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

PhD : selection by topics

Technological challenges >> Artificial intelligence & Data intelligence
23 proposition(s).

See all positions [+]

Fault injection and integrity of edge neural networks: attacks, protections, evaluation

Département Systèmes (LETI)

Laboratoire Sécurité des Objets et des Systèmes Physiques

01-02-2021

SL-DRT-21-0159

pierre-alain.moellic@cea.fr

Artificial intelligence & Data intelligence (.pdf)

One of the major trends of Artificial Intelligence is the large-scale deployment of Machine Learning systems to a large variety of embedded platforms. A lot of semi-conductor practioners propose "A.I. suitable" products, majoritarely with neural networks for inference purpose. The security of the embedded models is a major issue for the deployment of these systems. Several works raised threats such as the adversarial examples or the membership inference attacks with disastrous impact. These works consider the ML aglorithms through a pure algorithmic point of view without taking into consideration the specificities of their physical implementation. Moreover, advanced works are compulsory for physical attacks (i.e., side-channel and fault injection analysis). By considering a overall attack surface gathering the theoretical (i.e. algorithmic) and physical facets, this subject propose to analyze Fault Injection Analysis threats (FIA) targeting the integrity of the model (fooling a prediction) of embedded machine learning systems and the development of appropriate protections. Several works have studied physical attacks for embedded neural networks but with usually naive model architecture on 'simple' 8-bit microcontrolers, or FPGA or at a pure simulation level. These works do not try to link the fault models or the leakages with well-known algorithmic threats. Being based on the experience on other critical systems (e.g., cryptographic primitive), the main idea of this PhD subject will be to jointly analysis the algorithmic and physical world in order to better understand the complexity of the threats and develop efficient defense schemes. The works will answer the following scientific challenges: (1) Caracterization and exploitation of fault models: how to exploit fault injection mechanisms (laser, EM, glitching) to fool the prediction of a model with minimal perturbations. (2) Evaluation of the relevance of classical countermeasures (such as redundancy-based techniques) for this kind of systems and threats. (3) Develop new protections suitable to embedded neural networks.

Download the offer (.zip)

Lensless imaging and artificial intelligence for rapid diagnosis of infections

Département Microtechnologies pour la Biologie et la Santé (LETI)

Laboratoire Systèmes d'Imagerie pour le Vivant

01-10-2020

SL-DRT-21-0380

caroline.paulus@cea.fr

Artificial intelligence & Data intelligence (.pdf)

The objective of the thesis is to develop a portable technology for pathogen identification. Indeed, in a context of spread of medical deserts and resurgence of antibiotic-resistant infections, it is urgent to develop innovative techniques for rapid diagnosis of infections in isolated regions. Among optical techniques for pathogen identification, lens free imaging methods draws attention because they are the only ones currently able to offer simultaneous characterization of a large number of colonies, all with low-cost, portable and energy-efficient technology. The objective of the thesis is to explore the potential of lensless imaging combined with artificial intelligence algorithms to identify bacterial colonies present in a biological fluid. The thesis will aim to optimize the sizing of the imaging system (sources, sensors) and to study image processing and machine learning algorithms necessary for colony identification. Two cases of clinical applications will be studied.

Download the offer (.zip)

Oscillating neurons for computational optimization and associative memory

Département Composants Silicium (LETI)

Laboratoire d'Intégration des Composants pour la Logique

01-10-2021

SL-DRT-21-0393

louis.hutin@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Hopfield networks are a type of recurring neural network particularly well-suited to content-addressable associative memory functions. By giving its elements the ability to fluctuate at will, they can be adapted to efficiently solving NP-hard combinatorial optimization problems. Such problems, for which finding exact solutions in polynomial time is out of reach for deterministic Turing machines, find many applications in diverse fields such as logistic operations, circuit design, medical diagnosis, Smart Grid management etc. The frame of the proposed project is the search for hardware accelerators for Artificial Intelligence. In particular, we consider the use of injection-locked oscillators as neurons (ILO). The goals will be the design, fabrication and demonstration of such networks, featuring binary phase-encoded neurons coupled by adjustable synaptic weights, to carry out associative memory (e.g. pattern recognition) or combinatorial optimization tasks (ex: max cut, graph coloring,...).

