September 05, 2020 by Luke James
Inspired by the human brain, a research team from the Graz University of Technology has developed an energy-efficient artificial intelligence system.
One of the biggest hurdles to the widespread adoption of powerful artificial intelligence, or AI, solutions is the high amount of energy that artificial neural networks consume. This is especially relevant in smaller applications, such as mobile devices, where battery life is limited.
An approach that has long since been touted as a potential solution is one which is inspired by the human brain, which, despite being as powerful as a supercomputer, only requires 20 watts to operate (in other words: around one-millionth of the amount of energy used by a supercomputer). The reason for this low energy consumption is that information transfers at high efficiency in the human brain thanks to its neurons, which send short electrical pulses to other neurons only when it is necessary to do so.
Robert Legenstein and Wolfgang Maass from the Graz University of Technology (TU Graz) hold up a laptop to show information about their artificial intelligence algorithm. Image Credit: TU Graz via Informationsdienst Wissenschaft.
Now, a research team at the Graz University of Technology (TU Graz) has adopted this principle in the development of a new machine learning algorithm called e-propagation (e-prop). In this model, the researchers use spikes for communication between neurons in an advanced neural network. Much like in the human brain, these ‘spikes’ only become active when they are being used for information processing in the network.
Normally, the fact that spikes are inactive when they’re not being used would be bad news for machine learning. This is because it takes longer for observations to determine which neural connections improve overall network performance. In this study, however, the researchers solved this problem by using a decentralised method inspired by the human brain, known as an ‘eligibility trace’ or ‘e-trace’.
In e-tracing, each instance of a neural connection is documented by the neuron itself, allowing machine learning to take place as normal. According to the researchers, the e-trace method is just as powerful as current best-known learning methods.
By potentially solving such a prominent challenge, the TU Graz team hopes that its e-prop model will pave the way for the development of a new generation of mobile learning computing systems, which don’t even need to be programmed. Rather, they’ll learn according to the human brain model and adapt to dynamic requirements accordingly.
The ultimate goal, the researchers say, is to enable these computing systems to learn without being too energy-intensive and to detach them from cloud dependence. To achieve this, they’ll need to integrate the core part of the learning ability into mobile hardware components. This will be tricky, but the research team believes they can pull it off.
Currently, the TU Graz research team is working with the Advanced Processor Technologies Research Group of the University of Manchester to integrate the e-prop model into the neuromorphic SpiNNaker system. The team is also working with Intel researchers to integrate the algorithm into the next version of Loihi, Intel’s neuromorphic chip.
