Scientific direction Development of key enabling technologies
Transfer of knowledge to industry

PhD : selection by topics

Engineering science >> Mathematics - Numerical analysis - Simulation
14 proposition(s).

Inverse reinforcement learning of a task performed by a human

DPLOIRE (CTReg)

Autre

01-01-2019

SL-DRT-19-0262

laurent.dolle@cea.fr

Learning from demonstration involves an agent (e.g., a robot) learning a task by watching another agent (e.g., a human) performing the same task. It often uses reinforcement-learning methods to improve the robot's ability to perform a task in new situations (i.e., generalization). These methods involve providing a positive reinforcement (i.e., a reward) when the outputs of the algorithms help achieving the task, but require a human designed reward function. The more the task is complex the more difficult is the reward function to design, but it can be learned from a series of examples with methods called inverse reinforcement learning. The use, jointly or not, of these techniques has shown encouraging results, but which are limited to toy examples and cannot be adapted as such to tasks more representative of the industrial environment. During the thesis, the PhD student will analyze and test state-of-the-art previous works. S/He will then propose a method, combining inverse reinforcement learning to other algorithms (e.g., generative adversarial networks, GAN), so that the robot will understand the task performed by the operator (with as little explanation from the operator as possible), and will generalize enough to make the robot robust to dynamic environments (obstacles, moving objects?). This method should be suited for a "pick and place" task in an industrial environment and ensure a reasonable enough learning period (information a priori, feedback from the operator) for tasks of medium complexity.

3D Objects discovery in 3D scene

DPLOIRE (CTReg)

Autre

01-01-2019

SL-DRT-19-0269

anthony.mouraud@cea.fr

Object detection and localization in images is a problem studied since many years. The latest technological developments now allow the real-time acquisition of depth data coupled to color data (RGBD). At the same time, modern machine computing capabilities and intelligent image processing methods have led to significant advances in the detection / localization of 2D objects with many different approaches (bounding boxes, contours, from CAD models ...). An important step is being taken in recent years with the research conducted to directly extract the volume of detected objects and their position in 3D. These works are still in their infancy, but the first results are encouraging, both from 2D images (eg DeepManta) and from 3D images (eg Deep Sliding Shapes). However, there remain several identifiable scientific / technological barriers before allowing the democratization of this type of approach for the automatic extraction of objects in potentially unknown scenes. The objective of this work is to identify the current approaches of detection / localization of 3D objects, to target their weaknesses and work on new processing technologies to mitigate them. Moreover, the object discovery in unknown environments and the inference of the operator's intention by observation / location of his attention are two areas of interest that this work aims at addressing. Beyond their applications for demonstration learning, the software bricks resulting from this project can also be reused for other applications such as augmented reality ("smart" scanning, etc.), surveillance or mobile mobility for example.

SPAD Imager for HDR ToF using multimodal data fusion

Département Architectures Conception et Logiciels Embarqués (LIST-LETI)

Laboratoire Circuits Intégrés, Intelligents pour l'Image

01-10-2019

SL-DRT-19-0301

william.guicquero@cea.fr

Depth sensors are currently a very high trending topic. Indeed, in the fields of autonomous vehicles, portable electronic devices and the Internet of Things, new technology enablers now tend to provide handy 3D image data for future innovative end-user applications. There is a great diversity of 3D sensor types, either using passive imaging (depth from defocus, stereovision, phase pixels...) or using active imaging (ultrasounds, structured light, Time-of-Flight...). Each of these systems addresses specifications in terms of depth dynamic range (accuracy of the measurement versus maximum distance). In this thesis, we will study the specific case of Single Photon Avalanche Diodes (SPAD). Recent scientific results regarding this electro-photonic component demonstrate its relevance in the context of Time-of-Flight (ToF) imaging, especially in the case of integration in a 3D-stacked design flow exhibiting a pixel pitch of the order of ten micrometers. However, the nature of the data gathered by this type of component requires significant signal processing within the sensor to extract relevant information. This thesis will aim to revise traditional approaches related to histogram processing by directly extracting statistical features from raw data. Depending on the background and skills of the PhD candidate, two research axes would be investigated. First, on the hardware side, possible modifications of SPAD based sensor architecture in order to provide ?augmented? multi-modal information. Second, on the theoretical and algorithmic side, data fusion methods to improve the final reconstruction rendering of depth maps from sensed data.

