Emerging machine learning (ML) techniques are showing great potential in prediction performance. However, research and development is often conducted in an environment with extensive computational resources and blinded by prediction performance. In reality, computational resources might be contained on constrained hardware where energy and memory consumption must be restrained. Furthermore, shortages of sufficiently large datasets for ML is a frequent problem, combined with the cost of data retention. This generates a significant demand for sustainable ML. With sustainable ML, practitioners can train ML models on less data, which reduces memory and energy consumption during the training process. To explore solutions to these problems, this thesis dives into several techniques that have been introduced in the literature to achieve energy-savings when training machine learning models. These techniques include Quantization-Aware Training, Model Distillation, Quantized Distillation, Continual Learning and a deeper dive into Siamese Neural Networks (SNNs), one of the most promising techniques for sustainability. Empirical evaluations are conducted using several datasets to illustrate the potential of these techniques and their contribution to sustainable ML. The findings of this thesis show that the energy-saving techniques could be leveraged in some cases to make machine learning models more manageable and sustainable whilst not compromising significant model prediction performance. In addition, the deeper dive into SNNs shows that SNNs can outperform standard classification networks, under both the standard multi-class classification case and the Continual Learning case, whilst being trained on significantly less data. / Maskininlärning har i den senaste tidens forskning visat stor potential och hög precision inom klassificering. Forskning, som ofta bedrivs i en miljö med omfattande beräkningsresurser, kan lätt bli förblindad av precision. I verkligheten är ofta beräkningsresurser lokaliserade på hårdvara där energi- och minneskapacitet är begränsad. Ytterligare ett vanligt problem är att uppnå en tillräckligt stor datamängd för att uppnå önskvärd precision vid träning av maskininlärningsmodeller. Dessa problem skapar en betydande efterfrågan av hållbar maskininlärning. Hållbar maskininlärning har kapaciteten att träna modeller på en mindre datamängd, vilket minskar minne- och energiförbrukning under träningsprocessen. För att utforska hållbar maskininlärning analyserar denna avhandling Quantization-Aware Training, Model Distillation, Quantized Distillation, Continual Learning och en djupare evaluering av Siamesiska Neurala Nätverk (SNN), en av de mest lovande teknikerna inom hållbar maskininlärning. Empiriska utvärderingar utfördes med hjälp av flera olika datamängder för att illustrera potentialen hos dessa tekniker. Resultaten visar att energibesparingsteknikerna kan utnyttjas för att göra maskininlärningsmodeller mer hållbara utan att kompromissa för precision. Dessutom visar undersökningen av SNNs att de kan överträffa vanliga neurala nätverk, med och utan Continual Learning, även om de tränas på betydligt mindre data.

Bei der theoretischen Beschreibung von spektroskopischen Experimenten wird in der Regel das Materiesystem quantenmechanisch beschrieben, während das Strahlungsfeld klassisch behandelt wird. Diese semiklassische Näherung ist zur Beschreibung von Experimenten, bei denen eine starke Kopplung zwischen dem Matriesystem und einzelnen Photonen besteht, nicht mehr gültig. Dies kann beispielsweise innerhalb eines optischen Resonators der Fall sein. In dieser Arbeit wird am Beispiel eines Pump-Test- Experiments zum Nachweis des optischen Stark-Effekts untersucht, welche zusätzlichen Effekte sich bei einer quantisierten Beschreibung des Strahlungsfeldes ergeben. Ein signifikanter Effekt ist, dass die Photonenstatistik des Pumpfeldes sich in der Linienform der verschobenen Resonanzlinie widerspiegelt. Weiter wurde in dieser Arbeit bei kleiner Pumpverstimmung ein Verstärkungseffekt gefunden, der ebenfalls auf der quantisierten Behandlung des Strahlungsfeldes beruht (nichtklassische Verstärkung). Es treten ferner bei grosseren Ensemblen von Zwei-Niveau -Systemen zusätzliche Unterstrukturen und Resonanzen auf. Auch kann der Nachweis des optischen Stark-Effekts Aufschluss über die Nichtdiagonalelemente bezüglich der Photonenzahl des quantisierten Pumpfeldes geben.Im Hinblick auf die Beschreibung komplexer Materiesystemen wurde in dieser Arbeit auch eine näherungsweise Berechnung der Testabsorption mit quantisiertem Strahlungsfeld im Rahmen einer Dichtematrixtheorie untersucht. Insbesondere war hier für die quantitative Beschreibung der nichtklassischen Verstärkung eine Berücksichtigung hoherer Korrelationen zwingend erforderlich. Auch wurden näherungsweise Entkopp- lungen unter Berücksichtigung der Erhaltungsgrossen durch- geführt. Die Dichtematrixtheorie wurde auf die Untersuchung des optischen Stark-Effektes an storstellengebundenen Exzitonen in Halbleitern angewandt. Da diese Resonanzen vergleichsweise kleine homogene und inhomogene Linienbreiten aufweisen,ist hier experimentell zu erwarten, dass sich feine Effekte des quantisierten Pumpfeldes bemerkbar machen konnen. / The theoretical description of spectroscopic experiments usu ally relies on a semiclassical approach where the matter system is described in terms of quantum mechanics while the radiation field is treated classically. This approach does n ot work well for systems with a strong coupling between the matter system and photons of the radiation field. The latter can be the case within an optical resonator.In this thesis, additional effects of a quantized radiation field are inves tigated on a pump-probe experiment for detecting the optical Stark effect. One significant effect is that the lineshape of the shifted resonance displays the photon statistics of the pump field. For small pump detuning probe gain results in a frequency regime where the semiclassical treatment predicts absorption. This effect is refered to nonclassical gain. For larger ensembles of two-level systems, additional substructures and resonances appear within the probe absorption spectrum. Also non- diagonal elements of the field density matrix can be detected in such an experiment. In order to describe a more complex matter systems, the optical Stark effect has been treated in terms of a density matrix approach with quantized radiation fields. For a quantitative description of nonclassical gain, higher correlation terms had to be treated properly. Moreover, conserved quantities were taken into account in approximate decouplings. The density matrix approach was applied to the description of the optical Stark effect on impurity-bound excitons in semiconductors. These systems are of high interest as their narrow resonances might allow the demonstration of fine effects of the quantized radiation field.

