PhD defense

Deep learning algorithms allow computers to perform cognitive tasks ranging from vision to natural language processing with performance comparable to humans. Although these algorithms are conceptually inspired by the brain, their energy consumption is orders of magnitude higher. The reason for this high energy consumption is both architectural and algorithmic. The architecture of computers physically separates the processor and the memory where data is stored. This separation causes particularly intense and energy-intensive data movement for machine learning algorithms, limiting on-board or low-energy budget applications. One solution consists in creating new neuromorphic architectures where the memory is as close as possible to the computation units. However, existing learning algorithms have limitations that make their implementation on neuromorphic chips difficult. In particular, the algorithmic limitations at the heart of this thesis are catastrophic forgetting and non-local credit assignment. Catastrophic forgetting concerns the inability to maintain the performance of a neural network when a new task is learned. Credit assignment in neural networks is performed by Backpropagation. Although efficient, this algorithm is challenging to implement on a neuromorphic chip because it requires two distinct types of computation. These concepts are presented in details in chapter 1 of this thesis. Chapter 2 presents an algorithm inspired by synaptic metaplasticity to reduce catastrophic forgetting in binarized neural networks. Binarized neural networks are artificial neural networks with binary weights and activation, which makes them attractive for neuromorphic applications. The training process of binarized synaptic weights requires hidden variables whose meaning is poorly understood. We show that these hidden variables can be used to consolidate important synapses. The presented consolidation rule is local to the synapse, while being as effective as an established continual learning method of the literature. Chapter 3 deals with the local estimation of the gradient for training. Equilibrium Propagation is a learning algorithm that requires only one type of computation to estimate the gradient. However, scaling it up to complex tasks and deep architectures remains to be demonstrated. In this chapter, resulting from a collaboration with the Mila, we show that a bias in the estimation of the gradient is responsible for this limitation, and we propose a new unbiased estimator that allows Equilibirum propagation to scale up. We also show how to adapt the algorithm to optimize the cross entropy loss instead of the quadratic cost. Finally, we study the case where synaptic connections are asymmetric. These results show that Equilibrium Propagation is a promising algorithm for on-chip learning. Finally, in Chapter 4, we present an architecture to implement ternary synapses using resistive memories based on Hafnium oxide in collaboration with the University of Aix Marseille and CEA-Leti in Grenoble. We adapt a circuit originally intended to implement a binarized neural network by showing that a third synaptic weight value can be encoded when exploiting the low supply voltage regime, which is particularly suitable for on-board applications.The results presented in this thesis show that the joint design of algorithms and computational architectures is crucial for neuromorphic applications.

Link: https://www.youtube.com/watch?v=kCT6AZqZbRE

Find a clickable link in the attached file below

PhD defense

Composition du jury :

Président : Riad Haidar, professeur chargé de cours, Onera, Supoptique, école Polytechnique

Rapportrice : Christelle Aupetit-Berthelemot, professeure, Université de Limoges, XLIM

Rapporteur : Jeremy Witzens, professeur, RWTH Aachen University

Examinateur : Cyril Renaud, professeur, University College London

Examinateur : Guang-hua Duan, docteur, 3SP Technologies

Directeur de thèse : Laurent Vivien, directeur de recherche, CNRS, C2N

Encadrant industriel : Jérôme Bourderionnet, docteur, Thales Research & Technology

Abstract :

One of the key technological issues in the development of future wireless communications network like 5G is the settlement of the new frequencies higher than the classical sub-6 GHz bands, and called millimeter waves (above 30 GHz). Indeed, current frequency bands are close to saturation due to a high number of users and mm-waves represent a very wide unused bandwidth available for 5G and beyond. However, at these frequencies, electronic sources are very expensive, have a high-power consumption and occupy a lot of space, which make them unsuitable for embedded and remote applications. As a replacement optical coherent technology is in first position due to its high compatibility with the already fibered network while requiring very few elements at remote sites, like antennas for instance. Furthermore, the maturity level reached by photonic integrated circuits is now at a decisive moment when they can be used massively within coherent transmission components. This being combined with a downscale of antenna size, due to lower wavelength in millimeter wave, allows new architectures to introduce beamforming and spatial division multiplexing to increase network fronthaul capacity.

