Researchers from the Graz University of Technology created a new energy efficient artificial intelligence (AI) method. The team was inspired by how the brain operates at high power with low energy use.
TU Graz computer scientists Robert Legenstein and Wolfgang Maass (from left) are working on energy-efficient AI systems and are inspired by the functioning of the human brain. Source: Lunghammer - TU Graz
AI uses a lot of energy, which is a major hurdle especially in mobile applications. Researchers behind the new algorithm were inspired by the human brain’s ability to operate with the same power as a supercomputer but only consuming 20 watts of energy. This is thanks to the efficient transfer of information between neurons in the brain.
The team adopted this principle for their AI algorithm, called e-prop (e-propagation). Researchers used spikes in the model for communication between the neurons in the artificial neural network. Spikes are activated when information is processing.
Learning is challenging for less active AI networks because it takes longer to determine neuron connections and improve overall performance. e-prop solves this problem by means of a decentralized method copied from the brain. Each neuron documents when connections are used in the eligibility trace (e-trace).
e-prop is just as powerful as other AI algorithms. Most algorithms use machine learning to network activities stored centrally and offline to trace every few steps and show how connections are used during calculation. This requires constant data transfer, resulting in excessive energy consumption. e-prop works completely online and doesn’t require a separate memory, making it energy efficient.
The team hopes that their algorithm will drive development of a new generation of mobile learning computing systems with no need to be programmed. These systems learn according to a model of the human brain and adapt to constantly changing requirements. They also efficiently integrate a greater part of learning ability into mobile hardware components, leading to energy savings.
A paper on this research was published in Nature Communications.
