How liquid neural networks could boost AI capabilities

BOSTON: US researchers are working on new machine learning methods that could outperform anything available today.

Known as liquid neural networks, these systems are easily adaptable and flexible, even after training, and could find potential applications in autonomous transport and robotics.

For several years, the Computer Science and Artificial Intelligence Laboratory (CSAIL) at Massachusetts Institute of Technology (MIT) has been working to develop new technology based on a liquid neural network, recently demonstrating its first practical applications.

This type of network, directly inspired by the functioning of biological neurons, is today the most advanced form of machine learning.

Such systems use differential equations, which have the advantage of being easily modified in real time, to adapt to new data or new environments. All these equations are similar to those relating to the behaviour of fluids, which is why they have been named "liquid" networks. Their first applications are mainly in the field of mobility: today's drones, tomorrow's autonomous cars.

An initial test was carried out using a drone powered by this new form of intelligence. The MIT researchers were able to make it fly autonomously in an unseen environment, without incident, as it was able to adapt automatically to everything around it. Generally speaking, in robotics, this type of network could be used with machines operating in environments that are completely unknown or simply too dangerous for humans.

In the automotive field, autonomous cars could obviously benefit from these liquid neural networks. This technology could help such vehicles adapt to difficult weather conditions or unforeseen, dangerous situations. The idea would be to have safer autonomous transport.

Not only are these liquid neural networks more flexible and adaptable, but they are also faster in their calculations and even less data-intensive than the machine learning models used today. But there's still a long way to go to optimise these networks and make them usable in everyday life.
Autos News