What is Liquid AI, and could it prove more effective than other types of AI?
New research into neural nets and algorithms has revealed what some call “Liquid AI,” a more fluid and adaptable version of artificial intelligence.
In my previous episode, I discussed the basics of AI and the limitations that hold it back. It looks like Liquid AI could provide the very solutions that the AI community has been searching for.
In this episode of Short and Sweet AI, I explore the new research behind Liquid AI, how it works, and what it does better than other types of AI.
Listen to the episode below or keep reading to learn more.
Neural Nets have Traditional Limitations
As discussed in the last episode, neural nets are the powerhouses of machine learning. They can translate entire books in just seconds with Google Translate. They can transform written text into images with DALL·E or discover the 3D structure of a protein in hours with AlphaFold.
But even neural nets capable of all these great things have limitations.
Neural nets cannot do anything other than what they’re trained for. To function as well as they do, neural nets are programmed with parameters set to give the most accurate results. The problem is that this makes them brittle when presented with any new information they weren’t trained on.
Another big limitation in AI is that deep learning neural nets are so complex, with layers of billions of parameters, they require powerful supercomputers to run the algorithms. This means only big corporations and tech companies have the resources to unlock AI’s true powers.
A Neuroplastic Neural Net Based on a Nematode
Researchers from MIT and Austria’s Science Institute have created a new, adaptive neural network they describe as “Liquid AI.”
The researchers designed algorithms based on the nervous system of a simple worm – C. elegans.
This worm has only 302 neurons, but it’s very responsive and can display various behaviors. The research teams have mathematically modeled the worm’s neurons and build them into a neural network.
Computer Software with Neuroplasticity
The worm-brain algorithm has a much simpler design than the vast neural nets we’ve seen elsewhere – and yet it accomplishes similar tasks.
In current AI architecture, the neural net’s parameters are locked into the system after training. With Liquid AI, these parameters are fluid and can change with time and experience. That’s because it’s based on the mathematical models of the worm’s neurons.
So, when it encounters new information, it can adapt. While it may be an artificial brain built with software, it shows a kind of built-in neuroplasticity like a human brain.
A test using autonomous vehicles found that these new algorithms could keep a vehicle in its lane. It proved to be just as accurate and efficient as other more complex machine learning algorithms.
The worm-brain model could also adapt to new pathways. Researchers found that the algorithm could change its underlying mathematical equations when it had new information – for example, when there was rain on the vehicle’s windshield.
With this “neuroplasticity”, it means the neural net is less likely to break in the face of new data.
Liquid AI Uses Fewer Parameters and GPUs
Liquid AI has seen researchers reduce the neural net’s size. While it still has 75,000 trainable parameters, it’s considerably fewer than the millions or billions required in some machine learning algorithms.
With fewer parameters to support, this decreases the computing power needed to run the algorithms. As computing power is one of the primary limitations of AI, this has generated a great deal of excitement among researchers.
Liquid AI is an adaptable machine learning algorithm that consumes less power, uses a smaller neural net, and still manages to be as accurate as larger machine learning systems. So, is this the future of AI?
New Liquid AI is More Transparent
Saving the best for last, one of the other great benefits of Liquid AI is that it’s more transparent.
For years, AI ethicists and researchers have been deeply troubled by machine learning systems being “black boxes.” How they work and arrive at their results is largely impenetrable. No one can determine exactly what’s going on within the neural nets to get the results that it does.
The worry is, what happens when unsupervised machine learning models are trained on the unfiltered internet? There’s no way to know or control what they learn.
Liquid AI, on the other hand, is designed differently. It’s a new type of architecture that makes it more open to observation and study. This means that researchers can analyze the neural net’s decision-making process and understand how it arrived at those answers.
It’s more transparent, but not just that. Liquid AI is more adaptable, efficient, smaller, accurate, and transparent. This could have a massive effect on the AI world, which could shape it for years to come.