Fast Amortized Inference and Learning in Log-linear Models with Randomly Perturbed Nearest Neighbor Search
This addresses efficiency issues in natural language processing and computer vision tasks where output spaces are large but enumerable, offering incremental improvements for domain-specific applications.
The paper tackles the computational bottleneck of inference in log-linear models, which scales linearly with output space size, by proposing a method using Gumbel perturbations and nearest neighbor search to achieve sublinear amortized cost with provable guarantees, resulting in significant speedups in experiments on ImageNet and word embeddings.
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.