Download the offer (.zip)

Hybrid representation of polarising topics in political news

Département Intelligence Ambiante et Systèmes Interactifs (LIST)

Laboratoire Analyse Sémantique Textes et Images

01-02-2021

SL-DRT-21-0419

julien.tourille@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Existing recommendation algorithms tend to create filter bubbles which reduce the diversity of information proposed to users and lead to polarized opinions. This functioning mode is problematic particularly for domains such as politics, in which balanced and diversified information is needed to fuel a healthy debate. The main objectives of this PhD are to: (1) automatically detect topics which are prone to polarization, (2) propose new text representations which combine objective and subjective description criteria, (3) update the representations to take into account the highly dynamic nature of political texts and (4) propose new datasets and evaluation protocols which are adapted for the methods introduced in the paper. The results of the PhD work will constitute the entry point for diversified recommendation algorithms which aim to open the filter bubbles.

Download the offer (.zip)

Training and quantization of large-scale deep neural networks for transfer learning

Département Systèmes et Circuits Intégrés Numériques

Laboratoire Intelligence Artificielle Embarquée

SL-DRT-21-0446

johannes.thiele@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Training and quantization of large-scale deep neural networks for transfer learning Transfer learning is today a common technique in Deep Learning that uses the learned parameters of a generic network (a feature extractor) to accelerate the training of another network on a more specific task. This specialized network is subsequently optimized for the hardware constraints of the specific use-case. However, given that the representations of the feature extractor are often rather generic, it might be possible to optimize the parameters before the transfer, to avoid that each end-user has to perform this optimization by herself. In this context, the thesis has the following scientific objectives: - Using several ?unsupervised? learning methods (self-supervised, weakly supervised, semi-supervised) to train feature extractors on large datasets - Studying how common optimization methods (in particular quantization) can be applied on these extractors in a ?task-agnostic? fashion - Quantifying the influence of these optimizations on the transfer learning capacity, by benchmarking and theoretical analysis (e.g. information compression theory) Required competences: Master degree (or equivalent), machine learning (in particular Deep Learning), programming (Python, Pytorch, Tensorflow, C++), good English (French knowledge is not required, but helpful)

Download the offer (.zip)

Embedded autonomous incremental learning

Département Systèmes et Circuits Intégrés Numériques

Laboratoire Intelligence Intégrée Multi-capteurs

01-09-2021

SL-DRT-21-0465

anca.molnos@cea.fr

Artificial intelligence & Data intelligence (.pdf)

The recent development of incremental learning algorithms for deep neural networks is an opportunity to imagine new intelligent sensor applications deployed in real environments. By being incrementally able to learn new tasks, the sensor will be able to personalize its behavior to its specific deployment milieu, allowing it to adapt to slow variations of its targeted tasks (e.g. the detection of different types of anomalies) or learn new tasks that were not initially anticipated. This possibility would make the service rendered by the autonomous sensor more and more relevant. The objective of this thesis is the exploration of the means by which the intelligent sensor can become fully autonomous in its evolution while taking into account the limited processing capability of the embedded system. Also, seeing the limited power consumption of the platform, the idea is to associate two embedded systems, a first which is ?Always-on? and executes the nominal task of the application (e.g. the detection of different classes of events or anomalies), and a second which is ?On-demand? which would be executed now and then, in order to retrain the model of the ?Always-on? part. For coherence, it is necessary that the power consumption ratio of the two platforms be in a ratio of 1:100 to 1:1000 approximately. The challenges facing the design of such a system are many : The first is the design of detection mechanisms able to find false negative examples (slowly changing classes) as well as novel examples (new classes). These mechanisms must be executed on the ?Always-on? platform, with the associated implementation constraints. A second difficulty concerns the retraining phase which is executed on the ?On-demand? platform. This phase must take into account the structure of the ?Always-on? model in order to be able to retrain it with new examples. This both in order to slowly learn the modifications of the existing detection task or to learn a new task without forgetting the old ones. Since this a new application space, the PhD candidate must be able to have a wide understanding of the subject and will necessarily have to address a wide number of domains including different incremental leaning algorithms, different deep learning training algorithms, and the hardware requirements necessary for running these algorithms in the embedded context.

