Artificial intelligence may now be solving advanced mathematics, performing complex reasoning, and even using personal computers, but today’s algorithms can still learn a thing or two from tiny bugs.
Liquid AIa startup spun out of MIT, today announces several new AI models based on a new type of “liquid” neural network. This neural network could be more efficient, consume less power, and be more transparent than the one that underpins everything. From chatbots to image generators to facial recognition systems.
Liquid AI’s new models include models for detecting fraud in financial transactions, controlling self-driving cars, and analyzing genetic data. At an event at MIT today, the company touted a new model it is licensing to outside companies. The company has received funding from investors including Samsung and Shopify, both of which are testing its technology.
“We’re scaling,” he says Ramin Hassanico-founder and CEO of Liquid AI, co-invented Liquid Networks as a graduate student at MIT. Hassani’s research was inspired by: nematodeare millimeter-long insects usually found in soil or rotten vegetation. This nematode is one of the few organisms whose entire nervous system has been mapped, and despite having just a few hundred neurons, it is capable of highly complex behaviors. “Once just a science project, this technology is fully commercialized and fully poised to bring value to companies,” Hassani said.
Within a typical neural network, the properties of each simulated neuron are defined by static values or “weights” that influence its firing. within liquid neural networkthe behavior of each neuron is governed by an equation that predicts its behavior over time, and the network solves a cascade of linked equations as a network function. This design makes the network more efficient and flexible, and unlike traditional neural networks, it can learn even after training. Liquid neural networks can also be examined differently than existing models, as you can essentially rewind their behavior to see how they produced the output.
In 2020, researchers showed that such a network with just 19 neurons and 253 synapses, significantly small by modern standards, could control a simulated self-driving car. While regular neural networks can only analyze visual data at static intervals, liquid networks are very efficient at capturing changes in visual information over time. In 2022, the founders of Liquid AI came up with a shortcut This makes the mathematical effort required for liquid neural networks practical.
