Faites connaître cette offre !
Reference : UMR6004-VINLOS-003
Workplace : NANTES
Date of publication : Thursday, February 18, 2021
Type of Contract : FTC Scientist
Contract Period : 12 months
Expected date of employment : 1 August 2021
Proportion of work : Full time
Remuneration : 2648 € monthly gross (maximum of 2 years of research experience after the PhD)
Desired level of education : PhD
Experience required : 1 to 4 years
The TrAcS project stands for “Trainable Acoustics Sensors”. Its main goal is to design new sensors for audio event detection and classification. These sensors will comprise trainable modules which will be able to identify the most relevant sources in an acoustic stream while discarding the influence of background noise.
In this context, the first mission of the postdoctoral researcher will be to invent new models in statistical machine learning with audio signal processing. The originality of these models will reside in their “explainability”, i.e., their ability to produce a visualization of acoustic data that is humanly interpretable.
The main application scope for this study the condition monitoring of industrial machines. An extension of the proposed method to musical or biomedical could be envisioned as a secondary goal.
The TrAcS project faces two main challenges: the modeling of infrequent events and the statistical generalization to multiple sensing domains. To address these challenges, the postdoctoral researcher will be encouraged to consider alternative paradigms to supervised learning; specifically, self-supervised and semi-supervised learning.
In terms of fundamental research, the TrAcS project will contribute to the advancement of differentiable computing. This framework integrates the building blocks of deep learning within a global computational environnement, involving trainable parameters as well as non-trainable functions for feature extraction, data augmentation, physical simulation, control, or probabilistic inference.
In particular, the TrAcS project will apply time-frequency scattering as a mid-level audio representation for self-supervised and semi-supervised learning. The postdoctoral researcher will contribute to Kymatio, an open-source software library for differentiable time-frequency scattering with GPU acceleration and PyTorch/TensorFlow interoperability.
- scientific research in information science and computer science
- co-publication in scientific journals, for example: IEEE TPAMI, JASA, Applied Acoustics
- contribution to the development and maintenance of the Kymatio open-source software library
- presentation to scientific conferences, for example: IEEE ICASSP, DCASE, ISMIR, GRETSI
- participation to various team meetings and lab meetings
The candidate should hold a PhD in Computer Science or Applied Mathematics
- Mathematical proficiency in time-frequency analysis, specifically wavelet theory
- Methodological proficiency in deep learning, specifically convolutional neural networks
- Scientific computing in the Python language, specifically with libraries Librosa, NumPy, Scikit-Learn, and PyTorch
- Clear scientific expression in the English language
Experience with the Kymatio library is appreciated but not required.
Skills in acoustics or music are appreciated but not required.
Knowledge of the French language is not required.
The candidate will work at the Nantes Laboratory of Digital Sciences (LS2N). In particular, they will belong to the team “Signals, Images, and Sounds” (SIMS), which is integrated to the department “Signals, Images, Ergonomy, and Languages” (SIEL). Furthermore, the candidate will contribute to the activities of the “Industry 4.0” special interest group at LS2N.
For more details, please: https://www.ls2n.fr/presentation/
Constraints and risks
This position does not involve any particular professional constraints or risks.
We talk about it on Twitter!