Neuromorphic circuit can scale up more easily with disordered synapses

A human brain contains about 8.3 × 10⁹ neurons connected by about 6.7 × 10¹³ synapses, immense quantities which are impractically large for neuromorphic computing circuits. Uday Goteti and Robert Dynes have simulated a superconducting artificial neural network that can be scaled up much more easily, due to the disorder of its synaptic pathways.

Instead of establishing distinct synaptic connections between pairs of individual neurons, like a traditional network, this circuit connects neurons through a disordered array of superconducting loops that contain Josephson junctions. This allows for high connectivity with less hardware than traditional networks.

“We’re trying to replace individual synapses with a single device that can behave like all of them combined together,” Goteti said. “This allows us to build very large networks without taking up much real estate, and it could have other potential advantages in energy efficiency and computing power.”

Mathematicians have previously established that increasing the randomness or disorder in a network can increase the performance of that network. In the simulated network of Goteti and Dynes, current paths through the network are continuously changing depending on the memory state the device is in. Whenever a current in one path gets large enough, a Josephson junction changes its direction.

Because of the dynamic nature of the current paths, this kind of disordered network has a large number of memory states, which increase exponentially with a linear increase in network size. In their simulations, Goteti and Dynes have demonstrated that this kind of network is as functional as an individually connected network, with greater scalability of memory and power efficiency.

Goteti said the next step is to experimentally realize and demonstrate the proposed circuits, and then design experiments to study the emergent properties and learning behaviors of such disordered array synaptic networks.

Source: “Superconducting neural networks with disordered Josephson junction array synaptic network and leaky integrate-and-fire loop neurons,” by Uday S. Goteti and Robert C. Dynes, Journal of Applied Physics (2021). The article can be accessed at https://doi.org/10.1063/5.0027997 .