Artificial intelligence (AI) has become a seamless part of daily life, from unlocking our phones with facial recognition to predictive text in emails. Behind this technology lies deep learning and neural networks (NNs), complex systems designed to mimic the human brain’s ability to process data and make decisions. However, as these systems grow in capability, their unpredictable nature presents significant challenges, especially in high-stakes applications like space exploration and healthcare.
At Florida Tech’s NEural TransmissionS (NETS) Lab, researchers are tackling these challenges head-on, working to understand why neural networks sometimes fail and developing methods to mitigate these risks. Led by Ph.D. student Mackenzie Meni, the lab has pioneered a technique called PEEK (Peeking into the Inner workings of Neural Networks), which visualizes the decision-making process of NNs and identifies biases that may lead to errors. This groundbreaking technique has shown promise in identifying “correct” outputs even when the network fails, offering a potential fail-safe to improve AI reliability.
NNs work by simulating the behavior of interconnected artificial neurons, each with a “knob” that controls how signals are passed. During the “training” process, these knobs are adjusted to minimize errors, allowing the network to map vast datasets to their correct outputs. However, NNs can still make mistakes for reasons that are often difficult to pinpoint, making them risky for critical tasks. The NETS Lab’s research aims to unveil the hidden mechanisms behind these errors, improving both the safety and accuracy of AI systems.
The lab’s work extends across various fields, including aerospace and medicine. In collaboration with the Autonomy Lab, NETS researchers have developed vision and guidance algorithms for autonomous satellite swarms used by the Air Force Research Laboratory (AFRL). These algorithms enable satellites to perform complex tasks like navigation and image capture, with ongoing work focused on human-guided vision systems.
Additionally, the lab is working on real-time satellite tracking algorithms and 3D imaging projects for spacecraft inspection, further advancing spaceflight technology. Ph.D. student Arianna Issitt and her team are also exploring optimal satellite orbits for conducting inspections around space vehicles, crucial for maintaining space safety.
In healthcare, the NETS Lab is collaborating with the Multiscale Cardiovascular Fluids Laboratory to develop NNs that estimate blood flow dynamics within patient blood vessels. By analyzing real-time data noninvasively, these models could revolutionize cardiovascular diagnostics, enabling medical professionals to make quicker, more accurate decisions in treating heart conditions.
(Based on an article originally published in Science X)
