Title: Efficient Sampling-Based Bayesian Active Learning

Abstract: Bayesian Active Learning (BAL) is an efficient framework for learning the parameters of a model, in which input stimuli are selected to maximize the mutual information between the observations and the unknown parameters. However, its applicability to real experiments is limited by two main drawbacks. Firstly, it requires performing high-dimensional integrations and optimizations in real time: current methods are either too time consuming, or only applicable to specific models. Secondly, it only optimizes for the next stimulus input (an approach referred to as a myopic design), disregarding all future observations in the experiment. To address these issues, we propose an Efficient Sampling-Based Bayesian Active Learning (ESB-BAL) framework, which is efficient enough to be used in real-time biological experiments and to go beyond myopic approaches. We apply our method to the problem of estimating the parameters of a chemical synapse from its postsynaptic responses evoked by presynaptic action potentials. Synaptic parameter inference is a challenging application of BAL, as computation needs to be faster than the smallest inter-stimulation interval, which can be on the order of a few milliseconds. Using synthetic data and synaptic whole-cell patch clamp recordings in cerebellar brain slices, we show that our method has the potential to significantly improve the precision of model-based inferences, thereby paving the way towards more systematic and automated experimental designs.