Artificial intelligence may now be able to solve advanced math, perform complex reasoning, and even use personal computers, but today's algorithms can still learn a thing or two from microscopic worms.
Liquid AI, a startup spun out of MIT, will today unveil several new AI models based on a new type of “liquid” neural network that has the potential to be more efficient, less power-consuming and more transparent than the models that underpin everything from chatbots to image generators to facial recognition systems.
Liquid AI's new models include one for detecting fraud in financial transactions, another for driving self-driving cars and a third for analyzing genetic data. The company touted the new models, which it licenses to outside companies, at an event held today at MIT. The company has received funding from investors including Samsung and Shopify, both of which are also testing their technology.
“We are scaling,” says Ramin Hasani, co-founder and CEO of Liquid AI, who co-invented liquid networking as a graduate student at MIT. Hasani's research was inspired by the C. elegansa millimeter-long worm typically found in soil or decaying vegetation. The worm is one of the few creatures whose nervous system has been mapped in its entirety, and it is capable of remarkably complex behavior despite having only a few hundred neurons. “It was once just a scientific project, but this technology is fully commercialized and fully ready to create value for companies,” says Hasani.
Within a regular neural network, the properties of each simulated neuron are defined by a static value or 'weight' that influences its firing. Within a fluid neural network, the behavior of each neuron is determined by an equation that predicts its behavior over time, and the network solves a cascade of coupled equations as the network functions. The design makes the network more efficient and flexible, allowing it to learn even after training, unlike a conventional neural network. Fluid neural networks are also open to inspection in a way that existing models are not, because their behavior can essentially be rewound to see how it produced an output.
In 2020, researchers showed that such a network with just 19 neurons and 253 synapses, which is remarkably small by modern standards, could drive a simulated self-driving car. While a regular neural network can only analyze visual data at static intervals, the fluid network very efficiently captures how visual information changes over time. In 2022, Liquid AI's founders came up with a shortcut that made the mathematical work required for liquid neural networks feasible for practical use.
