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

PhD : selection by topics

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

AI processing of time series for smart sensors

Département Architectures Conception et Logiciels Embarqués (LIST-LETI)

Laboratoire Infrastructure et Ateliers Logiciels pour Puces

01-10-2020

SL-DRT-20-0261

marielle.malfante@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Today, sensors are used to acquire data of a given modality (acoustics, pressure, image, etc.). Usually, such data are stored before being analysed, for instance with machine learning methods. The relevant information is thereby extracted. A large variety of sensors and use cases can be considered: ? Microphones for the automatic classification of acoustic landscapes, ? Pressure sensors to study deformation and monitor various architectures (brides, dams, wind turbines, etc.), ? Seismometers to detect signals warning of a seisms or of a volcanic eruption, ? Smart watches or bracelets to detect stress phases, ? Etc. The issue of smart sensors consists in creating and designing sensors whose output is the relevant information, straightaway (see Fig. 1). Most of the time the raw signal no longer has to be transferred and stored. Smart sensors are a challenge in numerous fields, typically when sensors have to run autonomously in remote environments, or with limited power and storage access. For instance when studying acoustic landscapes pour environmental monitoring (forest, underwater areas, etc). IoT and wearable sensors are also targeted. Turning a sensor into a smart sensor presents a challenge at many levels. For instance, efficient AI methods to process the data need to be designed, with constraints in term of computation power and energy. Another challenge consists in building those analysis tools from small datasets, or from weakly supervised datasets. CEA is already conducting researches on those issues, and AI based methods are particularly relevant. This PhD subjects focuses on sensors recording time series: IMU, microphones, connected bracelets, etc. The core of the issue is to work on AI methods for time series, in one or several applicative fields. The PhD registers in a larger subject, namely AI reliability (anomaly detection, detection of events of class unseen during the training stage, etc.), but also the development of AI methods under labelling constraint. The topic is ambitious and several approaches are considered.

Download the offer (.zip)

Embedded AI for the semantic interpretation of a probabilistic environment model

Département Architectures Conception et Logiciels Embarqués (LIST-LETI)

Laboratoire Infrastructure et Ateliers Logiciels pour Puces

01-10-2020

SL-DRT-20-0291

tiana.rakotovao@cea.fr

Artificial intelligence & Data intelligence (.pdf)

The perception and modelling of an environment is a major issue when developing autonomous vehicles. How to model the surroundings of a vehicle? How to detect and identify the various obstacles? What about free spaces, and areas safe to drive on? Which sensor combination is the most appropriate to reach an exhaustive description and modelling of the environment? Those questions all have beginnings of answers, but still remain open and not yet solved. There also is a strong constraint regarding the need for embedding systems, which is one of the CEA focuses. Which processing and analysis can be considered while targeting embedded systems? Occupancy Grid is a model used to represent the surroundings of a vehicle and present various advantages. Several sensors of various modalities are used to compute the grid: each modality brings a specific information. For instance, infra-red is efficient by night, LIDAR offers a 360° field of view but is not robust to bad weather conditions, in which case a radar would be preferable. Ultrasound sensors on the contrary are used to analyse very short distances. CEA has developed approaches based on Bayesian fusion to produce SigmaFusion library. SigmaFusion is a tool to fuse the information of different sensors to produce an occupancy grid, which evolves with time. A strong point of SigmaFusion is the computing optimization: the technology is particularly efficient and competitive under strong embedded constraints (low cost integration with low energy consumption on micro-controller certified for critical task for the automotive market). An issue currently addressed is the use of EdgeAI methods to gain a semantic interpretation of an occupancy grid. A typical question is the level of knowledge and interpretation that can be reached while respecting the embedding constraint. Is it possible to detect the object evolving in a grid automatically, in real time and at low energetic cost (pedestrians, cyclists, cars, etc.)?

