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.
In this episode find out:
Important Links & Mentions
Hello to you who are curious about AI, I’m Dr. Peper. Machine learning algorithms are getting an overhaul from a very unlikely source. It’s a fascinating story.
Neural Nets have Traditional Limitations
Neural nets are the powerhouse of machine learning. They have the ability to translate whole books within seconds with Google Translate, change written text into images with DALLE, and discover the 3D structure of a protein in hours with AlphaFold. But researchers have struggled with neural networks because of their limitations.
Neural nets cannot do anything other than what they’re trained for. They’re programed with parameters set to give the most accurate results. But that makes them brittle which means they can break when given new information they weren’t trained on. Today the deep learning neural nets used in autonomous driving have millions of parameters. And the newest neural nets are so complex, with hundreds of layers and billions of parameters, they require very powerful supercomputers to run the algorithms.
A Neuroplastic Neural Net based on a Nematode
Now researchers from MIT and Austria’s Science Institute have created a new, adaptive neural network they’re describing as “liquid” AI. The algorithm’s based on the nervous system of a simple worm, C. elegans. And elegant it truly is. This worm has only three hundred and two neurons but it’s very responsive with a variety of behaviors. The teams were able to mathematically model the worm’s neurons and build them into a neural network. I’ve explained neural networks in my previous episode called A Simple Explanation of AI.
Computer Software with Neuroplasticity
The worm-brain algorithm is much simpler than the huge neural nets and yet accomplishes similar tasks. In current AI architecture, the neural net’s parameters are locked into the system after training. With liquid AI based on the mathematical models of the worm’s neurons, the parameters are able to change with time and with experience. This is a fluid neural net. As it encounters new information, it adapts. It’s an artificial brain created out of computer software but shows a kind of built-in neuroplasticity like a human brain.
When the algorithm was tested on the task of keeping an autonomous vehicle in its lane, it was just as accurate and efficient as more advanced and complex machine learning algorithms. The worm-brain model also adapts new pathways. In one example the researchers found the algorithm could change its underlying mathematical equations when it had new information like, there was rain on the autonomous vehicle’s windshield. This “neuroplasticity” means the neural net is less likely to break when it’s given data it hasn’t been trained on.
Liquid AI Uses Less Parameters, Less GPUs
Also, with this new approach, the researchers reduced the neural net’s size. It has only 75,000 trainable parameters instead of the million or billion parameters in some machine learning algorithms. This decreases the GPUs or computing power needed to run the algorithm. You can appreciate the excitement this has generated. Liquid AI is an adaptable machine learning algorithm that consumes less power, uses a smaller neural net, while being as accurate as the larger machine learning systems.
New Liquid AI is More Transparent
But I saved the best for last. For many years AI ethicists and researchers have been deeply troubled by machine learning systems being “black boxes” meaning how they work and arrive at their results is largely impenetrable. No one can determine exactly what’s going on within the neural nets that lead to the successful results. This can be a big problem when unsupervised machine learning models are trained on the unfiltered internet because there’s no way of knowing or controlling what they learn.
But this AI system was designed differently. It’s a new type of AI architecture. This liquid neural net is more open to observation and study. Researchers are able to analyze the neural net’s decision making and diagnose how it arrived at the answers. It’s more transparent. It’s adaptable, efficient, smaller, accurate, and transparent. Liquid AI, now that’s a good thing.
Thanks for listening. I hope you found this helpful. Be curious and if you like this episode, click the thumbs up button and leave a comment. From Short and Sweet AI, I’m Dr. Peper.