Fast Amortized Inference and Learning in Log-linear Models with Randomly Perturbed Nearest Neighbor Search

TL;DR

This paper introduces a novel method for fast inference and learning in large log-linear models by combining Gumbel perturbations with nearest neighbor search, achieving sublinear amortized runtime with proven guarantees.

Contribution

It presents a new approach that significantly speeds up inference and learning in large-scale log-linear models using random perturbations and efficient search structures.

Findings

Achieves sublinear amortized inference time.

Provides theoretical runtime and accuracy guarantees.

Demonstrates substantial speedups in experiments on ImageNet and Word Embeddings.

Abstract

Inference in log-linear models scales linearly in the size of output space in the worst-case. This is often a bottleneck in natural language processing and computer vision tasks when the output space is feasibly enumerable but very large. We propose a method to perform inference in log-linear models with sublinear amortized cost. Our idea hinges on using Gumbel random variable perturbations and a pre-computed Maximum Inner Product Search data structure to access the most-likely elements in sublinear amortized time. Our method yields provable runtime and accuracy guarantees. Further, we present empirical experiments on ImageNet and Word Embeddings showing significant speedups for sampling, inference, and learning in log-linear models.