Brain-Inspired AI Liquid Neural Networks Power Drones

Placidplace/Pixabay

Artificial intelligence (AI) is enabling drones to perform real-time analysis, mapping, and tracking of information from their cameras. Researchers from the Massachusetts Institute of Technology (MIT) Computer Science and Artificial Intelligence Laboratory (CSAIL) demonstrate how brain-inspired AI liquid neural networks enable drones to navigate unfamiliar places in their study published this week in Science Robotics.

“The results in this paper open the door to the possibility of certifying machine learning solutions for safety critical systems,” said senior co-author Daniela Rus, Director at MIT CSAIL, and the Andrew (1956) and Erna Viterbi Professor of Electrical Engineering and Computer Science at MIT.

Ramin Hasani is the study senior co-author and co-first author along with co-first author Makram Chahine, and researchers Alexander Amini, Mathias Lechner, Ryan Shubert, Aaron Ray, and Patrick Kao.

“Studying natural brains effectively narrows the search space of possible algorithms for acquiring intelligent behavior,” wrote the researchers.

For this study, the MIT CSAIL researchers used liquid neural networks to create a learning-based solution to enable drones to navigate in unfamiliar environments. The inspiration for liquid neural networks is from biological organisms, namely, the microscopic nematode C. elgans, which can produce complex dynamics with just a little over 300 neurons. This new class of AI algorithms was first introduced in 2020 by Hasani and Lechner, along with Rus, Amini, and Radu Grosu.

Liquid time-constant networks (LTCs) are a class of time-continuous recurrent neural networks (RNNs). Unlike standard artificial neural networks, liquid neural networks handle complexity better and are more fluid, robust, expressive, and interpretable. In effect, liquid neural networks exhibit artificial neuroplasticity. In standard neural networks, a weight reflects the strength of the connection between artificial neurons. In contrast, in liquid networks, each neuron is managed by a nonlinear function which allows for variability; hence it is a probabilistic process.

The researchers performed a series of tests to see how well the pre-trained neural model could generalize when faced with data it had not been trained on prior. These closed-loop tests included flying to the target, flight range, stress test, attention profile of networks, target rotation and occlusion, hiking with adversaries, a triangular multistep loop between objects, and dynamic target tracking.

The scientists compared six recurrent neural network architectures, which included temporal convolutional network (TCN), ODE–recurrent neural network (RNN), gated recurrent unit (GRU)–ordinary differential equation (ODE), closed-form continuous-time (CfC), neural circuit policy (NCP), and long short-term memory (LSTM). The closed-form continuous-time and neural circuit policy algorithms were developed by the MIT CSAIL researchers.

Source: MIT CSAIL

Each algorithm model was tested in four environments consisting of urban lawn, urban patio, training woods, and alternative woods. According to the researchers, their liquid neural networks were able to generalize to new tasks and novel environments better than the popular state-of-the-art recurrent neural networks.

“We had our reasons to hope liquid neural networks would outperform other architectures on the fly-to-target task, based on the favorable properties inherent to them by design,” said Chahine. “But all that is on paper. I was really surprised to see how consistently they performed flight navigation in the real world. As the tasks we considered got more and more complex, only liquid networks held their ground, while all other state-of-the-art architectures failed.”

Copyright © 2023 Cami Rosso All rights reserved.