Study and implementation of non-recurrent deep learning algorithms for temporal sequences processing

Département Architectures Conception et Logiciels Embarqués (LIST-LETI)

Laboratoire Calcul Embarqué

01-01-2019

SL-DRT-19-0393

david.briand@cea.fr

Recurrent neural networks - and notably the Long-Short Term Memory (LSTM) variant - are today at the state of the art for solving many temporal sequence classification problems and in particular used in speech recognition applications (from 2015 for Android) and automatic translation (from 2016 at Google, Apple and Facebook). This type of algorithm is also successfully applied in various applications such as audio event recognition, denoising, language modeling, sequences generations, etc. The success of these approaches comes however with the cost of huge computing power requirements. This is why most of this algorithms are run on the Cloud, and not on the Edge. Moreover, recurrent neural networks are very sensitive to training parameters and can be difficult to converge because gradients internal to their recurrent structure can easily explode or vanish to zero. The adaptation of these algorithms for an embedded implementation is therefore not straightforward, because the recurrence requires a high precision and partially sequential (large latency) computing. Some technics for overcoming these difficulties are starting to appear, but are still in their infancy. Among them, a non-recurrent technic allowing sequence processing with less constrain than LSTM seems promising: hierarchical networks. Temporal convolution networks (TCN) are one of their application. The advantages and drawbacks of this model are studied notably in "An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling" (Shaojie Bai, J. Zico Kolter, Vladlen Koltun). A basic implementation of each structure showed that TCN are more efficient in almost every test cases. Internal gradients are much more stable and computation can be easily parallelized thanks to the elimination of the recurrence.

Cross-layer security reinforcement of vehicular wireless communication protocols

Département Systèmes

Laboratoire Communication des Objets Intelligents

01-10-2019

SL-DRT-19-0515

benoit.denis@cea.fr

Vehicular wireless connectivity (also referred to as V2X for Vehicle to Everything) is seen today as a core enabler of future cooperative intelligent transport systems (C-ITS) (ex. highly autonomous driving, vulnerable users safety, fleet/trajectories coordination, vehicular mapping and vehicular Internet of Things?). ITS-G5, which relies on the IEEE 802.11p radio standard operating at 5.9 GHz, or C-V2X/LTE-V, which is an adaptation of 4G cellular solutions into the vehicular context (standard under definition), are two examples of relevant technologies frequently promoted in this context. However, related « open » V2X transmission modes are most often based on information broadcast (i.e., to reach the highest numbers of neighboring vehicles around) and as such, they are highly vulnerable (ex. with public control frequency channels). Accordingly, many kinds of attacks must be considered, including critical services denial (ex. through jamming, messages injection/interception, impersonations?). So far, most of the security schemes put forward in this context rely on conventional cryptographic techniques and tools (i.e., using non-specific keys, pseudonyms or signatures). On the one hand, the main security features (i.e., primitives, seeds and algorithms?), which are determined in a static way, can be over-sized in some particular vehicular use cases. On the other hand, the resulting cryptographic overhead (in terms of computational complexity and access to the core network) contribute to strongly increase the latency of protected systems, what may be not compliant with safety applications. In the frame of these PhD studies, we thus propose to define and evaluate new security mechanisms that could take benefits from different layers of the V2X protocol stack, as well as from the specificities of the vehicular application context itself (ex. « stealth » radio resource allocation, pseudo-random access and/or messages periodicity reducing the predictability of over-the-air data traffic, neighbors' trust assessment by cross-checking the consistency of exchanged application data...), while completing and reinforcing existing security schemes. A first step of these investigations with consist in conducting an in-depth risk analysis with respect to the specifications of current V2X standards. Then, some of the counter-measures proposed to mitigate most critical attacks will be validated by means of both simulations and field experimental data.