Download the offer (.zip)

Real-world data and AI for an innovative analytical approach to intonation and speech intelligibility in children with Cerebal Palsy

Département Intelligence Ambiante et Systèmes Interactifs (LIST)

Laboratoire d'Interfaces Sensorielles & Ambiantes

01-10-2021

SL-DRT-21-0539

margarita.anastassova@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Cerebral palsy (CP) is a developmental motor disorder affecting an individual's ability to move, and maintain balance and posture. It affects 2 to 4 children in 1,000, making this lifelong, chronic condition the most common motor disability in childhood. In addition to motor problems, many children with CP have difficulties with speaking, which severely impacts their access to social and educational activities. The most frequent speech impairment in CP is developmental dysarthria, characterized by limited movement of the jaw, lips and tongue; imprecise articulation; slow speaking rate; reduced intonation with a limited variation in pitch, rhythm and volume, which results in impaired speech intelligibility. Despite the frequent occurrence of problems in intonation and speech intelligibility in children with CP, the heterogeneity of the profiles and the central role of intonation in communication, few studies have examined intonation patterns in this population in order to characterize and classify them. As a result, little is known about intonational difficulties and their relationship to intelligibility in children with CP. Knowledge on the relation of intonational patterns and expressions with motor activities in real-world tasks is even more scarce. The aim of the research project is to fill in the above-mentioned knowledge gaps. This will be done using a real-world data-driven approach, combined with innovative analytical approaches based on AI and machine learning.

Download the offer (.zip)

Transfer Learning and Optimal Transport applied to the adaptation of models learnt on synthetic data

Département Métrologie Instrumentation et Information (LIST)

Laboratoire Science des Données et de la Décision

01-10-2021

SL-DRT-21-0563

fred-maurice.ngole-mboula@cea.fr

Artificial intelligence & Data intelligence (.pdf)

This PhD thesis aims at exploring possible contributions of optimal transportation field to transfer learning through the following directions: - building a knowledge transferability criterion between a source and a target task based on the regularity of the transportation plan between the source and the target data distributions; - integrating priors on the tasks similarity through the transportation ground metric; - applying Wasserstein barycenter to multi-task learning problems. These works might find multiple use-cases of interest in the lab, including adaptation of models learnt on synthetic data to real world systems. A more detailed presentation of this PhD thesis subject can be found via the following link: https://drive.google.com/file/d/13RAQEi0PdnkllM-MHxQS50WWUNUtGS07/view?usp=sharing

Download the offer (.zip)

Blending intuition with reasoning - Deep learning augmented with algorithmic logic and abstraction

Département Ingénierie Logiciels et Systèmes (LIST)

Labo.conception des systèmes embarqués et autonomes

01-03-2021

SL-DRT-21-0617

shuai.li@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Within machine learning, deep learning, based on neural networks, is a subfield that has gained much traction since several high-profile success stories. Unlike classical computer reasoning, the statistical method by which a neural network solves a problem can be seen as a very primitive form of intuition, as opposite to classical computer reasoning. However, so far the only real success of deep learning has been its ability to self-tune its geometric logic that lets it transform data represented as points in n-dimension, to data represented as points in m-dimension, if we provide enough training data. Unlike a human being, a neural network does not have the ability to reason through algorithmic logic. Furthermore, although neural networks are tremendously powerful for a given task, since they have no ability to achieve global generalization, any deviation in the input data may give unpredicted results, which limits their reusability. Considering the significant cost associated with neural network development, integrating such systems is not always economically viable. It is therefore necessary to abstract, encapsulate, reuse and compose neural networks. Although lacking in deep learning, algorithmic logic and abstraction are today innate to classical software engineering, through programming primitives, software architecture paradigms, and mature methodological patterns like Model-Driven Engineering. Therefore, in this thesis, we propose to blend reusable algorithmic intelligence, providing the ability to reason, with reusable geometric intelligence, providing the ability of intuition. To achieve such an objective, we can explore some ideas like integrating programming control primitives in neural networks, applying software architecture paradigms in neural networks models, and assembling modular systems using libraries containing both algorithmic modules and geometric modules. The results of this thesis will be a stepping stone towards helping companies assemble AI systems for their specific problems, by limiting the costs in expertise, effort, time, and data necessary to integrate neural networks.

