New Method Learns Task Relationships for Decentralized Multitask AI

Traditional multitask learning often assumes a predefined structure for how different tasks relate to each other. However, in many real-world scenarios, this underlying task relationship is unknown and needs to be discovered from the data itself, especially in decentralized systems where data is distributed. Researchers have developed a novel two-phase approach to address this challenge. The first phase involves estimating a generalized graph Laplacian, which represents the task relationships, directly from noisy, non-cooperative stochastic gradient updates across a network. Once this graph is learned, the second phase leverages it to enable cooperative multitask diffusion learning, allowing different tasks to benefit from shared insights. This framework is grounded in a Gaussian Markov random field prior, leading to a decentralized maximum likelihood estimator for the graph Laplacian. Simulations confirm that this cooperation, guided by the learned task graph, substantially enhances performance compared to independent learning, closely matching the performance achieved with a known, true task graph when estimation parameters are well-tuned.