Short wave Infra Red diffuse reflectance spectroscopy for noninvasive molecular detection medical devices

Département Microtechnologies pour la Biologie et la Santé (LETI)

Laboratoire Imagerie et Systèmes d'Acquisition

01-10-2019

SL-DRT-19-0562

anne.koenig@cea.fr

Diabetes is a major public health and industrial issue, with the number of diabetics worldwide estimated at 415 million. Until recently, to control their blood glucose, patients had to prick their fingertips. To avoid this inconvenience, laboratories have recently been offering minimally invasive measurement systems that can be interrogated using a smartphone. This evolution, although major, still poses many problems, such as its cost, its size, or its invasiveness, even reduced. As such, a medical optical sensor, reliable, inexpensive would represent a major breakthrough: many players in microelectronics such Apple or Google produce an effort in this direction. In vitro, the sugar measurement can be performed by diffuse spectroscopy in the SWIR domain (wavelength range 1 - 1.7 µm). A fraction of the photons produced by an immersed light source (emitter) diffuse into the liquid and emerge from it by the interplay of multiple reflections. The sugar absorption is on a band around 1.5 µm, the fraction of light emerging and detected in this range will be as lower as the sugar concentration is high. Multispectral analysis provides realistic concentration measurements. In vivo, the results are much worse because of the heterogeneity of the biological tissues and the presence of many interferents (other absorbents). The purpose of the thesis is to remove these biases by developing a new type of optical sensor comprising on the one hand several emitters, on the other hand a plurality of detectors. This subject concerns a candidate who must have received a physicist training with a solid module dedicated to optics / photonics, interested in a work at the interface between physics and biology.

A quantum algorithm toward a practical evaluation of Worst-Case Execution Times (WCETs) for real-time tasks

Département Architectures Conception et Logiciels Embarqués (LIST-LETI)

Laboratoire Calcul Embarqué

01-10-2019

SL-DRT-19-0568

sergiu.carpov@cea.fr

Schedulability analysis is an old research field related to both real-time systems and performance of systems. An important requirement of the schedulability analysis is the availability of upper bounds of execution times. As a consequence, the research field of Worst Case Execution Times (WCETs) started to flourish in the 90s. For safety critical real-time systems, missing a deadline can lead to catastrophic results, with important damages or even loss of human lives. As such event must be avoided, the schedulability analysis must provide a safe result. So if exact WCETs cannot be expressed, which is usually the case because they depends on too many parameters including a perfect model of inner-work of the target microprocessor, then a safe upper bound of the WCETs must be provided instead. However, if it is indeed easy to show overly pessimistic WCETs, e.g. by disabling caches, and serializing instructions in the pipeline, the usefulness of a given WCET estimation decreases as its accuracy is reduced. So a useful WCET should be both safe and as accurate as possible, which is a tricky problem because execution times in modern systems are heavily context dependent, and the mathematical models of WCETs are NP-hard problems. Within this PhD work we aim at two target objectives: The first si to find a way of modeling the evaluation of WCETs with a quantum algorithm, and the second is to open up another field of application of quantum computing outside of the usual fields of reseach.

CNN-3d-lensfree

Département Microtechnologies pour la Biologie et la Santé (LETI)

Laboratoire Imagerie et Systèmes d'Acquisition

01-09-2019

SL-DRT-19-0605

lionel.herve@cea.fr

At CEA-Leti, we are developing lensfree microscopy for the monitoring of cell culture. This technique overpass several limits of conventional microscopy (compactness, field of view, quantification, etc.). Recently we showed, for the first time, 3D+time acquisitions of 3D cell culture with a lens-free microscope. We observed cells without any labelling within the volume as large as several cubic millimeters over several days. This new mean of microscopy allowed us to observe a broad range of phenomena only present in 3D environments. However, two drawbacks are still present on the microscope prototype: a long reconstruction time (>1 hour/frame) and the reconstructed volumes present artefacts owing to the limited number of angular acquisitions. The thesis work will focus on the ability of deep learning technologies to overcome the above-mentioned limitations. Basically, a convolutional neural network will be trained on the basis of simulated 3D cell culture volume (ground truth) and simulated response of our current 3D lensfree microscope (input). This approach is expected to accelerate the reconstruction process and to allow full 3D reconstructions. Yet it poses two scientific questions: are simulated data pertinent to train a neural network and how can we assess the quality of 3D reconstruction obtained through a neural network? Profile of the candidate sought: - Engineering degree in applied mathematics or physical sciences. - Strong knowledge in image processing with skills in deep learning.

