Uncertainty In Artificial

We present a comprehensive study of theuse of generative modeling approaches forMultiple-Instance Learning (MIL) problems.In MIL a learner receives training instancesgrouped together into bags with labels forthe bags only (which might not be correctfor the comprised instances). Our work wasmotivated by the task of facilitating the di-agnosis of neuromuscular disorders using setsof motor unit potential trains (MUPTs) de-tected within a muscle which can be cast as aMIL problem. Our approach leads to a state-of-the-art solution to the problem of muscleclassification. By introducing and analyzinggenerative models for MIL in a general frame-work and examining a variety of model struc-tures and components, our work also servesas a methodological guide to modelling MILtasks. We evaluate our proposed methodsboth on MUPT datasets and on the MUSK1dataset, one of the most widely used bench-marks for MIL.