ENGLISH VERSION:
Time-predictable Machine Learning in TensorFlow/MLIR ==================================================== Contact: [email protected]
To apply: https://www.irt-systemx.fr/recrutement/ia-temporellement-predictible-en-tensorflow-mlir/
Partnership ----------- This work will take place in the framework of the Confiance.AI Grand Challenge [7] proposed and financed by the French government. The safe use of AI-based technologies is key in supporting engineering, industrial production, and the design of novel products and services. Thus, facilitating the use of AI in critical systems is one of the major objectives of Confiance.AI.
The program is coordinated by the Institut de Recherche Technologique SystemX (IRT SystemX, https://www.irt-systemx.fr/). Situated in the scientific campus of Paris-Saclay, IRT SystemX is an interdisciplinary thematic institute that develops economic sectors related to its field through a balanced strategic public/private partnership. For this, it manages research programs coupled with technology platforms, conducts research and development projects at the international level, contributes to the engineering of initial and continuous trainings (qualifying professional training and/or degree delivering) and ensures the exploitation of the results.
Inria (https://www.inria.fr/en)is the French national research institute for digital science and technology. World-class research, technological innovation and entrepreneurial risk are its DNA. In 200 project teams, most of which are shared with major research universities, more than 3,500 researchers and engineers explore new paths, often in an interdisciplinary manner and in collaboration with industrial partners to meet ambitious challenges. As a technological institute, Inria supports the diversity of innovation pathways: from open source software publishing to the creation of technological startups (Deeptech).
This PhD will be supervised by Dumitru POTOP BUTUCARU, Inria researcher of the Kairos team (https://team.inria.fr/kairos/). The work will take place at IRT SystemX, in Palaiseau, with regular visits to the Inria Paris research center. Inside the IRT SystemX, the PhD candidate will be part of the "Numerical infrastructures" department supervised by Makhlouf HADJI.
Significant interactions are also anticipated with industrial use case providers of Confiance.AI (e.g. Thales, Safran, Valéo...), with other academic partners of the programme, with the RITS Inria team[5,6] to take advantage of its know-how in applying ML technology to autonomous driving, and with the Open-Source TensorFlow/MLIR community.
Scientific context and objective -------------------------------- The know-how required for the specification and implementation of the high-performance, real-time embedded systems of the future (autonomous transportation, predictive maintenance and control...) is today split among two scientific and engineering communities:
- The "High-Performace Computing" (HPC) community has produced the Machine Learning (ML) frameworks (TensorFlow, PyTorch, N2D2…) which allow the efficient implementation of ML applications on virtually every type of architecture (Single- and Multi-Core processors, GPUs, TPUs, other accelerators). These frameworks provide optimized implementations of a large variety of advanced compilation algorithms meant to take advantage of the data parallelism of ML applications (tiling, vectorization, buffer allocation, parallelization...). They also provide full compilation flows obtained by pipelining these algorithms. One such framework is MLIR (https://mlir.llvm.org), part of the TensorFlow ecosystem and distributed along with the LLVM compiler suite (https://llvm.org).
- The "Real-Time Embedded" (RTE)community has produced languages, methods, and tools allowing the specification and the implementation of systems that must not only be functionally correct, but also satisfy stringent non-functional requirements, and in particular real-time requirements such as periods, I/O latencies, or throughputs. To provide such guarantees, applications, target architectures and implementation methods must provide support for precise worst-case timing and interference analysis. This explains the use of concurrent and deterministic domain specific languages, such as Lustre/Scade or Simulink, which allow the natural representation of task parallel applications. In this context, the Kairos team has produced the Lopht real-time parallelization tool [2,3] which reached TRL4 on large-scale critical avionics applications specified in Lustre, reaching very good parallelization speed-ups.
Our objective is to allow the joint application of methods of both fields, thus enabling the implementation of HPC real-time systems. The Kairos team has already identified [1] strong formal links between the Lustre and MLIR languages, which allow today the joint specification of both HPC data-parallel and reactive task-parallel aspects of an application, and their joint compilation to executable code.
The objective of this PhD is to extend the work of [1] by taking into account non-functional aspects, and in particular real-time aspects. More precisely, we need to:
- Extend the TensorFlow/MLIR framework with non-functional real-time annotations. These annotations must allow the specification of real-time properties and requirements in a way that is compatible with the SSA-based structure of the MLIR intermediate representation.
- Take into account these annotations during the compilation process to provide real-time guarantees while still preserving the code generation efficiency specific to the ML frameworks. To do this, we will build upon the real-time parallelization work of [2,3].