Spintronic Wake-Up Radio

Département Architectures Conception et Logiciels Embarqués (LIST-LETI)

Laboratoire Architectures Intégrées Radiofréquences

01-09-2019

SL-DRT-19-0645

dominique.morche@cea.fr

The increasing number of wireless connected objects and smart sensors requires defining components and operational schemes that drastically reduce the power consumption. Within such communicating networks the RxTx modules are the most power consuming elements. The solution actively searched for is to switch off the main RxTx module when no communication is requested and to use a low power, degraded wake-up radio receiver WuRx that will switch on the main module when receiving an according wake-up signal. The realization of robust and ultralow power WuRx is an active field of research. The thesis proposes to explore RF spintronic devices as such a compact and low power solution. Magnetic tunnel junctions, which are the main spintronics building blocks, are capable to passively convert an RF signal into a DC signal, with frequency selectivity and at relatively high output signal levels. LETI/DACLE and INAC/SPINTEC work together on the realization of such spintroncis based WuRx and the PhD project will be at the interface of the two laboratories. While SPINTEC will realize the devices and optimize their sensitivity to low input signal levels, the thesis will be carried out at LETI/DACLE to realize the corresponding antenna networks and rf electronics. In order to establish the performance parameters the student will first spend some time at SPINTEC to get trained on the characterization of spintronic based rf components. The student will also be involved in the testing of the developed rf circuits with the spintronics components to iteratively optimize the electronic circuits and adapt it to the spintronics device performances.

Self-adaptation and Resilience for Artificial Intelligence-based Robotic Systems

Département Ingénierie Logiciels et Systèmes (LIST)

Labo.conception des système embarqués et autonomes

01-04-2019

SL-DRT-19-0757

ansgar.radermacher@cea.fr

Robotic systems have to integrate more & more functionality including autonomous decisions how to adapt to changing environment conditions or failures. These systems have to respect classical requirements of embedded systems (resource, timeliness), but resilience to failures and safety requirements become very important. Stopping the system is not an option for instance for an autonomous vehicle or a drone, systems have to be fail-operational. Another aspect is the use of AI components (machine learning) in control algorithms and for taking autonomous decision require a specific validation of functional safety. The objective of this thesis is to embed validation mechanisms into the running system ? in addition to execute offline (model-based) validation. These could be based on runtime models that incorporate constraints and safety requirements defined in design models. The approach should also handle auto-adaptation in order to stay operational in the presence of faults. The work will examine use case from existing robotic projects in our laboratory, notably from collaborative robotics (robot arms) and the European project RobMoSys.

MILP models for optimal management of hybrid CSP plants

Département Thermique Biomasse et Hydrogène (LITEN)

Laboratoire Systèmes Solaires Haute Température

01-12-2019

SL-DRT-19-0769

valery.vuillerme@cea.fr

The hybridisation of CSP plants with "conventional" plants (existing or not) has many advantages, but in return brings more complexity in the control of the system and the steering strategy. At present, we do not find in the study literature dealing with this issue and highlighting the principles of predictive management of hybrid CSP plants. The planned work will allow the development of dynamic models of CSP plants by taking advantage of the special possibilities offered by the Cathare code in co-simulation environment (PEGASE). This environment will address the advanced control-command aspects of hybrid CSP systems (biomass, geothermal, incineration, coal, nuclear, H2 ...), and we will implement examples to demonstrate the contribution of predictive management to such systems constrained by variations in demand (market) and resource. For validation purposes, certain internal and external experimental means will be used, such as the prototypes of the solar zone at Cadarache for elements relating to CSP / storage pairs, or infrastructures made available under the SFERA III project to which participates the CEA. Ultimately, the numerical demonstrator developed will highlight the hybrid CSP systems of major interest at national, European and international level.

