liquid neural network

A continuous-time recurrent neural networkrecurrent neural network, n.: A neural network that reads a sequence, such as a sentence or an audio clip, one piece at a time, carrying a running summary forward so that what came earlier can shape how later pieces are understood., described by Ramin Hasani and colleagues, whose neurons each follow a simple first-order differential equation with a time constant that is not fixed: a learned nonlinear function of the current input and state speeds each unit up or slows it down, so its responsiveness shifts with the data, hence 'liquid'. Running the network means integrating these equations numerically, as with a neural ordinary differential equationneural ordinary differential equation, n.: A neural network that, instead of passing data through a fixed stack of layers, describes how the data changes continuously over 'time' and uses an equation solver to follow that change.; the authors proved that the states stay bounded and reported better time-series predictions than earlier recurrent models. A closed-form approximation (CfC, 2022) avoids the solver and runs much faster. Compact 'neural circuit policies' built from such neurons, with only a few dozen units, were demonstrated steering vehicles and drones with inspectable dynamics. The name is also associated with Liquid AI, a company co-founded in 2023 by several of the method's authors, whose foundation models use related but different hybrid architectures.

A liquid neural network with 19 control neurons kept the car in its lane from camera input.