Skip to Main content Skip to Navigation

Theory and algorithms for learning metrics with controlled behaviour

Abstract : Many Machine Learning algorithms make use of a notion of distance or similarity between examples to solve various problems such as classification, clustering or domain adaptation. Depending on the tasks considered these metrics should have different properties but manually choosing an adapted comparison function can be tedious and difficult. A natural trend is then to automatically tailor such metrics to the task at hand. This is known as Metric Learning and the goal is mainly to find the best parameters of a metric under some specific constraints. Standard approaches in this field usually focus on learning Mahalanobis distances or Bilinear similarities and one of the main limitations is that the control over the behaviour of the learned metrics is often limited. Furthermore if some theoretical works exist to justify the generalization ability of the learned models, most of the approaches do not come with such guarantees. In this thesis we propose new algorithms to learn metrics with a controlled behaviour and we put a particular emphasis on the theoretical properties of these algorithms. We propose four distinct contributions which can be separated in two parts, namely (i) controlling the metric with respect to a reference metric and (ii) controlling the underlying transformation corresponding to the learned metric. Our first contribution is a local metric learning method where the goal is to regress a distance proportional to the human perception of colors. Our approach is backed up by theoretical guarantees on the generalization ability of the learned metrics. In our second contribution we are interested in theoretically studying the interest of using a reference metric in a biased regularization term to help during the learning process. We propose to use three different theoretical frameworks allowing us to derive three different measures of goodness for the reference metric. These measures give us some insights on the impact of the reference metric on the learned one. In our third contribution we propose a metric learning algorithm where the underlying transformation is controlled. The idea is that instead of using similarity and dissimilarity constraints we associate each learning example to a so-called virtual point belonging to the output space associated with the learned metric. We theoretically show that metrics learned in this way generalize well but also that our approach is linked to a classic metric learning method based on pairs constraints. In our fourth contribution we also try to control the underlying transformation of a learned metric. However instead of considering a point-wise control we consider a global one by forcing the transformation to follow the geometrical transformation associated to an optimal transport problem. From a theoretical standpoint we propose a discussion on the link between the transformation associated with the learned metric and the transformation associated with the optimal transport problem. On a more practical side we show the interest of our approach for domain adaptation but also for a task of seamless copy in images
Document type :
Complete list of metadatas

Cited literature [182 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Wednesday, February 13, 2019 - 3:44:07 PM
Last modification on : Monday, January 13, 2020 - 5:46:07 PM
Long-term archiving on: : Tuesday, May 14, 2019 - 7:36:11 PM


Version validated by the jury (STAR)


  • HAL Id : tel-02018083, version 1


Michaël Perrot. Theory and algorithms for learning metrics with controlled behaviour. Machine Learning [cs.LG]. Université de Lyon, 2016. English. ⟨NNT : 2016LYSES072⟩. ⟨tel-02018083⟩



Record views


Files downloads