PhD in Computer Science
Machine Learning & Language Understanding

Who am I ?

I'm a young computer science enthusiast.
Always looking forward to improve state-of-the-art, and always looking to learn from the past.

About

I'm a PhD in computer science with an expertise on complex and hypercomplex neural networks applied to speech recognition, image processing and natural language understanding. I am also particularly attached to the concept of science for the social good. For seven years now (since the middle of the bachelor), I have been teaching at the Centre d'Enseignement et de Recherche en Informatique (CERI). Interested in computer science since my childhood, i just decided to follow the always learn and make more about computers way of life.

Education

  • PhD in computer science (CIFRE) - 2019
  • Thesis: Quaternion neural networks
  • Advisors: Georges Linarès, Mohamed Morchid
  • Reviewers: Thierry Artières, Allexandre Allauzen
  • Committee: Yoshua Bengio, Benjamin Lecouteux, Xavier Bost, Nathalie Camelin
  • Avignon Université, France & Orkis, Aix-en-provence, France
  • Master Research in computer science - 2016
  • Thesis: Quaternions and deep neural networks
  • Avignon Université, France
  • Bachelor in computer science - 2014
  • Avignon Université, France

Research experiences

  • CIFRE PhD Student - 2017 / 2019
  • Thesis: Quaternion neural networks
  • Advisors: Georges Linarès, Mohamed Morchid

  • PhD Student in machine learning working on quaternion-valued neural networks with applications to speech recognition, spoken language understanding and image processing. Research engineer at ORKIS, working on new techniques on ASR and documents management with deep learning. Teacher assistant for Master and Bachelor students. Given courses mainly focused on machine learning, programming and algorithms.

  • Avignon Université, France & Orkis, Aix-en-provence, France
  • Montréal Institute for Learning Algorithms (MILA) - Jan. 2018 / Apr. 2018
  • Advisor: Yoshua Bengio

  • Collaboration and release of the Pytorch-Kaldi toolkit. Research on quaternion convolutional neural networks for end-to-end automatic speech recognition. Introduction of quaternion-valued recurrent neural networks to speech recognition.

  • Université de Montréal, Québec, Canada
  • Research Engineer - Sep. 2016 / Mar. 2017
  • Advisor: Mohamed Morchid

  • Hyper-complex neural networks implementation on GPU (CUDA). Research on Spoken Language Understanding (Pattern Recognition).

  • Avignon Université, France

Research Interests

A word about the thesis - Quaternion Neural Networks


Supervised by Georges Linarès and Mohamed Morchid - University of Avignon - LIA


Machine Learning (ML) techniques have allowed a great performance improvement of different challenging tasks (Spoken Language Understanding, Image Processing, Motion Management, Robotic ...). Among these methods, Neural Networks (NN), recently received a great interest from researchers due to their representation capability of complex internal structures in a low dimensional subspace. However, NNs employ representations that are based on a basic level (because of their 1-dimensionality). Therefore, these basic representations reveal little in way of the internal relations that characterizes the entity that is processed by only considering each feature composing this entity as individual components, ignoring relations between them. My thesis propose to remedy this weakness by extending the hyper-complex numbers, also called quaternions, to neural networks called QNNs for Quaternion Neural Networks. Indeed, in a real-valued neural network, multidimensional features require to be reduced to a one dimensional vector before the learning process, while an appropriate solution is to process a multidimensional input as a single homogeneous entity. All these drawbacks can be alleviated thanks to the fourth dimensionality of quaternions and the hamilton product.

This thesis first proposes an integration of quaternions to state-of-the-art models such as : deep, recurrents, convolutionals, and adversarials neural networks. This integration rely on adapted and well-suited algorithms to fit the quaternion algebra specificities, and on a GPU implementation of quaternion neural networks. Then, a new type of neural network expressly designed for hyper-complex numbers wil be investigated, including a specific learning process and a well-adapted structure. This method has already reached promissing results that will be reported on my ResearchGate homepage and below.

AI for Social Good

One of the first question that I asked myself before starting to work on ML was "Why are we doing AI ?". Going to the NeurIPS 2018 conference held in Montréal, I have been shocked to realize that from the hundred of companies presententing their activities, less than 10% were in fact interested in making the world a better and fairer place. While I understand that AI is a formidable tool for growth and productivity, I trully can't get why there are almost no ressources involved in the application of AI to fundamental domains for the Human being such as health, climat change, environment, new energies, or even social fairness. At the dawn of the 6th mass extinction, we are more concerned by the prediction accuracies on insurance markets or by the quality of "fake" portraits, than by finding solutions to our actual threats trough the formidable power of AI. Consequently, to the question "Why are we doing AI ?", my answer is that I will dedicate my post thesis research to AI for social good.

Feel free to contact me for further informations !

Contributions

SpeechBrain: A Pytorch-based Speech Toolkit.


Invited Talks

University of Oxford (UK): "Quaternion neural networks", July 9th (2019), Computer Science Department.
Samsung AI Cambridge (UK): "Quaternion neural networks", July 10th (2019).

Scientific Responsabilities

Reviewer for international journals: IEEE Transactions on Neural Networks and Learning Systems | IEEE International Journal of Wavelets, Multiresolution and Information Processing.
Reviewer for International Conferences: NeurIPS, ICLR, ICASSP.

