Compositional, structured models are appealing because they explicitly decompose problems and provide interpretable intermediate outputs that give confidence that the model is not simply latching onto data artifacts . Learning these models is challenging because end-task supervision only provides a weak indirect signal on what values the latent decisions should take . In this work, we introduce a way to leveragepaired examples that provide stronger cues for learning latent decisions . We apply ourmethod to improve compositional question answering using neural module networks on the DROP dataset . We empirically demonstrate that our proposed approach improves both in- and out-of-distribution generalization andleads to correct latent decision predictions . We prove that ourproposed approach improves in-and-out-of the distribution generalization, and leads to correct decision predictions. We also demonstrated that our proposals that our proposal improved both in and out of-distributive generalization of the model’s ‘targets’
Author(s) : Nitish Gupta, Sameer Singh, Matt Gardner, Dan RothLinks : PDF - Abstract
Code :
https://github.com/oktantod/RoboND-DeepLearning-Project
Keywords : latent - decision - generalization - models - indirect -