Oregon State University (OSU) College of Engineering researchers have achieved a significant breakthrough in developing energy-efficient artificial intelligence. They’ve engineered an experimental light-sensitive memory device that integrates three distinct functionalities: sensing, data storage, and initial signal processing.
Current machine vision and AI accelerator systems operate on a segregated principle. An image is first captured by an image sensor, then transferred to memory, and subsequently processed by a separate processor. This method necessitates constant transmission of large data volumes between components, leading to reduced operational speed and increased energy consumption.
The OSU innovation tackles this challenge through the concept of in-sensor computing. The core idea is that a portion of the information processing occurs concurrently with data acquisition, directly within the sensor itself, eliminating the need for continuous data transfer between discrete modules. Professor Larry Cheng, the project lead, explains that this approach significantly bolsters information processing efficiency right at the sensor level.
At the heart of this device is a hybrid phototransistor structure, combining two distinct materials. The lower layer is composed of a semiconductor that facilitates rapid electrical current transfer, while the upper layer consists of an organic photosensitive material that responds to light by generating electrical charges.
When light strikes the device, some of these generated charges become trapped within the upper layer, persisting even after the light source is removed. These “trapped” charges continue to influence the conductivity of the lower layer, enabling the transistor to effectively retain a memory of the captured light signal.
A key characteristic of this development is that this memory is controllable. By applying a small voltage, the researchers can manipulate the position of these “captured” charges. Moving them closer to the conductive channel prolongs the signal’s persistence and strengthens the memory. Conversely, moving them away weakens the effect and gradually erases the information.
This behavior mirrors the way the human brain functions, where certain memories are reinforced while others fade over time. In this context, electrical control over the material’s state serves the role of biological processes, allowing for the definition of a “lifespan” for each signal.
This innovation falls under the domain of neuromorphic computing—the creation of systems that operate based on principles similar to the brain’s neural networks. For cameras, drones, autonomous vehicles, and robotics, this could usher in a new era of more efficient data processing, enabling systems to immediately distinguish critical information from secondary data and conserve resources by avoiding the storage of superfluous information.
Currently, the technology exists only as a laboratory prototype, and its widespread industrial application is still some way off. However, in the future, such solutions hold the promise of making AI systems considerably faster, more compact, and more energy-efficient.
About the Author
