Théo Delemazure - Projects

Projects

Representing Uncertain Databases as World-Set Decomposition

Data Wrangling

This work is based on the paper 10^10^6 Worlds and Beyond : Efficient Representation and Processing of Incomplete Information by Lyublena Antova, Christoph Koch, and Dan Olteanu, in which the authors present an efficient way to represent incomplete information into databases, using World-Set Decompositions (WSD). This model being a strong representation of incomplete databases for every query language, the result of every query on WSD can be expressed as a WSD. In this project, I implemented algorithms to answer relational algebra queries on WSD. Finally, I tested my algorithms on a dataset with real incompleteness taken from the Paris Open Data website and containing informations about all trees in Paris.

Robust Adversarial Networks for Image Detection

Data Science Projects

In this third project of the Data Science Projects course, our goal was to build neural networks for image classification which are robust to classic adversarial attacks, like PGD and FGSM presented in Explaining and Harnessing Adversarial Examples by Goodfellow and Al. We first build a basic neural network and tested the different attacks on it. Then, we tried to build the most robust neural network to these attacks. We got very good results when training a network with the contrastive loss on adversarial examples.

Unsupervised Word Translation

Data Science Projects

In this second project of the Data Science Projects course, our goal was to use word embedding to build an unsupervised world translator, like described in the paper World Translation Without Parallel Data by Conneau and Al (Facebook AI Research). We first implemented supervised word translation using a dictionary. Then, we tried the unsupervised method on different pairs of languages using generative adversarial networks.

Online-Learning for Recommendation Systems

Data Science Projects

In this first project of the Data Science Projects course, our goal was to familiarize ourselves with data science through a small project. We chose the project about reinforcement learning methods like linearUCB for recommendation systems. We based our work on the paper A Contextual-Bandit Approach to Personalized News Article Recommendation (Yahoo Labs) in which they use contextual bandit for article recommendation. Once we did that, we tried to implement an online-learning framework in which new movies and new users can come into the system at any time.

Electric Motor Temperature Prediction

Optimization for Machine Learning

For this project, the goal was to apply notions and algorithms covered by the course to a dataset of our choice on a classification or a regression task. I choose the Kaggle dataset Electric Motor Temperature for a regression task. I used classic methods like Ridge and Lasso and small neural networks entirely hand coded. I also tried to add a recursive unit to obtain better results.

Paper presentations

On the Representation and Querying of Sets of Possible Worlds

Data Wrangling (M2)
2 weeks
1

Report

Rationilizing Neural Predictions

Natural Language Processing (M2)
1 week
3

Slides

Pinpointing in the Description Logic EL

Knowledge Graphs, Description Logics and Reasoning on Data (M2)
1 week
1

Slides

Object Detection with Discriminatively Trained Part Based Models

Artificial Vision (M1)
1 week
1

Slides

Report

SIMBICON: Simple Biped Locomotion Control

Robotic (M1)
1 week
1

Slides

Report

Projects

Paper presentations

On the Representation and Querying of Sets of Possible Worlds link

Rationilizing Neural Predictions link

Pinpointing in the Description Logic EL link

Object Detection with Discriminatively Trained Part Based Models link

SIMBICON: Simple Biped Locomotion Control link

On the Representation and Querying of Sets of Possible Worlds

Rationilizing Neural Predictions

Pinpointing in the Description Logic EL

Object Detection with Discriminatively Trained Part Based Models

SIMBICON: Simple Biped Locomotion Control