Skills ----------- We seek a PhD candidate with an excellent level in computer science. Solid knowledge of at least one of the following fields is considered a plus: Machine Learning, Real-time systems, Distributed systems, Programming languages, Compilation, Parallelization, High-Performance Computing (HPC), Semantics, Formal Methods, Computer architecture.
I am advertising a postdoc position to work in Bordeaux, France, on the "Games for Synthesis" project (G4S).
What is the science about: G4S is a 3-year project to work on the theory and practice of both controller synthesis and program synthesis, using games and reinforcement learning. Feel free to ask me for details!
The non-scientific part. The position is for two years and can start as early as October 2021 (flexible). The salary is roughly 2100€ per month after tax (includes health insurance), which is comfortable to live in Bordeaux, a beautiful city close to the ocean and the Pyrenees Mountains (also, have you heard of the wine?). There are no teaching duties included, but there are many interesting teaching opportunities.
Please contact me if you have any questions ([email protected]), and forward this announcement to whoever may be interested.
Dear Colleagues, Please find below the description of a PhD position in “Joint embedded speech separation, diarization and recognition for the automatic generation of meeting minutes”.
Starting date: October 01, 2021
Deadline for Applications: July 16, 2021
All details are available at: https://recrutement.inria.fr/public/classic/en/offres/2021-03757
Keywords: diarization, speech separation, robust automatic speech recognition, transfer learning, deep learning
Context Founded in 2015 and awarded two CES Innovation Awards, Vivoka (https://vivoka.com/en/) has created and sells the Voice Development Kit (VDK), the very first solution allowing a company to design a voice interface in a simple, autonomous and quick way. Moreover, this interface is embedded: it can be deployed on devices without an Internet connection and fully preserves privacy. Accelerated by the COVID-19 health crisis and the need for "no-touch" interfaces, Vivoka is now optimizing this technology by developing its own speech and language processing solutions able to compete with the most efficient current technologies. This research project, which involves the entire Vivoka R&D team, is carried out within the framework of a long lasting partnership with Inria's Multispeech team (https://team.inria.fr/multispeech/).
The hired PhD student will share his/her time between Vivoka's R&D team and Inria's Multispeech team. He/she will benefit from the startup spirit of Vivoka, where he/she will interact with other PhD students, interns and researchers hired as part of the partnership and the engineers responsible for integrating their results into the VDK. He/she will also benefit from the skills of the Multispeech team, the largest research team in the field of speech processing in France, and the overall Inria environment.
Assignment Conversational Automatic Speech Recognition (ASR) has seen tremendous progress over the past decade, with a word error rate now similar to that of humans for a single speaker speaking close to the microphone [1]. As soon as the speaker moves away from the microphone, the error rate increases due to reverberation, ambient noise, and overlapping speech from other speakers. The automatic generation of meeting minutes thus involves solving a set of tasks: i) segmenting the signal according to the number of speakers and who is speaking at each time (diarisation) [2], ii) separating overlapping speech signals [3] and enhancing them with respect to ambient noise and reverberation, iii) ensuring the robustness of ASR with respect to diarization errors and signal distortions introduced by separation and enhancement [4], and iv) removing disfluencies from the word-for-word transcription in order to obtain readable minutes.
The objective of this PhD is to design a system which can jointly address the first three tasks given a single-channel or a multichannel signal and which can be embedded in a device with limited computing power (for example a mobile phone), while being able to compete with current Cloud-based technologies.
Skills Master 2 in computer science, data science or signal processing. Programming experience in Python and in a deep learning framework.
Previous experience in the field of speech processing or computational footprint reduction is a plus.
Instructions for applying Application deadline: July 16, 2021
Submit your complete application data online at https://recrutement.inria.fr/public/classic/en/offres/2021-03757 and send a copy to [email protected]
Applications will be considered on the fly. It is therefore advisable to apply as soon as possible.
Dear Colleagues, Please find below the description of a PhD position in “Multi-factor Data Augmentation and Transfer Learning for Embedded Automatic Speech Recognition”.
Starting date: October 01, 2021
Deadline for Applications: July 16, 2021
All details are available at: https://recrutement.inria.fr/public/classic/en/offres/2021-03756
Keywords: automatic speech recognition, speech synthesis, voice conversion, transfer learning, deep learning
Context Founded in 2015 and awarded two CES Innovation Awards, Vivoka (https://vivoka.com/en/) has created and sells the Voice Development Kit (VDK), the very first solution allowing a company to design a voice interface in a simple, autonomous and quick way. Moreover, this interface is embedded: it can be deployed on devices without an Internet connection and fully preserves privacy. Accelerated by the COVID-19 health crisis and the need for "no-touch" interfaces, Vivoka is now optimizing this technology by developing its own speech and language processing solutions able to compete with the most efficient current technologies. This research project, which involves the entire Vivoka R&D team, is carried out within the framework of a long lasting partnership with Inria's Multispeech team (https://team.inria.fr/multispeech/).
