Australian Firm Turns Brain-Inspired Tech to Robotics Applications
February 2, 2021
Australian-based Strategic Elements Ltd. (SOR) announced proof-of-concept work that highlights its printable neuromorphic technology’s potential for data processing and self-learning in soft robotics (touch sensing) and other signal processing applications such as computer vision.
The group said early-stage results show that in the case of computer vision, the technology uses less power to operate than the human brain (less than 10 watts), and is able to use multiple resistance states with the potential capacity to process multiple points of data. The technology was successfully operated at an ultra-low level of 0.8V and in the microamps range of current, SOR said.
The printable neuromorphic hardware is being developed from the company’s Nanocube Memory Ink technology and is conducted in the Nanoionics laboratory at the University of New South Wales (UNSW).
While artificial neural networks are not uncommon, most synapse networks exist only as software, the group said. The UNSW team is in early-stage development of neural network hardware designed to be printable, low cost, portable, ultra-low power, flexible, and semi-transparent. These features would be suited for robotics and computer vision applications, the group explained. “For example, the ability to place flexible neuromorphic hardware onto soft robotics and in health or manufacturing sectors or devices requiring such low power that battery or energy harvesting technology (e.g., humidity) could potentially be used as a power source,” SOR said in a statement.
The team was able to fabricate a memristor device with its Nanocube Ink. A memristor is an electronic memory device that mimics the information-transmitting synapses in the human brain to carry out complex computational tasks. Memristor devices are used for storing as well as processing information and are known to emulate the memory and learning properties of biological synapses.
The artificial synapse on the Nanocube Ink-created device was tested for endurance, known as long-term potentiation (learning) and depression (forgetting). SOR said this was done with a programming pulse algorithm to pulse the artificial synapse with a series of positive and negative voltages (also known as spikes), mimicking neurons firing in the human brain.
The artificial synapse was pulsed for 5,000 cycles (10 positive voltage spikes and 10 negative voltage spikes per cycle), and no significant degradation was observed after a total of 100,000 spikes, SOR reported. “The synaptic weight (conductance) of the artificial synapse was gradually increasing with positive voltage pulses, indicating the potentiation behavior and gradually decreasing with negative pulses, indicating the depression behavior,” SOR said. “This shows that the artificial synapse has good potential endurance after repeated learning and forgetting cycles.”
The memristors were programmed at an ultra-low level of 0.8V, but the company said early-stage work showed potential to program the memristors as low as 0.4V since the memristors’ values were successfully ready at 0.1V. In addition, the artificial synapse showed the ability to have multi-level switching, up to 10 resistance states per cell. “This is an important feature as neuromorphic computing systems designed for complex applications, such as image processing, require multiple resistance rates,” SOR said. “This is in contrast with other memory devices that only use two resistance states (high and low), which present disadvantages in synaptic weight performance such as low accuracy or area efficiency.
The company said it will assess a potential program of work between the computer vision and robotics team at a subsidiary, Stealth Technologies, and the materials team at UNSW to develop a prototype application. Future work aims to reduce the temperature required in the manufacturing process, fabrication on flexible substrates, and to increase the number of memristors or artificial synapses in the thousands, to meet requirements of image recognition and tactile touch sensors in robotics.