Download the offer (.zip)

Scalable and Precise Static Analysis of Memory for Low-Level Languages

Département Ingénierie Logiciels et Systèmes (LIST)

Laboratoire pour la Sûreté du Logiciel

01-10-2021

SL-DRT-21-0641

matthieu.lemerre@cea.fr

Artificial intelligence & Data intelligence (.pdf)

The goal of the thesis is to develop an automated static analysis (based on abstract interpretation) to verify, in large code base in low-level compiled languages (e.g. C, C++, assembly, Rust, Fortran), security properties that are related to memory, lke flow information properties and absence of memory corruption. This problem has many applications in cybersecurity, as most of the software-related cybersecurity issues, and those that have the highest severity, come from memory safety errors (e.g. (buffer overflows, use-after-free, null pointer dereferences, wrong type punning, wrong interfacing between several languages, etc). The three main issues when designing such an automated static analysis is to keep the verification effort low, to handle large and complex systems, and to be precise enough so that the analysis does not report a large amount of false alarms. The privileged approach in this thesis will draw on the success of a new method using abstract domains parameterized by type invariants, which found a sweet spot between precision (i.e. few false alarms), efficiency (in computing resources), and required effort (by the user). This method allowed in particular to fully automatically prove absence of privilege escalation and of memory corruptions of an existing industrial microkernel from its machine code, using only 58 lines of annotations. Many research questions remain, and we will explore how to extend the analyzer to improve scalability (using compositional analysis), how to improve its expressivity (to show complex security properties like non-interference), how to improve precision without degrading efficiency, or how to further reduce the amount of annotations (using automatic inference of more precise type invariants).

Download the offer (.zip)

Real Time Semantic extraction on sparse data for embedded perception

Département Systèmes et Circuits Intégrés Numériques

Laboratoire Intelligence Artificielle Embarquée

01-10-2021

SL-DRT-21-0656

mehdi.darouich@cea.fr

Artificial intelligence & Data intelligence (.pdf)

The thesis topic we propose is in the field of embedded architectures for the semantic analysis of sparse data in real-time. During the last decade, the analysis of image and videostreams has experienced a boom following the significant improvement of neural networks and the increased specialization of the associated processing architectures. Research has led to the development of more efficient networks, which are less memory-intensive and increasingly integrable into embedded hardware. Several works are currently in progress in the laboratory, around the N2D2 tool for the optimization and integration of neural networks on embedded hardware, as well as around the DNeuro embedded hardware architecture. Within embedded perception systems, the strong constraints on bandwidth and memory lead to the privileged use of sparse data (graphs, point clouds, etc.), reduced in quantity and containing particularly rich information on the environment to be analyzed. However, the non-contiguous and unpredictable structure of this sparse data is very different from a traditional image stream, making current hardware architectures unsuitable for their execution. However, these particular characteristics point to very interesting opportunities in terms of optimization and efficiency. This thesis work aims at exploring this class of algorithms and their capacity of integration under constraints in an embedded computing architecture. The scientific problems that arise here are how to perform efficient data management in a context of highly scattered computational distribution, the compatibility of sparse data analysis algorithms with execution on embedded targets, and the performances and precision achievable under these constraints.

Download the offer (.zip)

Abnormal event detection in videos

Département Intelligence Ambiante et Systèmes Interactifs (LIST)

Laboratoire Vision et Apprentissage pour l'analyse de scènes

01-10-2021

SL-DRT-21-0659

aleksandr.setkov@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Abnormal event detection in videos by Deep Learning is essential in applications such as video surveillance, road safety, or autonomous driving. However, there are numerous technical and scientific obstacles such as scarcity and heterogeneity of such events, as well as scene complexity. Supervised methods require labelled data in large quantity which is difficult to obtain in this context. Anomalies are rare compared to normal data and video annotation is a tedious task. Another group of methods, called One-Class, uses normal data only to model normality in the data. However, this modeling is often hindered by the heterogeneity of both normal and anomaly data. The PhD work will focus on the methods that require a few examples (Few-Shot Learning) and/or semi- or weakly-supervised methods. The proposed new approaches will be evaluated on academic datasets as well as on real data of scenarios observed by static (video surveillance, industrial inspection, or detection of incidents) and dynamic (autonomous driving) cameras.

