Researchers develop frugal artificial synapse for neural networks
Researchers at Stanford and Sandia have developed an organic synapse based on the design of a battery. The synapse could combine high performance and energy efficiency and could eventually be used for neural networks. TU/e wants to further develop the synapse.
The scientists have named their synapse Enode, which stands for electrochemical neuromorphic organic device. The goal is to arrive at an actual neural network instead of a simulation of a neural network, claims Yoeri van de Burgt, one of the researchers and now affiliated with Eindhoven University of Technology. They were inspired by the way in which synapses in the brain learn by sending signals back and forth. The synapse is the contact point of neurons between which the electrical and chemical impulses in the brain take place. They are the foundation of the nervous system.
The current computer architecture is not very suitable for imitating the neural system. “The deep learning algorithms are very powerful, but the calculations and simulation of electrical signals is done by processors, after which the storage on memory takes place elsewhere. That is not efficient in terms of energy and time,” says Van der Burgt. The brain does this much better: over time, connecting signals to the synapses takes less and less energy.
They therefore made a synapse based on a completely different architecture. The artificial synapse consists of two thin, flexible film layers with an electrolyte in between. The source and drain are arranged on the bottom layer and a contact point ensures the application of a voltage on the top layer. For example, the synapse can be charged or discharged a little bit, whereby the resistance of the bottom layer increases or decreases. The changed state of the resistance is retained when the voltage is lost and the memory can represent more values than just 0 or 1. In this way the synapse can be trained and it is possible to predict which voltage is needed to obtain a certain electrical state, similar to the ‘training’ of synapses in the brain.
Researchers have been experimenting for some time with non-volatile memory such as memristors for neural networks. However, the invention of the Stanford researchers would require considerably less energy and could also take on more than five hundred different values. It takes 10 picojoules of current to move the artificial synapse to another position. In addition, the organic synapses could be fabricated on flexible substrates, for use in stretchable electronic systems, and it involves the use of inexpensive polymers. Van de Burgt even hints at the possibility that they could one day be used for data connections to the brain.
For now, it is a single working synapse. The researchers deduced that their architecture can work from simulations of use in a network. That simulated network was able to correctly identify written numbers from 0 to 9 with an accuracy of 93 to 97 percent. Van den Burgt now wants to work on a series of a thousand artificial synapses in Eindhoven to test their operation in practice.
The research titled A non-volatile organic electrochemical device as a low-voltage artificial synapse for neuromorphic computing is published by Nature Materials.