Mixture Manifold Networks: A Computationally Efficient Baseline for Inverse Modeling

We propose and show the efficacy of a new method to address generic inverse problems. Inverse modeling is task whereby one seeks determine hidden parameters natural system that produce given set observed measurements. Recent work has shown impressive results using deep learning, but we note there trade-off between model performance computational time. For some applications, time at inference for best performing may be overly prohibitive its use. In seeking faster, high-performing model, present leverages multiple manifolds as mixture backward (e.g., inverse) models in forward-backward architecture. These backwards all share common forward their training mitigated by generating examples from model. The proposed thus two innovations: 1) Manifold Mixture Network (MMN) architecture, 2) procedure involving augmenting data demonstrate advantages our comparing several baselines on four benchmark problems, furthermore provide analysis motivate design.