Offres de Post doctorat
Postdoctoral Proposal in Artificial Intelli-gence and Machine Learning
Postdoctoral Proposal in Artificial Intelli-gence and Machine Learning
Runtime Veriﬁcation for Critical Machine Learning Applications
Net salary: Negotiable with a minimum of 2600e per month with some teaching (20 hours per year on average)
Duration: one year (renewable once)
In the last decade, the application of Machine Learning (ML) has encountered an increasing interest in various applicative domains, especially for a wide panel of complex tasks (e.g. image-based pedestrian detection) classically performed by human operators (e.g. the driver). In ML, the objective is to synthesize an intended function (e.g., detect a pedestrian on an image) through a set of examples (images of road). The massive usage of such techniques has demonstrated its eﬀectiveness way beyond other classical methods. Obviously, the designers of critical systems would like to beneﬁt from the eﬀectiveness of ML-based models mainly for complex image processing and model reduction. But, above eﬀectiveness, the designers of critical systems must demonstrate that the obtained models are reasonably safe. Providing elements demonstrating the safety of a system is a classical issue addressed by various techniques tailored to the nature of the system, and covered by many safety standards (DO178C in aeronautics, ISO26262 in automative, IEC61508 in electronics, etc.). Nevertheless, the speciﬁcities of ML-based software disclose new safety threats that are not addressed by classical techniques. Despite very good results during training and testing of a ML software, it is not possible to provide suﬃcient guarantees that the training data set would be suﬃcient for expected real life situations, that during operational life the system may not face adversary situations (situation slightly diﬀerent from training ones, but which lead to a complete diﬀerent result of the ML, also called adversaries attacks), or that the distribution of situations may be diﬀerent from the ones during training (distribution shift). All these threats are a major brake to ML deployment in safety critical applications.
Most of works are focusing on the training data quality in order to increase robustness of the ML algorithm. However, to avoid overﬁtting, it is accepted that developing the dataset is limited, and a promising approach is to monitor the system at runtime, during operational life, in order to keep the system in a safe state, despite errors of the ML. A ﬁrst approach inspired from fault tolerance and close to safety monitoring , is to adapt the simplex architecture to the monitoring of a neural controller, using a decision block able to detect an error and to switch from the neural controller to a high assurance controller (but less performant) . Some works as  may be used to monitor the distance of input/ouput distribution during the exploitation vs the one observed during the learning phase and raise an alert when a “signiﬁcant gap” is observed. Other works like , dedicated to Neural Networks, propose to collect neuron activation patterns during the learning phase and, thanks to an online monitoring, detect the occurrence of an unknown activation pattern that may indicate an erroneous prediction.
All these works are ongoing, with very preliminary results, and actually no safety model is integrated in such proposals. We propose in this post-doc to specify, implement and verify a new runtime veriﬁcation approach for ML, an adversarial runtime monitor. This approach is based on adversaries generated at runtime, and used to assess if the ML maybe fooled in an unsafe state. This might lead the monitor to detect if the ML is in a potential unsafe erroneous state, or in a potential erroneous state but safe. Once such a monitor would be designed, we also plan to use formal methods (veriﬁcation) to prove the correctness of the monitor. This work will be applied to a case study, a ML software for drone collision avoidance studied and deployed in the context of the Delta project1.
hdieh Abbasi, Arezoo Rajabi, Azadeh Mozafari, Rakesh B. Bobba, and Christian Gagn´e. Controlling over-generalization and its eﬀect on adversarial examples generation and detection. ArXiv e-prints, 1808.08282, 8 2018.
 Chih-Hong Cheng, Georg N¨uhrenberg, and Hirotoshi Yasuoka. Runtime monitoring neuron activation pat-terns. CoRR, abs/1809.06573, 2018.
 Mathilde Machin, J´er´emie Guiochet, H´el`ene Waeselynck, Jean-Paul Blanquart, Matthieu Roy, and Lola Masson. SMOF - A Safety MOnitoring Framework for Autonomous Systems. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 48(5):702–715, May 2018.
 Dung Phan, Nicola Paoletti, Radu Grosu, Nils Jansen, Scott A. Smolka, and Scott D. Stoller. Neural simplex architecture. ArXiv e-prints, abs/1908.00528, 2019.
