Tasks that are difficult for conventional computing systems can be performed by artificial neural networks (ANNs). ANNs have learning abilities and can performs tasks such as classification, pattern recognition and on-line learning. Practical implementation of ANNs is currently hampered by the lack of a key component that every ANN requires in large numbers – efficient hardware synapses.
Research done by the University of Southampton has recently shown that artificial systems using a nanoscale device could be used to mimic the human brain. The device is called a memristor. By doing experiments, the team demonstrated an ANN using memristor synapses that support sophisticated learning rules and perform reversible learning of noisy input data.
Limiting or regulating the flow of electrical current in a circuit, memristors can remember the amount of charge that was flowing through it and hold the data even when the power is removed.
Dr Alex Serb from Electronics and Computer Science at the University of Southampton and lead author of the paper noted that to build artificial systems that can mimic the brain in function and power, hundreds of billions artificial synapses are required. Many of these must be able to implement learning rules of different degrees of difficulty.
Currently available electronic components could conceivably be pieced together to create such synapses. The area efficiency benchmarks and required power will however be extremely difficult to achieve. This might not even be possible without designing new and bespoke ‘synapse components’.
Serb added that memristors support many fundamental features of learning synapses (two-terminal structures, on-line learning, memory storage and computationally powerful learning rule implementation). Memristors can also be manufactured in extremely compact volumes and have remarkably low energy costs, thereby offering a possible route towards creating efficient hardware synapses. Serb believes memristive synapses have to succeed for artificial brains ever to become reality.
The metal oxide memristor array was capable of learning and re-learning input patterns within a probabilistic winner take all (WTA) network in an unsupervised manner. It did this by acting like synapses in the brain. To enable the Internet of Things to process real time big data without any prior knowledge of the data, memristors would be very useful for enabling low-power embedded processors.
Dr Themis Prodromakis, Reader in Nanoelectronics and EPSRC Fellow in Electronics and Computer Science at the University of Southampton is the co-author of the paper. According to Prodromakis, the lack of practical demonstrations that display the technology’s benefits in practical applications often hampers the uptake of any new technology. He feels that the team proving that nanoscale memristors can be used to formulate in-silico neural circuits for processing big-data in real-time is a technological paradigm shift and solves a key challenge of modern society.
The team has demonstrated that such hardware platforms can adapt to its environment independently without any human intervention. The platforms are also resilient in processing even noisy data in real-time reliably.
A diverse range of applications ranging from real-time monitoring of fuel in harsh or inaccessible environments, to pervasive sensing technologies could utilize this new type of hardware. This is a highly desirable capability for enabling the Internet of Things vision.