Download the offer (.zip)

Gamification for empowering engineers collectives

Département Ingénierie Logiciels et Systèmes (LIST)

Lab.systèmes d'information de confiance, intelligents et auto-organisants

01-09-2021

SL-DRT-21-0692

sara.tucci@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Technologies, and in particular digital technologies, are part of the range of solutions proposed to address the many societal and environmental challenges, such as the 17 sustainable development goals listed by the United Nations. Faced with the complexity of the systems to be implemented, collective intelligence is a major key to success. Multidisciplinary and holistic, systems engineering practices as formalized by INCOSE, and more particularly in their model engineering based version, are similar to collective intelligence practices, and therefore its performance relies on the group's ability to communicate and therefore on the collaborative work tools shared by the team. However, users often consider the tools as complex and the resulting "misuse" becomes a hindrance to the performance of the collective instead of a stimulus. This thesis aims to explore the fields of serious games and game theory in order to reverse this trend and make software tools allies of development rather than enemies to be fought.

Download the offer (.zip)

Tactile-based learning and classification methods for task planning and verification ? applications to multi-digital and bimanual robotic manipulation

Département Systèmes (LETI)

Laboratoire Signaux et Systèmes de Capteurs

01-09-2021

SL-DRT-21-0803

saifeddine.aloui@cea.fr

Artificial intelligence & Data intelligence (.pdf)

The robotic manipulation of objects first of all requires a grasp planning of these objects, which is a function of the characteristic parameters of the considered hardware tools and the task to be performed (such as the accessibility areas or the level and direction of the efforts that may be involved in the tasks of assembly, insertion, dexterous manipulation, etc.). In addition, during the execution of the task, it is necessary to be able to ensure the nominal progress of the planned task, by detecting the occurrence of certain critical events necessary for its completion (such as the interaction of objects with each other, the loss of stability of the object, etc.) and then validating the actual completion of the planned task (via the classification of data that characterizes the success or failure of tasks such as insertion or assembly). These detection and verification steps, which are crucial when it comes to robotizing certain critical tasks requiring a high level of traceability, can be based in particular on the analysis and monitoring of data or signals specific to the handling system in question. The work requested will exploit an experimental system consisting of a two-handed station, equipped in particular with two multi-digital grippers equipped with multimodal tactile sensors developed by the CEA. This thesis work is essentially divided into two parts. The first part consists in the use of learning methods, which are able to take into account the capacities of the pluridigital manipulators and the imperatives of the task, to plan the grasping of the objects. The second part of the thesis aims at exploiting some methods based on the classification of tactile and proprioceptive signals of the system to validate the accomplishment of the task.

Download the offer (.zip)

Attention guided dynamic inference in perception neural networks for autonomous mobile systems

Département Systèmes et Circuits Intégrés Numériques

Laboratoire Intelligence Artificielle Embarquée

01-10-2021

SL-DRT-21-0816

karim.benchehida@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Autonomous mobile systems are becoming more and more present in a variety of domains such us the delivery, inspection or agriculture executing tasks of growing complexity. These systems need to have a precise positioning (localization, pose estimation) and environment perception (detection, classification, object tracking?) to take relevant navigation decisions for example. To be very efficient, recent state of the art approaches based on neural networks for these perception tasks tend towards the use of wider (in the number of channels and modalities) and/or deeper (in the number of convolution layers) networks with a direct impact on the computation complexity and the decision making latency. The research proposal concentrates on the enhancement of the computational efficiency of perception neural networks (complexity vs. accuracy). For that purpose, we intend to dynamically reduce the computation (via dynamic inference techniques) and focus it (via attentions mechanisms) to allow the use of large representation capability networks that usually cannot be embedded. An important part of this project will be on the implementation of the developed techniques on a real embedded mobile platform to demonstrate the effectiveness of the approach. Indeed, the Embedded Artificial Intelligence Laboratory of the CEA have available some mobile robotic platforms and a fully automated, autonomous ready electrical vehicle with multiple integrated sensors. The research results will enrich the perception modules of these systems.