Among all optically assisted millimeter wave generation methods the one we deal with in this work is optical phase-locked loop, which compatibility with standardized photonic integrated circuits is also investigated. The working principle relies on an electronic feedback loop making optical phase variations of two independent lasers correlated, and their resulting beat note being then artificially coherent. Thus, the latter is very easy to use in order operate optically assisted frequency up- and down-conversion of a millimeter wave signal. This method benefits from strong assets like its high available optical power and its flexibility toward complex photonics architectures which makes it very suitable to supply photonic integrated circuits dedicated to signal processing, as beamforming networks for instance.

In this work we developed an optical phase lock loop based on commercially available fibered components in order to demonstrate the viability of using semiconductor lasers for coherent applications. This loop has then be implemented in several transmission schemes in the K-band (20-30 GHz) for future 5G experiments, all within the blueSPACE European project. This has led us to investigate the effect of the loop phase noise on QAM modulation formats and OFDM method, both being standards for 5G and beyond. In parallel we designed an optical phase lock loop photonic integrated circuit, which fabrication has been subcontracted to a commercial indium phosphide foundry, in order to evaluate the maturity of the process towards high complexity devices. This semi-industrial approach is validated by state-of-the-art performances for that kind of circuits.

Thèse soutenue en français

Link: https://us02web.zoom.us/j/87121400526?pwd=NjBvSDRNV2xLWnhHSXVaVk5SUGM2Zz09

Passcode: 352864

Find a clickable link in the attached file below

Légende : Description of the designed optical phase-locked loop implemented on an indium phosphide photonic integrated ciruit and delivering a locked beating tone in the mm-wave frequency range.

PhD defense

Jury members :

Camille-Sophie Brès, Associate professor, EPFL : Rapporteure, examinatrice

Arnaud Mussot, Professeur des universités, Université de Lille : Rapporteur, examinateur

Béatrice Dagens, Directrice de recherche, C2N : Examinateur

Bart Kuyken, Associate professeur, Université de Ghent : Examinateur

Thibaut Sylvestre, Directeur de recherche, Université Bourgigne Franche-Comté : Examinateur

Sébastien Tanzilli, Directeur de recherche, Université Côte d'Azur : Examinateur

Quentin Wilmart, Ingénieur chercheur, CEA LETI : Invité

Laurent Vivien, Directeur de recherche, Université Paris-Saclay : Directeur de thèse

Abstract :

Nowadays, communication networks are becoming more and more important, with the massive emergence of social networks, "streaming" platforms, or even the "Internet of things" for example. In this context where the demand in terms of communications continues to grow, it is necessary to find new technological solutions to overcome the limitations posed by current systems. The emergence of silicon photonics meets all the requirements relating to these needs: this technology allows the creation of ultra-fast communication systems (thus making it possible to transmit more data on the same channel) and reduces the energy consumption of the components of the entire transmission chain. In addition, the use of silicon as the basis for these new systems makes it possible to take advantage of the great maturity of manufacturing processes already established in the microelectronics industry.
During my thesis work, I dedicated my efforts to the exploration of nonlinear optical effects in photonic components compatible with silicon technology. The first objective was to study the possibility of exploiting Pockels effect (a second order nonlinear effect) in silicon to produce ultra-fast and power-efficient electro-optical modulators. Due to the centro-symmetry of the crystal lattice of silicon, this effect is absent in this material. The solution envisaged in this work is to apply a stress on the waveguide in order to deform it and therefore break the crystalline symmetry. Thus, Pockels effect becomes theoretically possible. Modulation experiments in static regime as well as in hyper-frequency regime were carried out. The results show the difficulty of separating the different effects responsible for the observed electro-optical modulation. Indeed, semiconductor effects behave similarly to the Pockels effect, making interpretation of experimental results particularly difficult. Numerical models were set up to establish the contribution of each effect in the electro-optical modulation results.
The second axis approached in my thesis concerned the realization of broadband light sources that can work in classic telecommunication windows. Silicon nitride waveguides (a material compatible with silicon technology) were excited with a pulsed laser. By the combination of nonlinear effects and dispersion during propagation, the initial spectrum of the source is drastically broadened, forming what is called a supercontinuum. This type of source is very interesting for producing broadband frequency combs, useful for advanced communication systems, but also in spectroscopy, in imaging, or for chemical and biological sensors for example. In my work, several platforms for manufacturing silicon nitride waveguides were studied. Each platform showed extremely low propagation losses, and their nonlinear properties are compared through pulsed laser light experiments. The generation of supercontinuum as well as two-photon absorption were analyzed in order to conclude on the nonlinear performances of these platforms.

