Title: Neural Langevin Machine: A Local Asymmetric Learning Rule Fosters Creativity

Abstract: Recurrent neural networks (RNNs) can utilize their fixed points to both store and produce information. By capturing these fixed points through the Boltzmann-Gibbs measure, researchers have developed neural Langevin dynamics, a method applicable for identifying these states during the generative learning of real-world datasets. This approach, termed the "neural Langevin machine," yields a learning rule that is asymmetric and adjusts based on firing rates. Crucially, it relies solely on local neural signals, offering significant biological plausibility for local predictive learning. The study highlights an intriguing out-of-equilibrium regime within the generative process, alongside a transition from memorization to generalization as the volume of training data increases. Furthermore, this neuro-inspired model is capable of continuously exploring phase space to generate various types of images and can effectively remove noise from corrupted images.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC