AI's Artificial Synapse

2019-12-09

In science fiction novels, the "positronic brain" of robots enables them to work and live like humans, with great processing power and speed, making them even smarter than humans. However, in reality, the human brain still outperforms CPUs in cognitive tasks such as pattern recognition, requiring only 20 watts of energy to complete a task, while supercomputers require over 1 million watts of energy. For this reason, researchers are now focusing on neuromorphic computers and artificial neural networks, which work more like the human brain. However, with current technology, replicating the inherent spatiotemporal processes of the brain, such as short-term and long-term memory, in artificial neural networks (SNNs) is not only challenging but also very expensive. Dr. Xiong Feng from the University of Pittsburgh has been awarded a $500,000 career award from the National Science Foundation (NSF) for his research on a missing element - a dynamic synapse. This synapse can significantly improve energy efficiency, bandwidth, and cognitive capabilities of SNNs. "When the human brain sees rain, it feels wet, and when it sees fire, it feels hot. The synapses in the brain connect these two ideas, so in the future, it will associate rain with wetness and fire with warmth. These two ideas are closely linked in the brain," explained Dr. Feng about his research. "On the other hand, computers require a large amount of data input to perform the same task. Our dynamic synapse will simulate the brain's ability to create neural connections as a time difference between stimuli, significantly improving the energy efficiency required to perform tasks." Existing non-volatile memory devices have been studied as artificial synapses in SNNs, but unfortunately, they have not met the standards. Their design aims to permanently retain data and is not suitable for the spatiotemporal dynamics and high precision that the human brain can achieve. In the brain, not only the information is important, but also the timing of the information - for example, in certain cases, the closer two pieces of information are in time, the stronger the synaptic chain between them. By programming the conductors to carry more current and form stronger neural connections, its functionality becomes more like the synapses in the human brain, placing more weight on more closely related items during the learning process. The electrical conductivity changes in dynamic synapses will have both short-term and long-term components, simulating the memory and learning abilities of the human brain. Although researchers have previously demonstrated this technology in the laboratory, this is the first time it has been applied to SNNs. This novel attempt could lead to widespread applications in artificial intelligence, as well as revolutionary advancements in cognitive computing, autonomous vehicles, and automated manufacturing.

RECENT AI NEWS

RECENT AI TOOLS