Download the offer (.zip)

Exploring learning techniques for "Edge AI" taking advantage of Resistive RAM

Département Systèmes et Circuits Intégrés Numériques

Laboratoire Systèmes-sur-puce et Technologies Avancées

01-09-2021

SL-DRT-21-0825

Francois.RUMMENS@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Today's computer architectures are inefficient in handling the simulation artificial neural networks, hindering their application in power-constrained environments, such as edge computing and the Internet of Thighs. Dedicated hardware implementations of neural networks that combine the advantages of mixed-signal neuromorphic circuits with those of emerging memory technologies have the potential of enabling ultra-low power processing suitable for edge computing. These new circuits and technologies have the potential to endow the system with the ability to learn at the edge. This breakthrough, which is unattainable using conventional approaches, can have many advantages, as it enables adaptation to changing input statistics, reduced network congestion, and increased privacy. However, current approaches often focus on learning algorithms that cannot be reconciled with the non-ideal physical behaviour of resistive memories. This thesis aims at exploring various algorithmic solutions for inference and learning in order to propose neural network architectures more adapted to the reality of the resistive memory technologies developed at LETI.

Download the offer (.zip)

Ultra Low Power and High Performance Microphone Signal Processing for Speaker Localization and Auditory Attention Detection : Application to Next Generation Hearing Aids

Département Systèmes (LETI)

Laboratoire Signaux et Systèmes de Capteurs

01-10-2021

SL-DRT-21-0898

vincent.heiries@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Located on the MINATEC campus in Grenoble, CEA-Leti's main mission is to create innovation and transfer it to industry by generating research results that will be used in industry in the medium and long term, positioning its research between academic research and industrial R&D. Within LETI Systems Department, the mission of the Sensor Systems and Electronics Service is to design and produce innovative systems to meet the needs of industrial innovation in a wide range of fields, from the automotive industry to sports and the building industry. The skills involved range from electronics to physics, electromagnetism, magnetostatics, signal processing and applied mathematics. Hearing loss is a major public health problem, affecting about 10% of the world's population. This handicap has a strong impact on the comfort of patients who suffer from it, in many aspects of their lives. Furthermore, with increased stimulation of our hearing system over long periods of time through various digital uses, the trend of increasing prevalence of hearing loss is clearly on the rise. Many forms of hearing loss can be treated through the use of hearing aids that significantly improve the lives of millions of people with hearing loss around the world. These hearing aids have benefited from considerable efforts to improve the underlying technologies in recent years, and today offer very high performance in terms of audio signal quality, amplification, noise filtering, compactness, and autonomy. However, these devices still have several limitations. In particular, in certain sound environments, the separation between the useful signal to be amplified and the interfering acoustic signals to be filtered remains a challenge. In this study, we propose to focus on the Cocktail Party Problem. The Cocktail Party Problem (CPP), is a psychoacoustic phenomenon that refers to the remarkable human ability to listen and selectively recognize an auditory source in a noisy environment, where the overlapping auditory interference is produced by competing speech sounds or a variety of noises that are often assumed to be independent of each other. The resolution of this type of problem, also called Auditory Attention Detection, represents a major problem for which few solutions have yet been found and which is currently the subject of intense research. This PhD thesis, which is part of the "Cyber-Physical Systems" and "Edge AI" roadmap of the Systems Department of CEA-LETI (Grenoble), will aim to make a major contribution to this Auditory Attention Detection theme, for the automatic recognition of the speaker by future generation hearing aids. The thesis will be based on advanced technological solutions using embedded artificial intelligence (Edge AI). We will address the problem through a multi-sensor data fusion approach (acoustic, inertial, video sensors). Indeed, we will consider coupling a processing of acoustic voice signals thanks to high performance microphones with a video processing of faces to realize a vocal activity detection of the speaker (automatic lip reading). The sensor data will be processed and coupled by adapted artificial intelligence algorithms. It is also envisaged to use several microphones to perform acoustic beamforming processing, and to possibly hybridize with inertial sensors to reinforce the localization estimation of the speaker. The validation of the implemented methods and the developed algorithms will be realized thanks to test campaigns in instrumented acoustic chamber (high performance microphone, video captures, etc...). Keywords: hearing aid, audio signal processing, artificial intelligence, sensor fusion, cocktail party problem, auditory attention detection

