Revolutionising Aviation: MIT’s LNNs set to reshape AI application

MIT scientists have recently unveiled Air-Guardian, an advanced deep-learning system designed to enhance flight safety by collaborating with airplane pilots. This AI co-pilot is capable of identifying situations where human pilots might overlook essential details and taking action to prevent potential incidents.

Air-Guardian relies on Liquid Neural Networks (LNN), a novel deep learning system developed by the MIT Computer Science and Artificial Intelligence Lab (CSAIL). LNNs have already demonstrated their effectiveness in various fields and show promise, especially in domains that require efficient and transparent AI systems, potentially as an alternative to commonly used deep learning models.

Air-Guardian’s approach to improving flight safety is distinctive. It continuously monitors both the human pilot’s attention and the AI’s focus, swiftly identifying instances where their attention diverges. When a critical aspect eludes the human pilot, the AI system intervenes to manage that specific flight element.

The fundamental concept behind this human-AI partnership is to maintain the pilot’s control while allowing the AI to step in when necessary. For instance, if an aircraft flies dangerously close to the ground, resulting in unpredictable gravitational forces that might cause the pilot to lose consciousness, an Air-Guardian intervenes to prevent potential accidents. In other scenarios where the human pilot is overwhelmed by excessive on-screen data, the AI efficiently filters and highlights crucial information that might have been overlooked.

Air-Guardian uses eye-tracking technology to monitor human attention and employs heatmaps to indicate where the AI system’s focus is directed. When a disparity between the two is detected, Air-Guardian assesses whether the AI has identified an issue requiring immediate attention.

Air-Guardian, like other control systems, operates on a deep reinforcement learning model. This model comprises an AI agent powered by a neural network, making decisions based on observations in the environment. The agent is rewarded for making accurate choices, enabling the neural network to gradually learn a policy for making the right decisions in diverse situations.

The unique aspect of Air-Guardian is its reliance on LNNs at its core. LNNs are known for their transparency, enabling engineers to comprehend the model’s decision-making process.

LNNs excel in acquiring causal relationships within their data, unlike traditional neural networks that often grasp incorrect or superficial correlations, leading to unexpected errors in real-world applications.

Another valuable attribute of Liquid Neural Networks is their compactness. Unlike conventional deep learning networks, LNNs can handle complex tasks with fewer computational units or “neurons.” This compactness equips them to function effectively on devices with limited processing power and memory, such as self-driving cars, drones, robots, and aviation.

The insights derived from the development of Air-Guardian have the potential to be applied across a wide spectrum of scenarios where AI collaborates with humans.

Liquid Neural Networks could also play a pivotal role in the flourishing domain of autonomous agents. They may empower AI agents, such as virtual CEOs, capable of making and elucidating decisions for their human counterparts, aligning their values and goals with those of humans.

LNNs have the potential to become a versatile foundational model for various applications, paving the way for potent AI systems on edge devices like smartphones and personal computers.