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-20-0518

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)

Applied formal semantics in hardware compiler frameworks

Département Architectures Conception et Logiciels Embarqués (LIST-LETI)

Laboratoire composants logiciels pour la Sûreté et la Sécurité des Systèmes

01-10-2020

SL-DRT-20-0540

Mihail.Asavoae@cea.fr

Artificial intelligence & Data intelligence (.pdf)

The development of RISC-V instruction set architecture (ISA) is supported by new methodologies and tools which are dedicated to increase the productivity of hardware designs (i.e., high-level design languages and specialized compilation chains). At the language level, Chisel and FIRRTL Hardware Description Languages (HDLs) aim to raise the level of abstraction of hardware design. It thus becomes appealing to formally reason on functional and temporal properties of these high-level designs and rely on appropriate compilation extensions to transfer these high-level properties down to the level of generated Verilog, for example. In this PhD proposal, we target a formal verification framework for computer architectures to support the specification and verification of timing-related safety and security properties. The following two contributions are expected of this PhD thesis: 1) the design and implementation of a verification infrastructure based on formal executable semantics of Chisel and FIRRTL HDLs and 2) the design and implementation of an assertion language to express timing safety and security properties, which are to be verified with the aforementioned formal infrastructure. The scientific contributions of this thesis are expected to evaluated on a selection of the rich-set of architecture designs provided by the RISC-V ecosystem.

Download the offer (.zip)

Ferroelectric Tunnel Junctions (FTJs) for Memory Applications and Ultra-Low Power Neuromorphic Circuits

Département Composants Silicium (LETI)

Laboratoire de Composants Mémoires

01-10-2020

SL-DRT-20-0635

laurent.grenouillet@cea.fr

Artificial intelligence & Data intelligence (.pdf)

The recent discovery of ferroelectricity in hafnium oxide (HfO2) thin films generates a strong interest to save information in a non volatile way for ultra-low power memories, via the application of an electric field to switch the material's electrical polarization. More recently, preliminary results demonstrating HfO2-based ferroelectric tunnel junctions (FTJs) were reported with this CMOS-compatible and scalable material. Here the ferroelectric layer enables to modulate the tunneling current passing through the junction, depending on its polarization. This opens numerous perspectives to those new devices. The objectives of the PhD work will be to fabricate, characterize and model ferroelectric tunnel junctions to better understand the physics of those devices, and then optimize their performance. The optimized devices will then be co-integrated in the form of arrays above complex CMOS circuits to serve as artificial synapses in an ultra low power neuromorphic processor. This work will be performed in collaboration with european partners in the framework of H2020 BeFerroSynaptic EU project.

Download the offer (.zip)

Continual learning for multimodal dataset

Département Architectures Conception et Logiciels Embarqués (LIST-LETI)

Laboratoire Infrastructure et Ateliers Logiciels pour Puces

01-01-2020

SL-DRT-20-0686

marina.reyboz@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Like any embedded systems, edge AI (eAI) connects with its environment, via sensors and possibly actuators. It has to handle a variety of sensor inputs, in a multimodal environment. Even though several artificial neural networks (ANN) already exist, each of them handling one specific modality, there is still a huge challenge to build an ANN for multimodality. In the international state of the art, a spiking neural network classifying images and sound (MNIST dataset +sound) demonstrated better recongition rate and better robustness. The challenge is thus to find a generic approach, able to take State-of-the-Art modality-specific ANNs, and integrate them into a multimodal ANN. Another challenge for eAI is the capability to adapt to a new situation, e.g., a given user or a specific environment. An AI algorithm, even though it has been trained on a large global database, has to adapt. We mamed this property custumisation. The challenge follows: how an ANN, trained on a global database, could be fine-tuned for a specific use-case (e.g., a given user, a specific environment)? From an unimodal bio-inspired model of incremental Learning, the second part of the thesis will focus on coupling multimodal and custumisation aspects.

Download the offer (.zip)

Resistance level modulation in PCM memory for neuromorphic applications

Département Composants Silicium (LETI)

Laboratoire de Caractérisation et Test Electrique

01-03-2020

SL-DRT-20-0740

carlo.Cagli@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Since the last 50 years, the processors are based on the von Neumann architecture and the progress in the integration on a very large scale made it possible to realize this computational architecture on a suitable technological substrate. However, today the miniaturization of electronic components is no longer sufficient to increase performance and reduce the power consumption of conventional architectures. In addition new applications, first of all the artificial intelligence, requires completely new paradighms and calculation approaches. New computational architectures inspired by biology have recently been proposed to overcome these difficulties. The main difference between a neuromorphic circuit and a classical architecture is the memory organization: networks of biological neurons are characterized by co-localization of memory (synapses) and computing centers (neurons). PCM memories are see as strong candidates for emulation of synaptic behavior but the ability to modulate their resistance level is still a challenge. This is a key point to enable PCM cell as synaptic component. This work proposal will start with a characterization phase where the ability of PCM to be modulated in different resistance levels will be scrutinized. The collected data will feed a PCM multilevel model, which is essencial to enable new circuit architecture. In a final stage, innovative circuit can be proposed, based on PCM technology, as proof of concept for neurorphic applications.