Download the offer (.zip)

Development of analysis tools based on machine learning for gamma spectrometry dedicated to in-situ measurements

Département Métrologie Instrumentation et Information (LIST)

Laboratoire de Métrologie de l'Activité

01-09-2021

SL-DRT-21-0911

christophe.bobin@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Gamma-spectrum analysis is a classical technique applied for identification and quantification assessment of radionuclides in radioactive sources. Automatic identification for in-situ applications by non-expert users needs specific algorithms to meet the demand in various fields from environmental measurements to decommissioning of nuclear facilities. The Laboratoire national Henri Becquerel (LNHB), located at CEA/Saclay, is involved in the development of an automatic analysis tool of gamma spectra at low statistics based on spectral unmixing, that can be applied to scintillation detectors (NaI:Tl, plastic). The robustness of decision-making is generally limited by the variability of in-situ measurement conditions because of the deformation of gamma spectra due to attenuation and scattering phenomena in the surrounding of a radioactive source. The objective of the thesis is the development of new spectral unmixing methods based on machine learning allowing an automatic processing of these deformations of gamma spectra. The purpose is also to obtain an algorithm that can be implemented in embedded electronic systems of portable detection devices.

Download the offer (.zip)

Département des Plateformes Technologiques (LETI)

Laboratoire

01-10-2021

SL-DRT-21-1012

jerome.reche@cea.fr

Artificial intelligence & Data intelligence (.pdf)

With the rapid evolution of the semiconductor industry, each new technological node involves reduced size features and / or increasing complexities in the 3D morphology of nanostructures. This directly impacts the lithography step, the first link in manufacturing, in particular through the metrological control step which must be more complete but also more reliable. This results in very low allowable tolerances on measurement errors. Today, no single metrology technique can meet the stringent demands of the industry. In order to answer this, a hybrid metrology, based on the concept of data fusion, is very promising. It consists in jointly deploying different metrology techniques in order to combine their strengths while mitigating their weaknesses. At the stage of our current research developed as part of the ECSEL MadeIn4 project, a neural network (NR) approach has made it possible to design and optimize a neural structure for line patterns characterized by 3 dimensional parameters. The validation of the neural network is based on simulated data from physical models specific to each technique from which the dimensional parameters are extracted. The next natural step, which is the objective of this thesis, is to develop a neural network directly using raw data (image, spectrum, etc.) from different measurement instruments. The RN learning phase will be based on virtual data from physical modeling and the final validation of the RN will be performed on data from state-of-the-art metrology instruments available in our clean rooms and measured on samples optimized for this study.

Download the offer (.zip)

Symbolic Execution for Neural Networks

Département Ingénierie Logiciels et Systèmes (LIST)

Laboratoire exigences et conformité des systèmes

SL-DRT-21-1061

arnault.lapitre@cea.fr

Artificial intelligence & Data intelligence (.pdf)

To evaluate the confidence that we can have in a neural network, one of the issues is to understand how the network classifies the data. The size of neural networks used industrially, in terms of number of execution paths, but also in terms of number of parameters manipulated, is such that a human analysis is not possible. It is therefore necessary to automate this analysis and among the approaches recently explored in the state of the art, we find techniques based on symbolic execution. These approaches, based on the exploration of the different execution paths of a program, allow to perform structural coverage, to implement heuristics dedicated to the reachability of parts of the program or to the verification of properties, and to generate test cases for these objectives. The purpose of this thesis is to propose adaptations of existing symbolic execution techniques to apply them to the verification of neural networks that offer new challenges in terms of program exploration, in order to reveal and analyze undesirable behaviors after having characterized them.

