Active learning of functional networks from spike trains

Abstract

Learning functional networks from spike trains is a fundamental problem with many critical applications in neuroscience. However, most of existing works focus on inferring the functional network purely from observational data, which could lead to undiscovered or spurious connections. We demonstrate that by adopting experimental data with interventions applied, the accuracy of the inferred network can be significantly improved. Nevertheless, doing interventions in real experiments is often expensive and must be chosen with care. Hence, in this paper, we design an active learning framework to iteratively choose interventions and learn the functional network. In particular, we propose two models, the variance model and the validation model, to effectively select the most informative interventions. The variance model works best to reveal undiscovered connections while the validation model has the advantage of eliminating spurious connections. Experimental results with both synthetic and real datasets show that when these two models are applied, we could achieve substantially better accuracy than using the same amount of observational data or other baseline methods to choose interventions.