Hello, I'm Thomas Robert

I’m a PhD student in Deep Learning at the LIP6 lab (Sorbonne University, Paris).

I work on deep learning for computer vision under the supervision of Matthieu Cord and Nicolas Thome. I will defend my PhD in autumn 2019.

This page is a quick overview of me. For more details, please see my resume.

A few things about me...

I love Computer Science

I started coding 10 years ago through web development and keep working on it has a hobby

I love Machine Learning

It's an incredible tool to solve complex problems

I love challenges

I am looking for innovative research & development projects using deep learning

I love learning and sharing

That's why I did a PhD and taught at my university

Research

Summary

I am working at Sorbonne University (previously Pierre and Marie Curie University) in Paris, at the LIP6 lab, in the MLIA team managed by Patrick Gallinari. My PhD research was done under the supervision of Matthieu Cord and Nicolas Thome. My thesis is focused on improving latent representations of images, in particular to improve classification by deep convolutional neural networks. To this end, we worked on regularizing latent representation to improve their invariance (SHADE). We then proposed a new way to structure the representations by using two complementary spaces. First, we proposed HybridNet that uses a first class-invariant latent space for classification and a second complementary latent space with instance related information, improving the classification results in the context of semi-supervised learning based on reconstruction. Then, we proposed DualDis that learns to disentangle two information domains, i.e. person identity and visual attributes (glasses, hair style, etc.). This is done with two latent spaces, one for each domain, that can then be manipulated to change the information represented and produce new images using a decoder.

Publications & research work

DualDis: Dual-Branch Disentangling with Adversarial Learning

T. Robert, N. Thome, M. Cord

Under review at Neural Information Processing Systems (NeurIPS) (2019)

Paper

HybridNet: Classification and Reconstruction Cooperation for Semi-Supervised Learning

T. Robert, N. Thome, M. Cord

European Conference on Computer Vision (ECCV) (2018)

Paper

SHADE: Information-Based Regularization for Deep Learning

M. Blot, T. Robert, N. Thome, M. Cord

IEEE International Conference on Image Processing (ICIP) (2018) Best paper award

Paper

M2CAI workflow challenge: classification of surgical operation steps based on endosopic videos

R. Cadene, T. Robert, N. Thome, M. Cord

M2CAI workshop @ MICCAI (2016) 2nd best model submitted to the challenge

Technical paper Slides

VISIIR: deep learning for recipe image classification

T. Robert, R. Cadene, N. Thome, M. Cord

(2016) Unpublised extension of published work

Project website & Demo Related publication

My main skills

Machine Learning & Data Science

  • Deep Learning
  • Data Science
  • PyTorch
  • TensorFlow
  • Keras
  • Matlab
  • Hadoop / Spark

Programming

  • Theoretical knowledge
  • Python
  • Java
  • Scala
  • C / C++
  • SVN / Git

Web

  • Theoretical knowledge
  • PHP
  • MySQL
  • JavaScript
  • HTML5
  • CSS3
Beginner / Proficient / Advanced or expert

See all my skills

Latest experiences

PhD in Deep Learning

LIP6 lab, Sorbonne University • Paris, France

2016-2019 Current

Supervised by Matthieu Cord and Nicolas Thome

  • Teaching work (64h / year): in charge of multiple courses, writing of convolutional networks practical sessions subjects
  • Submission to the workflow challenge of the M2CAI (MICCAI 2016) workshop with Remi Cadène : classification of surgical operation step based on endosopic videos. 2nd best model submitted.
  • SHADE method for regularization of deep neural networks with Michael Blot. Published at ICIP 2018 (best paper award).
  • Developpement of HybridNet, a semi-supervised learning model based on auto-encoders. Published at ECCV 2018.
  • Work on DualDis, a disentangling model for generation. Under review at NeurIPS 2019.

Double degree in Data Science

INSA Rouen & University of Rouen

2015

Master’s degree (“diplôme d’ingénieur”) in Data Science &
Master’s degree in Multimedia Information Processing System

Data Science research internship

Conviva • San Francisco Bay, California, USA

2015 6 months

Work on various research problems about video streaming quality optimization using data science

  • Prediction of the optimal video quality for a given session using Machine Learning and A/B testing
  • Formulation and resolution of an optimization problem to determine optimal decisions Conviva's products should take
  • Statistical study of bandwidth evolution in the context of Conviva's products
  • Technologies: Python, Hive, Hadoop MapReduce, Spark

Read more about my experiences