MIT solved a century-old differential equation to overcome the computational barrier of ‘liquid’ AI

MIT created an AI/ML system last year that can learn and adapt to new information while on the job, not only during its initial training period. These “liquid” neural networks (in the Bruce Lee sense) essentially play 4D chess with time-series data, making them perfect for application in time-sensitive tasks such as pacemaker monitoring, weather forecasting, investment forecasting, or autonomous vehicle navigation. However, data flow has become a barrier, and expanding these systems has become prohibitively costly in terms of computing.

On Tuesday, MIT researchers claimed that they had developed a solution to that limitation, not by expanding the data flow, but by solving a differential equation that had puzzled mathematicians since 1907. The researchers specifically addressed “the differential equation underpinning the connection of two neurons through synapses… to unleash a new kind of rapid and efficient artificial intelligence systems.”

“The new ‘CfC’ machine learning models.” “[closed-form Continuous-time] replaces the differential equation defining the computation of the neuron with a closed form approximation, preserving the beautiful properties of liquid networks without the need for numerical integration,” said MIT professor and CSAIL Director Daniela Rus in a press release on Tuesday. “Causal, compact, explainable, and efficient to train and forecast CfC models.” They pave the door for dependable machine learning in safety-critical applications.”

Differential equations, for those of us without a PhD in Really Hard Math, are formulae that may describe the state of a system at numerous discrete points or stages along the process. For example, if you have a robot arm that is going from point A to point B, you may use a differential equation to determine where it is in space between the two spots at any given phase in the process. However, calculating these equations for each step soon becomes computationally costly. MIT’s “closed form” method solves this problem by functionally modelling the full system description in a single computing step.

The team hopes that by solving this equation at the neuron level, they will be able to build models of the human brain that measure in the millions of neuronal connections, which is now not conceivable. The team also mentions that this CfC model may be able to take visual training learnt in one setting and apply it to a completely different circumstance without any further effort, a process known as out-of-distribution generalisation. That is something that current-generation models cannot perform and would be a big step towards the generalised AI systems of the future.