This project studies how to improve decentralized stochastic gradient descent when data are unevenly distributed across agents. The goal is to learn a sparse communication topology that connects agents in a way that reduces the negative effect of data heterogeneity while keeping communication costs low. We will evaluate the method on standard image classification benchmarks, mainly CIFAR-10, using a 100-agent decentralized learning setup with label-skewed data. The experiments will compare the learned topology with standard fixed topologies, including random and fully connected graphs. The requested NAISS resources will be used to run GPU-based training experiments and measure convergence speed, accuracy, and communication cost. Main supervisor: Erik G. Larsson