Liquid AI Is Transforming the Neural Network
Motivated by tiny worms, the founders of Liquid AI created a neural network that is more adaptable and consumes less energy. The MIT spin-off is now unveiling multiple new highly efficient models.
ARTIFICIAL INTELLIGENCE may currently be tackling advanced mathematics, engaging in intricate reasoning, and utilizing personal computers, yet the algorithms of today could still gain knowledge from tiny worms.
Liquid AI, a company that originated from MIT, will unveil multiple new AI models today that utilize an innovative type of “liquid” neural network, which could prove to be more efficient, less energy-consuming, and more transparent than existing networks that support various applications such as chatbots, image generators, and facial recognition technologies.
Liquid AI’s latest models consist of one for identifying fraud in financial transactions, another for regulating self-driving vehicles, and a third for examining genetic information. The company promoted the new models, which it is licensing to external firms, at an event hosted at MIT today. The firm has obtained backing from investors such as Samsung and Shopify, both of which are also evaluating its technology.
“We are expanding,” states Ramin Hasani, cofounder and CEO of Liquid AI, who was a co-inventor of liquid networks while studying at MIT. Hasani’s study was inspired by C. elegans, a tiny worm measuring one millimeter usually located in soil or decaying plant matter. The worm is among the rare organisms that have had its entire nervous system mapped, and it can exhibit surprisingly complex behavior even though it possesses only a few hundred neurons. “It started as merely a science project, but this technology is now fully commercialized and prepared to deliver value for businesses,” Hasani states.
Within a typical neural network, the characteristics of each simulated neuron are determined by a fixed value or “weight” that influences its activation. In a liquid neural network, the actions of each neuron are regulated by an equation that forecasts its behavior over time, with the network resolving a series of interconnected equations as it operates. The design enhances the network’s efficiency and flexibility, enabling it to continue learning post-training, which is not the case with a traditional neural network. Liquid neural networks can be examined more transparently than current models, as one can essentially rewind their behavior to understand how they generated an output.
In 2020, researchers demonstrated that a network comprised of just 19 neurons and 253 synapses, which is exceptionally small by today’s measures, was capable of managing a simulated self-driving vehicle. Although a standard neural network can assess visual data solely at fixed intervals, the liquid network effectively captures how visual information evolves over time. In 2022, the founders of Liquid AI discovered a way to streamline the mathematical effort required for liquid neural networks, making it practical for real-world applications.
Liquid AI claims it has expanded on this research and developed innovative models that it is currently keeping confidential. This September, the firm unveiled several extensive language models derived from its network architecture. The startup claims that its language model, which has 40 billion parameters, surpassed Meta’s Llama 3.1 version with 70 billion parameters on a shared set of challenges referred to as MMLU-Pro.
“The benchmark outcomes for their SLMs appear highly encouraging,” states Sébastien Bubeck, a researcher at OpenAI investigating how the architecture and training of AI models influence their abilities.
“Discovering a novel kind of foundation model is not a routine occurrence,” mentions Tom Preston-Werner, a GitHub cofounder and an early backer of Liquid AI, who points out that the transformer models that support large language models and various AI systems are beginning to reveal their constraints. Preston-Werner emphasizes that enhancing AI efficiency ought to be a major focus for all. “We must take all possible actions to ensure we do not operate coal plants for an extended period,” he states.
A limitation of Liquid AI’s method is that its networks are particularly optimized for specific tasks, especially those involving temporal data. For the technology to function with various data types, custom code is needed. Naturally, another obstacle will be convincing large corporations to ground significant projects on an entirely fresh AI framework.
Hasani states that the objective currently is to show that the advantages—such as efficiency, transparency, and energy expenses—surpass the difficulties. “We are reaching points where these models can address many of the socio-technical issues of AI systems,” he states.
