In the realm of artificial intelligence, bigger is supposed to be better. Neural networks with billions of parameters power everyday AI-based tools like ChatGPT and Dall-E, and each new large language model (LLM) edges out its predecessors in size and complexity. Meanwhile, at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), a group of researchers have been working on going small.
In recent research, they demonstrated the efficiency of a new kind of very small—20,000 parameter—machine-learning system called a liquid neural network. They showed that drones equipped with these excelled in navigating complex, new environments with precision, even edging out state-of-the art systems. The systems were able to make decisions that led them to a target in previously unexplored forests and city spaces, and they could do it in the presence of added noise and other difficulties.
Neural networks in typical machine-learning systems learn only during the training process. After that, their parameters are fixed. Liquid neural networks, explains Ramin Hasani, one of the CSAIL scientists, are a class of artificial intelligence systems that learn on the job, even after their training. In other words, they utilize “liquid” algorithms that continuously adapt to new information, such as a new environment, just like the brains of living organisms. “They are directly modeled after how neurons and synapses interact in biological brains,” Hasani says. In fact, their network architecture is inspired by the nervous system of living creatures called C. elegans, tiny worms commonly found in the soil.
“We can implement a liquid neural network that can drive a car, on a Raspberry Pi”. —Ramin Hasani, MIT’s CSAIL
The purpose of this experiment wasn’t just the robust autonomous navigation of a drone, Hasani says. “It was about testing the task-understanding capabilities of neural networks when they are deployed in our society as autonomous systems.”
As training data for the neural networks that would control the drone, the researchers used drone footage collected by a human pilot flying toward a target. “You expect the system to have learned to move towards the object,” Hasani says, without having defined what the object is, or provided any label about the environment. “The drone has to infer that the task is this: I want to move towards [the object].”
The team performed a series of experiments to test how learned navigational skills transferred to new, never-seen-before environments. They tested the system in many real-world settings, including in different seasons in a forest, and in an urban setting. The drones underwent range and stress tests, and the targets were rotated, occluded, set in motion, and more. Liquid neural networks were the only ones that could generalize to scenarios that they had never seen, without any fine-tuning, and could perform this task seamlessly and reliably.
The application of liquid neural networks to robotics could lead to more robust autonomous navigation systems, for search and rescue, wildlife monitoring, and deliveries, among other things. Smart mobility, according to Hasani, is going to be crucial as cities get denser, and the small size of these neural nets could be a huge advantage: “We can implement a liquid neural network that can drive a car, on a Raspberry Pi.”
Beyond Drones and Mobility
But the researchers believe liquid neural networks could go even farther, becoming the future of decision making related to any kind of time series data processing, including video and language processing. Because liquid neural networks are sequence data-processing engines, they could predict financial and medical events. By processing vital signs, for example, models can be developed to predict the status of a patient in the ICU.
Over and above their other advantages, liquid neural networks also offer explainability and interpretability. In other words, they open the proverbial black box of the system’s decision-making process. “If I have only 34 neurons [in the drone system], I can literally go and figure out what is the function of each and every element,” says Hasani. That’s something that would be virtually impossible in a large-scale deep neural network. The smaller size of liquid neural nets also massively reduces the computational cost, and therefore the carbon footprints, of machine-learning models.
Hasani and his colleagues are looking for ways to improve liquid neural networks. “This paper covered a very controlled and straightforward kind of reasoning capability, but real-world interactions require more and more sophisticated reasoning problems,” he says. The team would like to design more complex tasks and test liquid neural networks to their limit, while also figuring out why liquid neural networks perform so much better than their competitors in reasoning tests.
Hasani explains liquid neural networks in this video:
[embedded content]Liquid Neural Networks | Ramin Hasani | TEDxMITyoutu.be
Original Source: https://spectrum.ieee.org/liquid-neural-networks