New Chips for Neural Networks to Improve Energy Efficiency of AI Systems


A nanochip. / Photo by: science photo via Shutterstock


Neural networks are the great boon of systems equipped with artificial intelligence. While the technology enables AI computers to do intense multitasking such as data gathering and analysis, neural networks have one major downside – withering hardware components.

Neural networks are computer systems modeled after the human brain and the nervous system. They replicate the synapses of nerve cells, connectors that enable communications between cells in the brain. Most AI systems are built and implemented in the software level because mimicking the human central nervous system requires a high amount of power and specialized hardware.

Researchers at IBM developed microelectronic synapses, chips based on the two types of synapses in the brain. The first type of synapses is short-term ones that can work for computation, while the second type is the long-term one that can function for memory. In neural networks, the two types solve common known issues like low accuracy.

The research team tested microelectronic synapses in a neural network capable of two image recognition tasks -- handwriting and color image classification. When they used the chips, the system was able to perform the tasks with the same accuracy level of a software-based neural network. But the chip only consumed one percent as much energy as a standard neural network. Although the result seems simple and small, a large-scale neural network utilizing the chips would be able to perform tasks with 100 times more energy efficiency, compared to the usual neural networks.

"A factor of 100 in energy efficiency and in training speed for fully connected layers certainly seems worth further effort,” explained Michael Schneider, a researcher at the National Institute of Standards and Technology.

The new chips from IBM still require further testing before these can be commercialized. Also, the chips also need some redesign tweaks since the sizes are quite heavy due to the five transistors and three other parts.