Link: https://us02web.zoom.us/j/87102803431?pwd=T0cxTVVlQUltWGJOOTBTcUNWUzd5UT09

 Find a clickable link in the attached file below

PhD defense

Quantum dots are a key building block for quantum technologies as they are able to generate single photons or entangled photon pairs. The deterministic fabrication of sources based on quantum dots in cavities paves the way towards scalability: we show for the first time that it is possible to obtain homogeneous properties, by studying in detail 15 of our sources. Each characteristic of the quantum dots' emission is measured with a specific setup. Among them, the indistinguishability is accessible through the visibility of the Hong-Ou-Mandel interference. We study how this latter quantity is affected by the presence of additional photons, depending on the nature of that noise. We thus derive a formula to deduce the true single-photon indistinguishability from the measured visibility and apply it to our sources. Finally, the properties of the emission from quantum dots is deeply related to their charge state and symmetry. We propose a new way to control the exciton's symmetry by applying three voltages via a specific structure. This experimental study opens the way to the reproducible generation of entangled photon pairs with a high brightness as well as frequency-encoded qubits.

Jury composition

Rosa TUALLE-BROURI,  IOGS, Université Paris-Saclay, Présidente

Carole DIEDERICHS, Sorbonne Université , Rapportrice

Jan SUFFCZYNSKI,  University of Warsaw , Rapporteur

Jean-Michel GERARD, CEA Grenoble, Examinateur

Rinaldo TROTTA, Université La Sapienza de Rome, Examinateur

Pascale SENELLART, CNRS-C2N Palaiseau, Directrice de thèse

PhD defense

Jury members

M. Alfredo DE ROSSI, Directeur de thèse, Université Paris-Saclay GS Physique

M. Philippe BOUCAUD, Rapporteur, CNRS - CRHEA / Université Côte d'Azur

M. Dario GERACE, Rapporteur, Università di Pavia

M. Allard MOSK, Examinateur, Universiteit Utrecht

M. Xavier CHECOURY, Examinateur, Université Paris Saclay

Mme Christelle MONAT, Examinatrice, Ecole Centrale Lyon

M. Fabrice RAINERI, Co-encadrant de thèse, Université de Paris

M. Sylvain COMBRIE, Invité, Thales Research and Technology 

Abstract

All optical signal processing could drastically reduce the power consumption and increase the data rates allowed by its electronic counterpart. This approach requires the integration of multiple photonics systems on a chip. Components exploiting parametric interactions can perform various tasks such as wavelength conversion, amplification, sampling and switching as well as provide non-classical state–of-light for quantum information. Nonlinear micro-resonators are attractive in order to reduce the footprint of the component and the power required to activate the nonlinear effects. Photonic crystals (PhC) cavities seem to provide a very interesting platform due to their ability to strongly confine light in a close to diffraction-limited volume. However, due to the difficulty to engineer the dispersion of the photonic crystal cavity modes, the nonlinear efficiencies have remained well below the theoretical maximum allowed by these structures. Based on the previous work realized by the research group, the aim of this thesis is to harness the potential of nonlinear PhC cavities.
The cavities studied in this work are by design the optical analogous of the quantum harmonic oscillator. When the photons are submitted to a parabolic electromagnetic potential, the modes of the cavity are equally spaced in frequency, which matches the energy conservation requirement for parametric interactions. However, the linear characterization of these resonators shows that the structural disorder induces a deviation on the targeted frequency that requires a compensation technique.