Download the offer (.zip)

Learning a class parameterized detector with 3D models

Département Intelligence Ambiante et Systèmes Interactifs (LIST)

Laboratoire Vision et Apprentissage pour l'analyse de scènes

01-10-2021

SL-DRT-21-1071

bertrand.luvison@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Object detectors have been signicantly improved thanks to deep neural network. It is now quite simple to detect on object as far as images with this kind of object in various conditions are available. These datasets exist for some classes (21 classes in PascalVOC, 80 in COCO, etc) but in practive, it is usually necessary to create a specific dataset for classes of interest and then to finetune or adapt an already learnt network. The constitution of this dataset is extremely tedious, time-consuming. Moreover, adding to detect a new class for such model is not flexible at all since you need to fully learn from scratch with the extended dataset. The challenge of this PhD is to free the detector from this constraint of fix set of detectable classes. Such a formulation could be seen as "Zero-shot learning" but this kind of modelling relies on an auxiliary semantic knowledge such as attribut definition. In this way any image can be described through these attributes and semantical definition closer to the one found in dictionnaries can be made event on class of object never seen. The objective of this PhD is not to tackle this problem through "Zero-shot learning" perspective but rather to provide the list of object type, one wants to detect using 3D models. The aim is not build a representative semantic ontology of the world but rather to learn a neural network to find in an image, whichever defined 3D model. In the way old "template matching" methods worked, the aim is now to use the strength of neural network to process this task efficiently on several classes of object simultaneously.

Download the offer (.zip)

Unsupervised monocular moving object detection using spatio-temporal constraints

Département Intelligence Ambiante et Systèmes Interactifs (LIST)

Laboratoire Vision et Apprentissage pour l'analyse de scènes

01-09-2021

SL-DRT-21-1072

camille.dupont@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Traditional object detection techniques based on supervised learning require a large amount of training data. This observation has motivated the scientific community to investigate learning methods with the lowest possible supervision. In this context, there is a growing popularity for self-supervised learning which automatically generates training annotations by exploiting the relationships between input signals. Among the principal tasks that can benefit from learning, we note especially low-level tasks with little semantics such as optical flow estimation, depth estimation from monocular video or pre-learning tasks. The objective of this thesis is to go even further in the exploitation of this learning family in an object detection context. In particular, the work expected in the thesis is to design an object detection model trained from monocular video data which can then be used in the image or video to detect all the object classes present in training data.

Download the offer (.zip)

Deep learning for privacy-oriented user profiling

Département Intelligence Ambiante et Systèmes Interactifs (LIST)

Laboratoire Analyse Sémantique Textes et Images

01-12-2021

SL-DRT-21-1126

adrian.popescu@cea.fr

Artificial intelligence & Data intelligence (.pdf)

A wide majority of online platforms implement a business model in which access to services is free in exchange for the right to exploit user data. User profiles are created from shared data and exploited to personalize the proposed services (search results, recommendations, etc.) and thus improve their effectiveness. Two important downsides of the current business model are that (1) its economic performance depends on the level of detail of the data and (2) profiles are created by the platforms and stored on the server side. This leads to important privacy and transparency problems since users have little control over their shared data and the profiles derived from them. Moreover, intrusiveness will increase with the widespread usage of AI, techniques to infer actionable information from raw user data. Alternative solutions are necessary in order to take into account users' expectations in terms of personal data protection and to comply with the European regulatory framework (GDPR1, Digital Services Act2, Artificial Intelligence Act3). The main objective of this PhD is to propose a privacy- and transparency-oriented user profiling methodology, while maintaining the advantages of online service personalization. The fulfillment of this objective entails the creation and storage of user profiles on their devices. These profiles will never be exposed to the services which exploit them because the personalisation step will also be realized at the edge. The mobilization of recent advancements in several areas of incremental learning is necessary : 1. Semi-supervised learning ? in order to learn effective deep models using a low amount of manually annotated data which are available for the application domains. 2. Transfer learning ? in order to leverage the strength of deep representations learned with large datasets, while also adapting these representations to the applicative context. 3. On-device learning ? in order to infer knowledge uniquely on the users' mobile devices. 4. Multimedia learning ? in order to integrate data of different types (images, vidéos, texts, geolocation etc.) The PhD will build on existing works done by CEA LIST LASTI and will explore axes such as: 1. Exploiting the models proposed by domain experts in order to optimize the performance of deep models learned in a semi-supervised manner. 2. Adapting the local inferences for each user by combining public and personal data for training whenever such a combination is useful. 3. Proposing deep model compression methods which are optimized for different data types and a variety of mobile devices. 4. Providing the users with effective profile understanding control mechanisms in order to increase their trust in the profiling and, ultimately, personalization processes.

Download the offer (.zip)

See all positions