivalent to the lifetime emissions of five cars. Researchers have been pondering on how to reduce this enormous carbon emission, where they found a device that was developed back in 2008. Memristors is a novel electronic device, which has the capability to reduce a fraction of ton of carbon dioxide emitted, which roughly equals to an afternoon’s drive. Another advantage of memristors is that they can contain large amounts of computing power than conventional computing systems, typically in handheld devices. This, in turn, eliminates the need for an internet connection. The UCL researchers published a study in ‘Nature Communications’, where they argued about how accuracy can be improved massively by getting memristors to work together in various subgroups of neural networks and taking the mean of their calculations. This means errors in each of the networks can be canceled out. Memristors are often described as “resistors with memory” since they can recall the amount of electric charge passed through them. They can remember this charge even after the current is turned off.
Memristors were introduced a decade ago where they were used in electronics to support resistors, capacitors, and inductors. Since they have been heavily used in memory devices. However, engineers believe they can be used to enhance AI systems. Memristors can vastly improve energy efficiency due to the fact that they operate in binary code of ones and zeros as well as several levels of ones and zeros at the same time. This means, more information can be picked into each bit. Besides, memristors have been touted as neuromorphic or brain-inspired. Similar to the brain, processing, and memory can be implemented in the same adaptive building blocks. On the contrary, the current computer system wastes a lot of energy in data movement. Engineers have found that their new method increased the accuracy of the neural networks of typical AI tasks.
According to Dr. Mehonic, the director of the study, their team has found a novel approach that can enhance device-level as well as system-level behavior. They also found that arranging the neural network into multiple smaller networks instead of one big network can lead to greater accuracy. Co-author prof. Tony Kenyon believes memristors can take a leading role in more energy-suitable times of edge computing and IoT devices.