Download the offer (.zip)

Transfer Learning and Optimal Transport applied to additive manufacturing process driving

Département Métrologie Instrumentation et Information (LIST)

Laboratoire Science des Données et de la Décision

01-10-2020

SL-DRT-20-0792

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 practical use-cases of interest in the lab, including additive manufacturing. A more detailed presentation of this PhD thesis subject can be found via the following link: https://drive.google.com/file/d/1TmoIYeK9RKRWV7aHFGeqY48tpobuBJeB/view?usp=sharing

Download the offer (.zip)

Few-shot Learning and Domain Adaptation for Information Extraction

Département Intelligence Ambiante et Systèmes Interactifs (LIST)

Laboratoire Analyse Sémantique Textes et Images

01-10-2020

SL-DRT-20-0862

romaric.besancon@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Automated Information Extraction from texts is necessary for all applications that require the processing of large amounts of textual data, but the development of such systems, adapted to a specific domain, is still very costly because it requires to specify very precisely the information to extract or to produce lots of annotated data in order to train machine learning models. The goal of this thesis is to work on new ways to reduce the cost of domain adapation for information extraction systems. The proposed approach is based on two axes: (1) the development of a generic information extraction model, based on a model of semantic frames representing general events that can be made more generic by the use of word embeddings, such as BERT or ELMO and (2) the study of methods to automatically adapt this generic model to a new domain, exploiting only few annotated examples: this adaptation will rely on approaches such as distant supervision, active learning or transfer learning.

Download the offer (.zip)

Optimized coding technics for the design of deep neural network hardware accelerators

Département Architectures Conception et Logiciels Embarqués (LIST-LETI)

Laboratoire Adéquation Algorithmes Architecture

01-09-2020

SL-DRT-20-0869

johannes.thiele@cea.fr

Artificial intelligence & Data intelligence (.pdf)

Artificial neural network based approaches have significantly improved performance in many areas such as classification, segmentation, and so on. The effectiveness of this approach is well established and the number of future applications increases. However, due to their computational complexity and their memory need, these networks are difficult to embed on low power platforms. When porting these networks on embedded platforms, a large variety of hardware constraints have to be taken into account. To overcome these difficulties several research works have produced different techniques that allow to reduce memory and computation footprint of artificial neural networks: reduction of the number of parameters, low precision quantization, etc. This thesis aims to go further into the optimization by working on the data coding. On this thesis, we proposed to explore a new method by working directly on the information coding of the neural network. This coding method would aim to unify two existing coding models: the vector model and the spike model, while keeping in perspective the hardware implementation.

Download the offer (.zip)

Distributed Memory-centric Computing architecture for AI applications, using advanced 3D and NVM Technologies

Département Architectures Conception et Logiciels Embarqués (LIST-LETI)

Laboratoire Intégration Silicium des Architectures Numériques

01-09-2020

SL-DRT-20-0913

ivan.miro-panades@cea.fr

Artificial intelligence & Data intelligence (.pdf)

With the revolution of AI applications, AI algorithms are getting more and more demanding in terms of computing and memory requirements, while it is envisioned that new devices implementing these AI features should be available at the ?edge?, meaning close to the final user (portable devices, automotive, IoT, etc) and not anymore only in the cloud. This implies very strong requirement in terms of memory capacity to enable learning at the edge with compliant energy efficiency. The PhD consists in proposing and exploring new Distributed Memory-centric Computing architecture for AI applications, using advanced technologies to overcome the current issues (memory capacity, memory bandwidth, energy per inference, and learning capability). Recent Non Volatile Memory (NVM) technologies and 3D integration technologies offer dense memory integration while bringing the memory closer to computing cores. The architecture challenge consists in defining the adequate system partitioning, the distributed communication mechanisms, the memory/computing ratios requirements, to in-fine obtain the targeted distributed memory centric computing architecture. The work will consist in architecture and application system exploration through system modeling, and may lead to a testchip to validate the proposed concept. The PhD will take place in an active collaboration between University of Stanford (CA, USA) and CEA (Grenoble, France).

Download the offer (.zip)

Voir toutes nos offres