The hired PhD student will share his/her time between Vivoka's R&D team and Inria's Multispeech team. He/she will benefit from the startup spirit of Vivoka, where he/she will interact with other PhD students, interns and researchers hired as part of the partnership and the engineers responsible for integrating their results into the VDK. He/she will also benefit from the skills of the Multispeech team, the largest research team in the field of speech processing in France, and the overall Inria environment.
Assignment Conversational automatic speech recognition (ASR) has seen tremendous progress over the past decade, with a word error rate now similar to that of humans [1]. This is explained by the maturity of deep neural networks but above all by the increase in the size of the training corpora available as open or proprietary data. These corpora must be annotated, that is to say transcribed manually in textual form. The cost of this operation means that the amount of proprietary data collected and annotated by large industry players, in the order of 10,000 hours or more for languages such as French, is inaccessible to SMEs. It is also reflected in the fact that current business solutions are only available for about 100 languages out of the 7,000 languages spoken in the world.
The objective of this PhD is to design an embedded ASR system capable of competing with current solutions while being trained on non-proprietary data only, for example open data from the Mozilla Common Voice initiative [2], i.e., less than 1,000 h annotated data in French or less than 100h for less-resourced languages.
Skills Master 2 in computer science or data science.
Programming experience in Python and in a deep learning framework.
Previous experience in the field of speech processing or computational footprint reduction is a plus.
Instructions for applying Application deadline: July 16, 2021
Submit your complete application data online at https://recrutement.inria.fr/public/classic/en/offres/2021-03756 and send a copy to [email protected]
Applications will be considered on the fly. It is therefore advisable to apply as soon as possible.
Dear All, On behalf of *Prof. Mauro Conti*, I'd like to inform you that the *SPRITZ Security and Privacy Research Group* @ University of Padua (Italy), is looking for *Research Fellows/PostDoc* in the field of Cybersecurity.
Two positions are *now open*:
- Secure Future Internet (deadline Jul 22nd, 2021, 1 pm CEST) https://www.math.unipd.it/news/bando-n-14-2021-per-n-2-assegni-di-ricerca-transizione-sicura-verso-linternet-del-futuro-securing-the-transition-toward-the-future-internet/
If interested, please apply now, since *deadline is close*! For more info, please drop an email to [email protected]. Best, Elena
Info about the SPRITZ Group: https://spritz.math.unipd.it/ Info about University of Padua: https://www.unipd.it/en/
Good morning, We are proposing a post-doc (announcement below), you could distribute it to interested people. Thanks in advance.
Post-Doc: Deep learning enhanced by computer graphics for video analysis Subject description
The CLIRIS company in collaboration with the Multidisciplinary Laboratory for Research in System Engineering, Mechanics and Energetics (PRISME) are looking for a graduate doctor during the academic years 2019-2020 and 2020-2021 for a post-doc.The CLIRIS company offers distribution networks solutions for controlling and optimizing their physical networks, comparable to those available to them for their merchant sites, based on image processing and deep learning algorithms.
For deep learning methods, the size of a database is an essential part of a training pipeline. However, in the field of object detection, public databases are often limited by a fixed context. For example, all the existing annotated data processes the images taken from the front, whereas in a real situation, the scenes are more complex.
In this context, we want to offer a system for generating high-volume synthetic images while providing a variety of content. The aim is to automatically provide videos of moving objects on still scenes. Several methods based on GAN [1] and VAE [2,3] have been used for the representation of synthetic images. However, these methods rely solely on CNNs, which makes capturing information interactions in a high-resolution image difficult to achieve. In [4], the authors combined CNN to reduce computational cost and Transformer [5] to include information merging from several convolutional responses at the same time. In this context, the first objective of this work is to create a system for generating image sequences based on three stages:
1. Generation of moving skeletons based on pre-trained models to generate objects guided by skeletons
2. Design a set of background scenes based on the semantic information
3. Combine the two steps to produce a variety of image sequences including moving objects in a scene
The second objective is to allow a diversification of the situations and bring a certain generalization to the learning algorithms.
Remuneration € 35-40k Gross
Candidate profile Training : Doctorate degree in France during the university years in 2019-2020 and 2020-2021 ( compulsory )
Required : Experience in image processing and deep learning. Mastery of python
Mastery of the machine learning framework (pytorch, tensorflow / keras ...)
Application procedure Send the candidate's detailed CV to the contacts below, motivation letter explaining the candidate's interest in the post-doc subject
Contacts : [email protected] , [email protected] , [email protected]