Researchers Push the Boundaries of Neuromorphics With Retina-inspired UV Sensor
In this article, we will look at a retina-inspired photonic synapse that may further develop sensory AI systems.
Researchers at Seoul National University and Inha University in South Korea have recently developed a retina-inspired photonic synapse that may help further develop sensory AI systems.
This isn't the first breakthrough drawing inspiration from human physiology, though.
In our article discussing recent research that stretches the meaning of “sensor” and “battery,” we learned about a hydrogel substance called "AISkin" that acts as a sensor while mimicking the flexibility and durability of human skin. Another instance is "organs-on-a-chip" or chips capable of recreating core physiological functions.
The researchers who discovered the retina-inspired photonic synapse launched the project as a response to shortcomings of neural nets in both software and hardware.
The Need for Neuromorphic Designs
Executing neural nets can be done in software or hardware, but each has its own problem. A software-based neural net execution system has the capability to adapt and learn because software itself is re-writeable. However, software execution often involves plenty of instructions, which significantly increases execution time.
Diagram of one level of a neural processing system. Image (modified) used courtesy of Carver A. Mead, California Institute of Technology
This execution time can be reduced with the use of hardware AI accelerators, but these have their own problems. Generally speaking, AI accelerators are very good at executing a neural net pattern with weighted functions, but cannot be reconfigured at the hardware level.
This is where neuromorphic designs come in.
By replicating how neurons work in the brain—with neurons being able to forge new connections—a system could run an AI system and improve its behavior on the fly. Major semiconductor companies have taken note of this powerful technology including Intel. Intel has used neuromorphic technology to design Loihi, a "manycore processor with on-chip learning" capabilities.
Loihi is a self-learning neuromorphic research chip. Image used courtesy of Intel
Traditional silicon does not, however, allow for the removal of electrical connections between transistors with ease. While FPGAs can simulate this behavior by making dynamic changes to how logic blocks are connected, it still doesn’t properly simulate neuron behavior, especially considering the supporting hardware needed to perform dynamic changes.
The Retina-inspired Synapse
Researches at Seoul National University and Inha University in South Korea developed a retina-inspired UV sensor based on carbon nitride to solve problems in sensor-related AI applications.
In their study, the researchers explain that photo-sensitive neuromorphic electronics are, in their estimation, the "core technology for applications of next-generation smart sensors."
The researchers further explain that photo-sensitive neuromorphic electronics "can efficiently replicate the functions of biological synapses (interconnection between two neurons, key roles in learning and memorizing) and detect various types of external light information."
The sensor is said to be capable of selectively responding to UV light thanks to a carbon nitride nanodot (C3N4). The carbon nitride nanodot is used as a floating gate in a transistor and exhibits a response similar to a neuron found in the eye.
The artificial photo-sensitive neuromorphic devices are designed to mimic a biological retina. Image used courtesy of Seoul National University
The wavelength of light that the UV layer is sensitive to is between 100 nm and 400 nm, which is a band of UV light that is harmful to human health. The idea behind the neuron-like sensor is that smart systems involving sensors can adapt to their environments at any moment.
While a typical sensor will record immediate levels, this sensor will, over time, store electrons in its floating gate—acting as a form of memory. Therefore, the conduction of the transistor will directly correspond to exposure.
This is a trivial form of data-processing and the output of the transistor could be directly coupled to an analog circuit, thereby removing the need for a processor.
Will We See This Technology in the Future?
Professor Tae-Woo Lee at Seoul National University explains relevant applications in which we might see this unfolding technology.
"This smart system platform will be widely applicable to advanced electronic skin that is able to automatically adapt to the changing light-dose environment, smart windows that can selectively control transmittance of strong UV lights, smart glasses that detect and block harmful UV rays, smart sensors, artificial retinas for soft humanoid robots, and neural prostheses compatible with biological optic nerves."
This new technology has a similar end goal to Bosch's new Virtual Visor, which uses an LCD panel, an AI algorithm, and liquid crystal technology to selectively filter or block light—in Bosch's case, to shadow a driver's eyes without obstructing his or her view.
Bosch's Virtual Visor uses an LCD panel, an AI algorithm, and liquid crystal technology. Image used courtesy of Bosch
Neuromorphic systems can be thought of as analog circuits whose characteristics change as their environment changes. Components such as memristors will be key to the neuromorphic circuit since their properties can be changed without a processor.