It is no secret that artificial and biological neural networks operate in very different ways. And these differences are clearly visible when these systems are observed in action. When a person sees a car, for instance, they immediately recognize it for what it is, even if it looks quite different from every other example they have previously seen. The process is robust against significant environmental differences and is highly energy-efficient, as well. This stands in stark contrast to artificial systems, which perform poorly if they have not been trained on similar examples, and which consume a great deal of energy by comparison.
In an effort to address this performance gap, researchers have been working to design neuromorphic hardware, which seeks to more closely approximate the function of the human brain. Now, a team of engineers at RMIT University has developed a neuromorphic device that brings this vision closer to reality. Specifically, they have developed a compact, energy-efficient vision processor made from molybdenum disulfide (MoS₂), a two-dimensional material only a single atom thick. This proof-of-concept device mimics how the human eye and brain work together to perceive, process, and remember visual information, without the need for a conventional computer.
A comparison between a biological neuron and a spiking model (📷: T. Aung et al.)
The team has demonstrated that their device can detect movement, process that information, and store it as a memory. Unlike traditional digital systems that rely on capturing images frame by frame, this device uses edge detection to identify changes in a scene, such as a waving hand, with significantly less computational load.
The device does this by emulating the behavior of leaky integrate-and-fire neurons — an important component of spiking neural networks. These neurons accumulate incoming signals until a threshold is reached, then emit a spike and reset. The MoS₂ device mirrors this behavior through its photoelectric response, capturing light and translating it into electrical impulses, just like biological neurons do.
This advance could lead to neuromorphic vision systems that respond almost instantly to environmental changes, with applications ranging from autonomous vehicles to collaborative robotics. The ability to operate without power-hungry digital processing makes the technology particularly attractive for real-time, energy-sensitive tasks in unpredictable environments.
A schematic of the device design (📷: T. Aung et al.)
Using recent developments in chemical vapor deposition techniques, the researchers aim to create larger arrays of MoS₂ devices in the future, enabling the development of more complex and higher-resolution neuromorphic vision systems. This would expand the technology’s capabilities and integration into hybrid systems combining analog and digital computation.
The work is still in the prototype stages, and practical applications are likely still many years away. But with additional work, this technology may ultimately change how machines see and understand the world around them.
