With the advent of ChatGPT, the use of artificial intelligence (AI) has taken the world by storm. However, data inaccuracies and logical errors can often cause AI models to ‘hallucinate’ or fabricate false information, which in turn handicaps their potential in critical applications. This can be partly due to noise: not just the audio kind, but any unwanted disturbances or fluctuations in the hardware models operate on, or the data they process.
“Today’s AI systems are very powerful, but they often assume the world is perfectly clean and precise,” said Nanxi Li, a Senior Scientist at the A*STAR Institute of Microelectronics (A*STAR IME). “In reality, both data and hardware are noisy and uncertain, which causes AI systems to make confident but wrong decisions.”
While noise is an inherent problem with any computing hardware, it can be more severe on photonic devices compared to conventional electronics, due to fabrication imperfections and temperature drift. Yet researchers like Li are interested in the potential of photonic neural networks (PNNs) on photonic devices, which could enable faster and more efficient AI models than electronics-based ones.
However, rather than seeking to reduce the noise in photonic devices, Li and A*STAR IME colleagues worked with collaborators from the National University of Singapore to look into weaponising these fluctuations in PNN architecture. By deliberately treating these fluctuations as a source of randomness, they created photonic Bayesian neural networks (PBNNs)—similar to random number generators—using photonic structures on an aluminium nitride/silicon (AIN/Si) platform, which could then make probabilistic calculations.
“When humans are unsure, they do not rely on a single exact answer—they consider a range of possibilities and the probabilities thereof,” said Li. “Similarly, our photonic architecture samples a distribution of probabilities to determine its confidence in a certain decision. It’s like the difference between someone saying ‘I’m sure’ and someone saying ‘I think this is likely, but I’m not fully certain.”
The researchers tested their PBNN on the MNIST dataset—a common benchmark for handwritten digit recognition—and achieved a classification accuracy of 98 percent, remaining stable even when the device’s temperature fluctuated. The neural network was also able to identify inputs it had not previously encountered, and responded by giving low probability scores to a wide range of possible decision outcomes for such inputs.
Li noted that the system’s behaviour closely resembled how humans react when faced with unfamiliar situations. It also demonstrated this robust uncertainty awareness at a hardware level, rather than through software post-processing. “This represents an important step toward trustworthy and dependable AI systems,” Li added.
The team is confident that their architecture can support critical applications requiring rigorous uncertainty awareness, such as self-driving vehicles, medical diagnostics and edge AI systems in unpredictable settings. To realise this vision, Li and colleagues are currently scaling up the system and refining its on-chip capabilities, aiming to deliver a general-purpose photonic platform for robust and trustworthy AI.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR Institute of Microelectronics (A*STAR IME).
