Jonas Gehring's Homepage

About

This page is a short listing of my academic, professional and voluntary work. Furthermore, it contains some programming-related stuff.

Data

Name: Jonas Gehring
Locations: Freiburg, Germany; Paris, France; Zürich, Switzerland
Work: Phd Student at Facebook AI Research and ETH Zürich
Research interests: Neural Networks, Machine Learning, Cognitive Systems, Algorithms
Programming interests: Parallelization, Optimization, Languages
Analogue interests: Reading, Guitar (Acoustic, Electric), Drums, Biking, Snowboarding
Linux flavour: Arch

Contact details:

E-mail address: jonas at this domain

Academia

In 2019, I started a PhD with Andreas Krause at ETH Zürich, and Nicolas Usunier and Gabriel Synnaeve at Facebook AI Research.

In 2013, I worked as a research assistant at the Interactive Systems Lab at Karlsruhe Institute of Technology on speech recognition and deep learning.

In 2012, I finished my Master's degree at the Karlsruhe Institute of Technology, focussing on Machine Learning, Cognitive Systems and Theoretical Computer Science. In my final thesis, I applied deep learning techniques to feature preprocessing in automatic speech recognition. With support from my advisors, Alex Waibel and Florian Metze, as well as from InterACT, I was able to perform my thesis at Carnegie Mellon University in Pittsburgh, PA.

In 2009, I finished my undergraduate studies in Computer Science at the University of Freiburg. I wrote my Bachelor's thesis, "Intelligente Objekterkennung für ein lernfähiges Carrerabahn-System" at Prof. Martin Riedmiller's Machine Learning Lab. It documents the development of an intelligent object tracking system for Prof. Riedmiller's NeuroRacer, a reinforcement learning slotcar system.

Publications

"RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning" Jonas Gehring, Kunhao Zheng, Jade Copet, Vegard Mella, Taco Cohen, Gabriel Synnaeve [arXiv]
"What Makes Large Language Models Reason in (Multi-Turn) Code Generation?" Kunhao Zheng, Juliette Decugis, Jonas Gehring, Taco Cohen, Benjamin Negrevergne, Gabriel Synnaeve [arXiv]
"The Larger the Better? Improved LLM Code-Generation via Budget Reallocation" Michael Hassid, Tal Remez, Jonas Gehring, Roy Schwartz, Yossi Adi (COLM 2024) [arXiv]
"Code Llama: Open Foundation Models for Code" Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, Gabriel Synnaeve [blog] [arXiv] [code, models]
"Leveraging Demonstrations with Latent Space Priors" Jonas Gehring, Deepak Gopinath, Jungdam Won, Andreas Krause, Gabriel Synnaeve, Nicolas Usunier (TMLR 03/2023) [website] [arXiv] [code]
"Hierarchical Skills for Efficient Exploration" Jonas Gehring, Gabriel Synnaeve, Andreas Krause, Nicolas Usunier (NeurIPS 2021) [website] [arXiv] [code]
"Forward Modeling for Partial Observation Strategy Games - A StarCraft Defogger" Gabriel Synnaeve, Zeming Lin, Jonas Gehring, Daniel Gant, Vasil Khalidov, Nicolas Carion, Nicolas Usunier (NeurIPS 2018) [arXiv]
"High-Level Strategy Selection under Partial Observability in StarCraft: Brood War" Jonas Gehring, Da Ju, Vegard Mella, Daniel Gant, Nicolas Usunier, Gabriel Synnaeve (RLPO@NeurIPS 2018) [arXiv]
"STARDATA: A StarCraft AI Research Dataset" Zeming Lin, Jonas Gehring, Vasil Khalidov, Gabriel Synnaeve (AIIDE 2017) [arXiv] [code]
"Convolutional Sequence to Sequence Learning" Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, Yann N. Dauphin (ICML 2017) [arXiv] [Lua code] [Python code]
"A Convolutional Encoder Model for Neural Machine Translation" Jonas Gehring, Michael Auli, David Grangier, Yann N. Dauphin (ACL 2017) [arXiv] [code]
"Models of Tone for Tonal and Non-Tonal Languages" Jonas Gehring, Kevin Kilgour, Florian Metze, Quoc Bao Nguyen, Van Huy Nguyen, Zaid A. W. Sheikh, Alex Waibel (ASRU 2013) [pdf]
"DNN Acoustic Modeling with Modular Multi-Lingual Feature Extraction Networks" Jonas Gehring, Quoc Bao Nguyen, Florian Metze, Alex Waibel (ASRU 2013) [pdf]
"Modular Combination of Deep Neural Networks for Acoustic Modeling" Jonas Gehring, Wonkyum Lee, Kevin Kilgour, Ian Lane, Yajie Miao, Alex Waibel (INTERSPEECH 2013) [pdf]
"Extracting Deep Bottleneck Features Using Stacked Auto-Encoders" Jonas Gehring, Yajie Miao, Florian Metze, Alex Waibel (ICASSP 2013) [pdf]

Work

Starting from 2016, I am a Research Engineer at Facebook AI Research (FAIR), initially working on neural machine translation and for some time now on reinforcement learning for games.

In 2013 I worked for Mobile Technologies, which offers speech-to-speech translation for smartphones via the Jibbigo app. In Fall 2013, we were acquired by Facebook and bootstrapped its speech recognition and machine translation efforts.

Throughout my studies, from 2006 to 2012, I've been working at Phönix PACS, a Freiburg-based company producing PACS software. I mainly worked on 3D reconstruction and visualization of medical image data as well as segmentation and analysis of anatomical structures.

In 2005 I did my alternative service at the Anesthesiological ICU of the Universitätsklinikum Freiburg.

Programming

Here are some of my personal projects that are publicly available as free software:

rsvndump, a remote Subversion repository dumping tool
pepper, a scriptable source code repository statistics and report tool
qmpwidget, a Qt widget for embedding MPlayer
hcxselect, a CSS selector engine for C++ working with htmlcxx
pgpry, an OpenPGP private key recovery tool using Pthreads
qICCreator, a simple IC drawing application (out of date)
cstune, a command-line audio player with a small number of features
Many programs for graphing calculators, including games as well as utility programs written in TI-BASIC and C.

Numerous other small things can be found at my GitHub page.

On rare occasions, I stick with existing software instead of re-inventing the wheel. Here are some patches I've submitted:

Support for custom URL scheme handlers for the Arora web browser
Custom button actions for menus and icons for PieDock, a pie-shaped task bar and application launcher for X11
Support for HTTP "Referer" field for the VLC media player
Interactive desktop switching in Openbox. The patch wasn't applied, but fortunately the behaviour is the same now.

Articles

"Meaning in Large Language Models: Form vs Function" lays out philosophical arguments on why or why not LLMs can develop understanding from text alone. Term paper for 263-5353-00L .

"Towards Natural Natural Language Processing" discusses textless NLP work from a structuralist perspective. Term paper for 263-5353-00L .

Slides from a talk about Haskell arrows (in German)
Beginner's tutorial for C programming on TI 68k calculators (in German). The English translation of the first parts can be found here.
Small article on gray-scale graphics on TI 68k calculators (in German)
How to setup Laudio with lighttpd describes the setup of a web-based audio player using the lighttpd web server.

Activities

I used to be an active member and administrator at the largest German message board about graphing calculators from Texas Instruments, mobifiles.de. During that time, Martin Huber from mobifiles.de and I founded boolsoft.org, an open-source programming group for TI 68k calculators. However, I've been rather inactive at calculator programming for the last couple of years.