Machine Learning

Pierre-Edouard Honnet, Pedro Oliveira Pinheiro, and Olivier Canévet just completed their PhD theses at Idiap (Martigny) affiliated with EPFL. Their research is in the field of Artificial Intelligence (Speech for Pierre-Edouard, and Computer Vision for Pedro and Olivier). They present their research in the frame of a small "Machine Learning" seminar.

----------------------------------------------------------------------

Speaker: Pierre-Edouard Honnet
Title: Intonation Modelling for Speech Synthesis and Emphasis
Preservation
Presentation in English

Speech-to-speech translation is a framework which recognises speech in
an input language, translates it to a target language and synthesises
speech in this target language. This presentation will deal with
aspects of speech-to-speech translation which are lost in traditional
systems. Motivated by the Swiss multilingual context, the development
of an intonation model will be presented, and some of its application
for speech synthesis. The model is physiologically plausible, and an
automatic extraction method is proposed to retrieve its parameters. In
a conversation scenario, it is interesting to be able to preserve word
emphasis, which indicates what in the sentence is important, or what
is the implicit message of the speaker. Following, we apply the model
to word emphasis synthesis, using random forest to predict word level
intonation contours.

----------------------------------------------------------------------

Speaker: Pedro Oliveira Pinheiro
Title: Large-Scale Image Segmentation with Convolutional Neural Networks
Presentation in English

Object recognition is one of the most important problems in computer
vision. A main challenge is the problem of variability: objects can
appear across huge variations in pose, appearance, illumination and
occlusion, and a visual system need to be robust to all these
changes. In this presentation, I will show how we can leverage
information of large-scale datasets to deal with pixel-level
recognition. We aim to algorithms that require the least amount of
feature engineering and are easy to scale.

----------------------------------------------------------------------

Speaker: Olivier Canévet
Title: Object Detection with Active Sample Harvesting
Présentation en français

En vision par ordinateur (computer vision) ou en apprentissage
automatique (machine learning), que ce soit pour entraîner un
classifieur d'images ou un détecteur d'objets, la phase
d'apprentissage se résume à trouver une frontière de décision optimale
entre les classes. En pratique, les exemples d'apprentissage n'ont pas
tous la même importance. Certains sont aisément classifiés, tandis que
d'autres, proches de la frontière ou mal classés, sont ceux qui ont de
l'importance. Cependant, la plupart des méthodes d'apprentissage
sélectionnent les exemples et les images de manière uniforme, et leur
accordent la même importance. Le but de nos travaux a été de mettre au
point des méthodes pour trouver efficacement les exemples
d'apprentissage les plus informatifs, sans jamais accéder à la
totalité de l'ensemble d'apprentissage.