IIT researchers develop new device for AI, smart computing applications

Researchers from IIT Guwahati and IIT Kharagpur have developed an ultrathin device that combines visual sensing, memory and processing, using 500 times less energy in lab tests.

Researchers from the Indian Institute of Technology (IIT) Guwahati, in collaboration with IIT Kharagpur, have developed an ultrathin, light-sensitive electronic device that can sense visual information, retain it as memory and process it using significantly less energy than conventional devices, opening up possibilities for more efficient artificial intelligence (AI), robotic vision systems and smart computing applications.

 

The new device can replicate a fundamental capability of the human brain — recognising and learning from visual information without repeatedly processing every detail from scratch — by integrating light sensing, memory and information processing into a single platform.

 

Researchers said the innovation addresses a key limitation of conventional electronic vision systems, in which cameras, memory units and processors perform separate functions, requiring continuous data transfer between components and consuming substantial energy.

 

 

In laboratory tests, the device consumed 500 times less energy than conventional MXene-based devices. The researchers also demonstrated its potential in facial recognition and traffic scene segmentation, achieving 90 per cent accuracy in facial recognition and 85 per cent accuracy in traffic segmentation.

 

Explaining the difference between conventional electronic vision and biological vision, P K Giri, lead author of the research and professor in the Department of Physics at IIT Guwahati, said a conventional system first captures an image through a camera, converts the light into digital information and transfers it to a microprocessor, which analyses the image pixel by pixel to identify an object.

 

The human brain, by contrast, can recognise a familiar object by drawing on previously acquired information about its colour, shape and context. "This ability to combine perception with memory enables rapid recognition without requiring the same sequential processing of visual information. We have attempted to reproduce aspects of this mechanism using an artificial synapse, an electronic component designed to emulate how connections between neurons in the brain transmit and retain information," he said.

 

The device has been built using MXene nanosheets, an electrically conductive, ultrathin two-dimensional material, and poly(3-hexylthiophene), or P3HT, an organic semiconductor that generates electrical charges in response to light. When the two materials come into contact, light-induced electrical charges separate at their interface and remain separated for an extended period, creating a persistent electrical response even after the light source is switched off.

 

According to Giri, a brief light signal produces an electrical response that persists after illumination ends. When the device receives a second light signal, its response becomes 180 per cent stronger than that generated by the first signal, demonstrating its ability to register the effect of repeated stimulation.

 

The combination of persistent memory and a strengthening response is significant because it could allow future visual computing systems to recognise recurring patterns and learn from repeated inputs while reducing the need to transfer raw image data repeatedly between separate hardware components, he said.

 

The innovation comes amid growing interest in developing computing hardware that can perform AI-related tasks more efficiently. Modern AI systems usually rely on cameras and other sensors to interpret their surroundings, from facial recognition and autonomous navigation to industrial inspection, surveillance and traffic monitoring. Processing these large volumes of data can place considerable demands on computing power, memory bandwidth and electricity consumption.

 

In conventional computing architectures, the separation of processing and memory consumes energy in addition to that required for computation itself, a challenge commonly associated with the so-called von Neumann bottleneck. Neuromorphic computing seeks to address this limitation by bringing aspects of memory and processing closer together. The IIT researchers' device follows this approach by combining light-responsive sensing with memory-like behaviour in a single material platform.

 

Chandra Sekhar Tiwary, a professor at IIT Kharagpur, said if developed further and integrated into practical systems, such devices could reduce the energy required to process visual inputs, particularly in applications where electronic equipment must operate continuously or within strict power constraints.

 

The findings, published in the journal Carbon, could contribute to the development of neuromorphic computing systems, which mimic the information-processing mechanisms of the human brain. Such systems are being explored as a way to reduce the energy demands of AI applications as the volume of visual data processed by intelligent machines continues to grow, researchers said.