To tackle this issue, a thermal tuning process that exploits the inhomogeneous spatial distribution of the electromagnetic modes is introduced. It allows to systematically put the cavity in a triply resonant configuration. This tuning technique is used to observe stimulated and spontaneous four wave mixing with record efficiency. Parametric oscillation in a PhC cavity is demonstrated for the first time in a sample with higher quality factors.
A second platform is also developed, based on the hybrid integration of a PhC cavity on a silicon circuitry, with a very low footprint and high integration capability. Efficient four wave mixing is again observed using the same tuning technique.  

Lien visio : https://u-paris.zoom.us/j/88469313083?pwd=cE5OTmdyS0lOZlVld0ROL21tOTNSUT09

Figure 1: First demonstration of parametric oscillation in a photonic crystal cavity

PhD defense

Composition du Jury
Sonia GARCIA BLANCO, Professeur, University of Twente, Rapporteur
Aline ROUGIER, Directrice de Recherche, CNRS, (ICMCB), Rapporteur
Odile STEPHANE, Professeur, LPS- Université Paris-Saclay, Examinatrice
Blas GARRIDO Professor, Universitat de Barcelona, Examinateur
Stefan ABEL Co-CEO and Co-founder, Lumiphase Corporation, Examinateur
Laurent VIVIEN, Directeur de Recherche, CNRS, C2N, Directeur de thèse
Sylvia MATZEN, Maître de Conférences, Université Paris-Saclay, Co-encadrante
Thomas MAROUTIAN, Directeur de Recherche, CNRS, Invité

Abstract
Les oxydes transparents avec des indices de réfraction modérés sont ciblés pour l'intégration hybride en raison des avantages. L'adaptation de réseau à l'aide d'une couche avec paramètre de maille entre le silicium et l'oxyde fonctionnel est une solution élégante pour faire croître par épitaxie des films d'oxyde fonctionnel cristallin de haute qualité. Cette technique nécessite d’un buffer pour visualiser les interfaces de films minces sans défaut, la diffusion de la lumière et la densité des défauts sont maintenues au minimum. La zircone stabilisée à l'yttria (YSZ) est un oxyde buffer bien connu pour d'autres oxydes fonctionnels. Parmi ses propriétés, il faut noter : la stabilité chimique, la transparence du visible au moyen-IR, un indice de réfraction autour de 2.15, ce qui rend cet oxyde fonctionnel intéressant pour le développement de guides d'ondes à faibles pertes lorsqu'il est cultivé sur un substrat à faible contraste. Après avoir démontré faibles pertes de propagation de 2 dB/cm à une longueur d'onde de 1480 nm dans des guides d'ondes de YSZ hautement monocristallins, nous avons envisagé d'explorer ce matériau pour l'amplification optique. Dans cette thèse, nous explorons le dépôt de couches minces YSZ dopées Er3+ par dépôt de couche pulsée (PLD) fournissant une luminescence en correspondance avec la bande C de la fenêtre de télécommunication (λ = 1,5 μm) et dans le domaine du visible de longueurs d'onde. La caractérisation structurelle et optique conduit à un matériau hautement optique efficace à des longueurs d'onde de 1530 nm. Après la caractérisation structurel et optique, nous avons intégré notre matériau comme cladding sur des guides d'ondes SiNx. Les pertes de propagation et l'amplification optique à 1530 nm en utilisant différentes techniques de pompage sont discutées dans cette thèse.
Language defense : Anglais

Lien visio: https://us02web.zoom.us/j/83238551183?pwd=M002Z05nU2RwUW9qMVVvem0xY3ZGUT09

Find a clickable link in the attached file below