Researchers Discover a More Flexible Approach to Machine Learning | Quanta Magazine

Abstraction: MIT liquid neural networks inspired by C. elegans offer adaptive continuous-time inference

Key points:

• Ramin Hasani and Mathias Lechner at MIT CSAIL created liquid neural networks inspired by Caenorhabditis elegans, the millimeter roundworm with a fully mapped nervous system (302 neurons)
• Each neuron is governed by a differential equation predicting behavior over time; unlike traditional networks that output at fixed intervals, liquid networks describe system state at any continuous time point (e.g., 0.53 s or 2.14 s)
• Synaptic connections use nonlinear probabilistic functions—not single weight values—making the network's response variable based on input, hence "liquid"
• Key 2022 breakthrough: Hasani found a closed-form (pencil-and-paper) approximate solution to the otherwise intractable nonlinear synapse equations, eliminating the need for iterative solvers and speeding computation by "several orders of magnitude" with no accuracy loss (per Sayan Mitra, UIUC)
• A 19-neuron, 253-synapse liquid network successfully steered an autonomous car; a drone trained in forests was then deployed in urban Cambridge with encouraging early results
• Planned improvements: determine minimum neurons needed per task and adopt selective C. elegans-like wiring (not fully connected); applications include power grids, financial transactions, weather, and brain activity simulation

Connections: Mit · Ramin Hasani · Mathias Lechner · Liquid Neural Networks · Neural Networks · Machine Learning

Source: https://www.quantamagazine.org/researchers-discover-a-more-flexible-approach-to-machine-learning-20230207/