Highlights of publications

Quaternion-valued neural networks suffer from the fact of being new and mainly unexplored, therefore I consider as an achievement the first oral presentation of one of our work entitled "Quaternion Convolutional Neural Network for End-to-End Speech Recognition" at the major INTERSPEECH 2018 conference. More recently, this interest from the community in quaternion-valued neural networks have been verified with the acceptance of our work on " Quaternion Recurrent Neural Network " at the ICLR 2019 conference. A good summary of our work on speech recognition with quaternion numbers called " Speech Recognition with Quaternion Neural Networks" has also been accepted at the IRASL workshop of the NeurIPS 2018 conference. The acceptances of our ideas in such famous and competitive venues validate the assumption that the research community is interested in such innotivates, and unusual approaches.

As described in the Collaborations section of this website, I spent 4 months in Montréal at the Montréal Institute of Learning Algorithms (MILA) working on Quaternion-valued Neural Networks for Speech Recognition. Many fruitful collaborations started with this internship, and a remarkable one is the development of "The Pytorch-Kaldi Speech Recognition Toolkit" with Mirco Ravanelli and Yoshua Bengio. This toolkit will provide to the community a very important bridge between two widely used framework for both ML and Speech Recognition, to enable fast and reliable iterations of the research in these fields. The activty of the GitHub repository (GitHub) of the project valide this keen interest.

International Journals

International conferences

  • Quaternion Recurrent Neural Networks
  • Titouan Parcollet · Mirco Ravanelli · Mohamed Morchid · Georges Linarès · Chiheb Trabelsi · Renato De Mori · Yoshua Bengio
  • ICLR 2019
  • May 6-9 2019, New Orleans, (USA)

National conferences

Collaborations

SpeechBrain - A PyTorch-based Speech Toolkit

The SpeechBrain project aims to develop an open-source and all-in-one toolkit based on PyTorch. The goal is to develop a single, flexible, and user-friendly toolkit that can be used to easily develop state-of-the-art speech systems for speech recognition (both end-to-end and HMM-DNN), speaker recognition, speech separation, multi-microphone signal processing (e.g, beamforming), self-supervised learning, and many others. The project is funded by the Montréal Institute for Learning Algorithms (MILA), Samsung, NVIDIA, and Dolby. SpeechBrain will also benefit from the collaboration and expertise of other partners such as the University of Avignon (LIA), Facebook, IBM Research, and Fluent.ai Inc. I'm co-leading SpeechBrain with Dr. Ravanelli Mirco, currently a post-doctoral researcher at MILA.

Thanks to our sponsors, we are hiring talented interns (3-6 months internships) that will work at Mila (Montréal) with the core development team. The ideal candidate is a PhD student with a strong experience in both PyTorch and speech technologies. Send us your CV if you are interested in this opportunity!

Have a look at: SpeechBrain!

ORKIS - PhD Fellowship

Orkis is a a french company based on Aix-en-provence, France. They are specialised in assets and data managements with well-known customers such as L'Oréal, Dassault, or the governement of France. The collaboration between the LIA (and thus the University of Avignon) started with a previous PhD fellowship which has now ended. In order to enhance their tools, they decided to pursue this collaboration with a new PhD student, me. My job involves many taks, from automatic indexation on speach recognition to image processing and machine learning. For more informations about the company, feel free to visit the Orkis website.

Montréal Institute for Learning Algorithms (MILA), Montréal - Internship

When it comes to ML and deep learning, the MILA is known to shine. I have been invited for a 4 months internship in this laboratory to work on Quaternion-valued Neural Networks under the supervision of Prof. Yoshua Bengio. These 4 months have been four of the most intensive months of my life, with three papers and many different projects started. This period also was the most collaborative period of my thesis, since sharing and collaborations are at the heart of the MILA research. The first focus was on the fusion of an already existant end-to-end model (MILA) based on convolutional neural networks, with quaternion numbers (LIA) for the first exploration of a fully quaternion-valued convolutional neural network for three dimensional accoustic quaternions processing. Then, alongside with the development of the PyTorch-Kaldi toolkit, we investigated and proposed quaternion quaternion recurrent neural networks with a detailled backpropagation trough time (BPTT) algorithm on the quaternionic domained. Fortunatly, we are still collaborating with the MILA on ongoing projects including new quaternion-valued end-to-end models based on SyncNet networks, or even multichannel speech recognition.

Teaching

All the tutorials and practical work classes are dispensed in addition to my standard research job

Year 2018/2019

TeachingTutorialsPractical WorkTotal
Computer Science Basics10h10h
Object Oriented Programming C++14h14h28h
Tools for Machine Learning12h12h
Total50h

Year 2017/2018

TeachingTutorialsPractical WorkTotal
Advanced programming and projects (B.Sc.)12h19h31h
Total31h

Year 2016/2017

TeachingTutorialsPractical WorkTotal
Advanced programming and projects (B.Sc.)12h18h30h
Agorithms and programming (B.Sc.)12h18h30h
Total60h

Year 2015/2016

TeachingTutorialsPractical WorkTotal
C2I Certification (B.Sc.)60h
Total60h

Year 2013/2014

TeachingTutorialsPractical WorkTotal
C2I Certification (B.Sc.)60h
Total60h

Contact

Lab address

LIA - CERI
339 Chemin des Meinajariès
84911 Avignon, France

Mail & Phone

titouan.parcollet@alumni.univ-avignon.fr
+33 490 843 577