Readspeaker menu

Uncertainty Prediction with Large-Scale GP and HIK

Alexander Freytag, Erik Rodner, Paul Bodesheim, and Joachim Denzler

 

[Freytag12:RUC]

Alexander Freytag and Erik Rodner and Paul Bodesheim and Joachim Denzler. Rapid Uncertainty Computation with Gaussian Processes and Histogram Intersection Kernels. Asian Conference on Computer Vision (ACCV). 2012. (Oral). [pdf] [bib] [poster] Best Paper Honorable Mention
Corresponding project: Large-scale Gaussian Processes with Flexible Adaptive Histogram Kernels


Code available!

We released the software for our ECCV and ACCV paper on fast Gaussian process inference with histogram intersection kernels on github. The software is written in C++ and requires our computer vision library NICE-core.

lgpl-logo

Problem Statement

Although we reduced computational costs for learning, classification, and hyperparameter optimization for Gaussian Processes equipped with histogram intersection kernels by orders of magnitudes in our ECCV'12 paper, computing the predictive variances is not possible with these techniques. In addition, a system shall be able to be non-stationary, i.e., it should be able to incorporate new training examples over time. Telling a long story short, we tackle the following problems

  • Reducing the costs for computing the predictive variance (different levels of granularity)
  • Allowing the GP-HIK framework for incremental learning
  • Developing a new query strategy for active learning derived from GP mean and variance


Summary

An important advantage of Gaussian processes is the ability to directly estimate classification uncertainties in a Bayesian manner. In this paper, we develop techniques that allow for estimating these uncertainties with a runtime linear or even constant with respect to the number of training examples. Our approach makes use of all training data without any sparse approximation technique while needing only a linear amount of memory. To incorporate new information over time, we further derive online learning methods leading to significant speed-ups and allowing for hyperparameter optimization on-the-fly. We conduct several experiments on public image datasets for the tasks of one-class classification and active learning, where computing the uncertainty is an essential task. The experimental results highlight that we are able to compute classification uncertainties within microseconds even for large-scale datasets with tens of thousands of training examples.


Downloads


Back to main project page.