Novel Transistor Is a Real Gate Changer

In traditional computing architectures, sensing (i.e. data input) occurs through specialized hardware such as cameras or microphones, while processing occurs through a central processing unit or graphics processing unit, and memory is stored separately in RAM or a hard drive. This separation is a sensible arrangement given the greatly differing technologies that power each of these functions.

This differs from biology, where sensing, information processing, and memory are not physically separated. Biological systems, such as the brain, are able to process information in a distributed and parallel manner, with sensory information being processed and stored simultaneously. This allows for efficient processing and recall of information.

These differences present a major problem when we are trying to emulate biological processes, as is the case with artificial intelligence. No matter how good the algorithms that we develop are, or how closely they mirror actual biological processes, they still have to be adapted to run on traditional computing architectures. Unfortunately, this is inefficient and slow for machine learning applications.

Combining sensing, memory, and processing into a single unit would greatly reduce the latency of a system and go a long way towards building artificial systems that can rival the efficiency of their biological counterparts. Researchers at Xi’an Jiaotong University and The University of Hong Kong have published the results of their recent work that may bring us closer to that goal. They have described the development of an electrochemical transistor that is capable of multi-modal sensing, memory storage, and processing.

This feat was achieved by creating organic electrochemical transistors with a vertical traverse architecture and a special type of channel that can be selectively doped by ions. This doping allows the transistors to switch between two states, simulating a volatile receptor or a non-volatile synapse. As a receptor, the device can act as a sensor, responding to different types of stimulation, like light, temperature, or ions. As a synapse, the electrochemical transistors can store 10 bits of data in an analog state for up to 10,000 seconds.

In simulation, it was shown that these transistors can be used to build an efficient spiking neural network (SNN). The network was tasked with classifying handwritten digits using the MNIST image database. The results were compared with a traditional artificial neural network, and it was found that the classification accuracy of both systems was very similar. However, the SNN based on electrochemical transistors is more efficient and faster, which means a physical implementation of such an SNN could be very useful for edge AI applications.

In another simulation, it was shown that this novel technology could also be used for more serious real-world problems, beyond the toy example of classifying handwritten digits. This demonstration showed how a neural network could be created for the efficient and real-time diagnosis of cardiac diseases. Reductions in power consumption could conceivably allow for a wearable device to be developed that can be worn for the entire day without needing frequent battery recharges.

Developing biologically inspired hardware that fuses sensing, memory, and processing is very challenging, but the efficiencies it offers could transform edge AI applications. Time will tell if this work can be developed to the point that it becomes relevant outside of a laboratory environment.