Making a Science of Model Search

TL;DR

This paper introduces a meta-modeling approach for automated hyperparameter optimization in computer vision, aiming to replace manual tuning with a reproducible, unbiased process that improves evaluation and development of algorithms.

Contribution

The work formalizes a meta-model that enables automated, objective hyperparameter tuning by exposing the computation graph of performance metrics, achieving state-of-the-art results across diverse vision tasks.

Findings

State-of-the-art results on LFW, PubFig83, and CIFAR-10.

Automated hyperparameter optimization improves reproducibility.

Formalization supports more objective evaluation.

Abstract

Many computer vision algorithms depend on a variety of parameter choices and settings that are typically hand-tuned in the course of evaluating the algorithm. While such parameter tuning is often presented as being incidental to the algorithm, correctly setting these parameter choices is frequently critical to evaluating a method's full potential. Compounding matters, these parameters often must be re-tuned when the algorithm is applied to a new problem domain, and the tuning process itself often depends on personal experience and intuition in ways that are hard to describe. Since the performance of a given technique depends on both the fundamental quality of the algorithm and the details of its tuning, it can be difficult to determine whether a given technique is genuinely better, or simply better tuned. In this work, we propose a meta-modeling approach to support automated hyper parameter optimization, with the goal of providing practical tools to replace hand-tuning with a reproducible and unbiased optimization process. Our approach is to expose the underlying expression graph of how a performance metric (e.g. classification accuracy on validation examples) is computed from parameters that govern not only how individual processing steps are applied, but even which processing steps are included. A hyper parameter optimization algorithm transforms this graph into a program for optimizing that performance metric. Our approach yields state of the art results on three disparate computer vision problems: a face-matching verification task (LFW), a face identification task (PubFig83) and an object recognition task (CIFAR-10), using a single algorithm. More broadly, we argue that the formalization of a meta-model supports more objective, reproducible, and quantitative evaluation of computer vision algorithms, and that it can serve as a valuable tool for guiding algorithm development.