Formal applications should include detailed CV, a motivation letter and transcripts of bachelors’ degree. Samples of published research by the candidate and reference letters will be a plus. Applications should be sent by email to: advisor email
More information about ANITI: https://aniti.univ-toulouse.fr/
This position is linked to a research project funded by the civil aviation administration in France, coordinated by the LATECOERE company located in Toulouse. This project aims at developing an assistance system for the control of an aircraft in airport areas, provided overall for bad visibility conditions (night, fog, rain…); multi-spectral images are acquired, fused and analyzed in order to create a view both synthetic and realistic of the environment where the aircraft is navigating. This view, augmented by symbolic and textual information, is sent to a monitor in the cockpit, in order to inform the pilot about all potential risk when he controls the aircraft from the parking area until the runway, and reciprocally.
The applicant will be integrated in the RAP research team (Robotics, Action and Perception) in the Robotics department of LAAS-CNRS.
The applicant will participate to the design of the architecture of the heterogeneous system (FPGA, CPU, DSP…) with respect to a functional specification and to critical real-time constraints. An SME partner of the project will develop the hardware system.
The applicant will have also to understand the image processing algorithms developed in the prototyping step, so that he could propose optimizations required for their implementation on the embedded architecture satisfying the real time constraints.
He will be responsible for the technical management of the project (doc, source code…), for the deliverables and for the technological transfer respecting the rules defined in the project.
With the involved permanent researchers of the RAP team, he will supervise the work of two PhD students or engineers funded by the project for the algorithmic development: image enhancement, filtering, fusion, and interpretation in order to detect objects typical from airport areas. He will guide the choice of these algorithms with respect to their migration on the physical system, specifically to take advantage from the processing parallelism. He will participate and will coordinate the algorithm migration on the system.
He will participate to the validation works supervised by the project coordinator and by aircraft companies. Validations will be done either with simulations (using HIL method) or in real conditions with a system embedded on an aircraft or a vehicle navigating on the taxiways of an airport.
Finally the applicant will have to interact with the project partners, preparing formal presentations for internal meetings and workshops. He will also participate to the scientific dissemination with the RAP permanent researchers.
- WORK TO BE DONE
- State of the art and contributions on the match between physical architecture and algorithms for real-time vision.
- Monitoring of the algorithmic development in C/C++
- Low level optimization and migration of algorithms on a heterogeneous architecture selected to respect real-time constraints.
- Hardware specification (operators, memories, bus, etc)
- Project management (doc, source code, meetings, etc)
Applicants must have a PhD in embedded system, electronics…
All of some of these competences will be appreciated:
- Autonomy, team work, supervision
- Project management and communication
- Algorithms for real time vision (OpenCV, Matlab), with basic knowledge in projective geometry (camera model, calibration, homography…)
- Embedded software development (C, optimization SIMD NEON)
- Hardware development (VHDL, Xilinx Vivado, test-bench programming)
- Migration of real-time vision algorithms on heterogeneous architectures FPGA+CPU.
- Methodology for the design of heterogeneous system.
One year contract extended at least one year
- CONTACT :
Michel Devy : email@example.com, +33 5 61 33 63 31
Jonathan Piat : firstname.lastname@example.org, +33 5 61 33 63 44
Location: LAAS-CNRS, Toulouse, France
Duration: 24 months
We are seeking a strongly motivated and mature post-doctoral researcher to take in charge DFT Modelling for depicting complex chemistries at metal oxide surfaces and interfaces. Position is in Toulouse, under the supervision of A. Estève. The fellowship, funded by the IDEX – University of Toulouse should start in 2016, at the earliest convenience.
The post-doc researcher will be fully integrated within a highly multi-disciplinary project’s team (chemists, physicists and technologists) working on the development of multifunctional & performance-tailored nanomaterials for energy.
The applicant will have a pivotal role in the project by bringing fundamental aspects of chemical processes at metal-oxide surfaces (mostly ZnO) and their interfaces with metal and organics. The applicant will work in close collaboration with experimentalists located in Toulouse (LCC, CIRIMAT, CEMES) and at the University of Texas at Dallas.
Among aspects which will be of interest, there are: reaction mechanisms, i.e. atomic or molecular (organic, bio, inorganic) interactions with surfaces, solvent issues, cooperative effects, role of defects are to be investigated in close relation with experimental developments.
Required Education and Experience - A recent PhD degree (within last three years) in Materials Science, noticeably Chemistry or related disciplines is required. We seek for a strongly motivated student with strong background in computational materials sciences, particularly in atomic scale modelling; skills in manipulating Density Functional Theory codes (both periodic or cluster packages) is mandatory. The applicant should send a detailed CV, including a list of publications and communications and a motivation letter to email@example.com.