Learning Randomly Perturbed Structured Predictors for Direct Loss Minimization

Abstract

Direct loss minimization is a popular approach for learning predictors over structured label spaces. This approach is computationally appealing as it replaces integration with optimization and allows to propagate gradients in a deep net using loss-perturbed prediction. Recently, this technique was extended to generative models, while introducing a randomized predictor that samples a structure from a randomly perturbed score function. In this work, we learn the variance of these randomized structured predictors and show that it balances better between the learned score function and the randomized noise in structured prediction. We demonstrate empirically the effectiveness of learning the balance between the signal and the random noise in structured discrete spaces.

Proceedings of the 38th International Conference on Machine Learning, 2021

References in corpus (11)

• The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables
• Learning Latent Permutations with Gumbel-Sinkhorn Networks
• Ranking via Sinkhorn Propagation
• Smooth Loss Functions for Deep Top-k Classification
• SparseMAP: Differentiable Sparse Structured Inference
• Learning with Differentiable Perturbed Optimizers
• Stochastic Optimization of Sorting Networks via Continuous Relaxations
• Differentiable Top-k Operator with Optimal Transport
• Improved Gradient-Based Optimization Over Discrete Distributions
• Direct Optimization through $\arg \max$ for Discrete Variational Auto-Encoder
• Direct Policy Gradients: Direct Optimization of Policies in Discrete Action Spaces

Cited by in corpus (1)

• Combinatorial Optimization for Panoptic Segmentation: A Fully Differentiable Approach