Robustness and privacy of graph neural networks: homomorphic encryption and randomization

DM2I (LIST)

Laboratoire d'Analyse des Données et d'Intelligence des Systèmes

01-10-2019

SL-DRT-19-0907

cedric.gouy-pailler@cea.fr

In various domains, graphs represent a useful representation for many types of data. Prominent examples entail behavioural analyses performed in cybersecurity or social network analysis. In the former, internet user behaviour can be observed by monitoring DNS requests, interpreted as successive steps of a random walker on a graph in which nodes represent domain names and edges represent population-level average behaviour. Therefore studying user behaviour can be done by analyzing the subgraph induced by specific user movements. In the latter, graph representations naturally emerge from user interactions. For example nodes can represent users, and any relation between two users (messages or common interests) can be interpreted as edges. Understanding and analyzing graph structures appear to be a key tool in many real-world applications. It is thus essential to find efficient and robust methods for tasks such as node or graph classification. In this context graph neural networks appear as a key technology, but raise crucial questions about their robustness to adversarial attacks and privacy. The goal of this thesis is to explore robustness and privacy of graph-neural network-based approaches, by considering solutions combining randomization and homomorphic encryption to ensure a satisfying compromise between performance, robustness and data privacy.

Machine learning methods using uncertain labels, human stress estimation application

Département Systèmes

Laboratoire Signaux et Systèmes de Capteurs

01-10-2019

SL-DRT-19-0915

christelle.godin@cea.fr

With wearable sensors development, it is now possible to monitor physiological parameters and activity. Many studies show that stress level assessment can be done by using those measurements. Supervised machine learning methods used for it are flourishing. They suppose that for each measurement, a ?ground truth? stress level is available. However, while doing experiments for database construction it is not possible to attribute an exact stress label to each event but subjective values are easily available. The goal of the PhD is to take into account data with uncertain, fuzzy, redundant, contradictory or missing values in order to obtain a better stress estimator. This kind of approach should be useful for a lot of applications including other mental state estimation like drowsiness detection, mental disease diagnosis, emotion estimation.

Spike based processing chain for signal classification

Département Architectures Conception et Logiciels Embarqués (LIST-LETI)

Laboratoire Architectures Intégrées Radiofréquences

01-09-2019

SL-DRT-19-0990

dominique.morche@cea.fr

The expansion of the internet of things is conditioned by our ability to develop innovative systems able to apprehend and understand the environment while having an ultra low power consumption, compatible with energy harvesting. To reach such a goal, one of the solution which is knowing a considerable renewed interest is the use of acoustic signals. Their low frequencies undoubtedly induces a low power consumption in the circuit interface and their low cost eases the dissemination of this solution. There's a huge applicative potential: wake-up by key words (the well known ?ok google?), choc detection, source localization, event classification, surveillance, and machine health monitoring. In order to implement such complex functions in an energy efficient manner, the potential of neural networks is more and more considered. However today, these solutions are too power consuming. To reduce this power, several alternatives are considered. One of the most promising is the coding of the signal in spike, coherently with neuromorphic architecture. Recently, CEA-LETI has developed a new ADC architecture which directly generate some spike and the best power efficiency in the state of the art has been reached. The aim of this PhD is to follow up this work by implementing in the analog domain some feature extraction in order to reduce the complexity of the neural network processing. To reach the best energy efficiency, a joint optimization between the analog, digital and algorithmic part is mandatory. In the scope of this PhD, CEA-LETI and EPDFL are collaborating to develop this new analog processing interface, adapted to neural networks based on spike processing. The main objective is ti setup a methodology to reduce the power consumption in all the sensing systems. The automotive applications will be particularly considered. Other application areas and different kind of signal might be also studied.

Voir toutes nos offres