<p>This thesis is divided into two parts. In the first part we study thermodynamics of the two-dimensional Heisenberg ferromagnet with dipolar interaction. This interaction breaks the conditions of the Mermin-Wagner theorem, resulting in a finite transition temperature. Our calculations are done within the framework of the self-consistent spin-wave theory (SSWT), which is modified in order to include the dipolar interaction. Both quantum and classical versions of the Heisenberg model are considered.</p><p>The second part of the thesis investigates the chemical bonding in solids from the first principles calculations. A new chemical bonding indicator called balanced crystal orbital overlap population (BCOOP) is developed. BCOOP is less basis set dependent than the earlier indicators and it can be used with full-potential density-functional theory (DFT) codes. We apply BCOOP formalism to the chemical bonding in the high-T_c superconductor MgB2 and the theoretically predicted MAX phase Nb3SiC2. We also study how the chemical bonding results in a repulsive hydrogen–hydrogen interaction in metal hydrides. The role of this interaction in the structural phase transition in Ti3SnHx is investigated.</p>

This thesis is divided into two parts. In the first part we study thermodynamics of the two-dimensional Heisenberg ferromagnet with dipolar interaction. This interaction breaks the conditions of the Mermin-Wagner theorem, resulting in a finite transition temperature. Our calculations are done within the framework of the self-consistent spin-wave theory (SSWT), which is modified in order to include the dipolar interaction. Both quantum and classical versions of the Heisenberg model are considered. The second part of the thesis investigates the chemical bonding in solids from the first principles calculations. A new chemical bonding indicator called balanced crystal orbital overlap population (BCOOP) is developed. BCOOP is less basis set dependent than the earlier indicators and it can be used with full-potential density-functional theory (DFT) codes. We apply BCOOP formalism to the chemical bonding in the high-T_c superconductor MgB2 and the theoretically predicted MAX phase Nb3SiC2. We also study how the chemical bonding results in a repulsive hydrogen–hydrogen interaction in metal hydrides. The role of this interaction in the structural phase transition in Ti3SnHx is investigated.

In this work a numerical approach for the determination of the energy spectra and the calculation of thermodynamic properties of magnetic molecules is presented. The work is focused on the treatment of spin systems which exhibit point-group symmetries. Ring-like and archimedean-type structures are discussed as prominent examples. In each case the underlying spin quantum system is modeled by an isotropic Heisenberg Hamiltonian. Its energy spectrum is calculated either by numerical exact diagonalization or by an approximate diagonalization method introduced here. In order to implement full spin-rotational symmetry the numerical approach at hand is based on the use of irreducible tensor operators. Furthermore, it is shown how an unrestricted use of point-group symmetries in combination with the use of irreducible tensor operators leads to a reduction of the dimensionalities as well as to additional information about the physics of the systems. By exemplarily demonstrating how the theoretical foundations of the irreducible tensor operator technique can be realized within small spin systems the technical aspect of this work is covered. These considerations form the basis of the computational realization that was implemented and used in order to get insight into the investigated systems.

In this work we studied magnetic properties of molecular magnets of the new heteropolyanion {Cu20}, dodecanuclear cluster {Ni12}, and the heterometallic {Cr7M} wheels, in which one of the CrIII ions of Cr8 has been replaced by a Fe, Cu, Zn, Ni, ion with this extra-spin acts as local probe for the spin dynamics.Such systems have been synthesized recently and they are well described using the Heisenberg spin Hamiltonian with a Zeeman term of an applied magnetic field along the z-axis. Using the numerical exact diagonalization method, we have calculated the energy spectrum and the eigenstates for different compounds,and we have used them for reexamining the available experimental susceptibility data to determine the values of exchange parameters.We have studied the thermodynamic properties such magnetization, susceptibility, heat-capacity. At low temperature regions molecular magnets act as individual quantum nanomagnets and can display super-paramagnetic phenomena like macroscopic quantum tunneling, ground state degeneracy, level-crossing. A crucial issue for understanding these phenomena is the coupling between magnetic molecular levels and the environment such as nuclear spins. We have modeled the behavior of the proton spin lattice relaxation rate as a function of applied magnetic field for low temperatures as it is measured in Nuclear Magnetic Resonance (NMR) experiments.

No description available.

Cette thèse est consacrée à la description de la physique à une particule ainsi qu'à celle de fluides quantiques bosoniques dans des systèmes topologiques. Les deux premiers chapitres sont introductifs. Dans le premier, nous introduisons des éléments de théorie des bandes et les quantités géométriques et topologiques associées : tenseur métrique quantique, courbure de Berry, nombre de Chern. Nous discutons différents modèles et réalisations expérimentales donnant lieu à des effets topologiques. Dans le second chapitre, nous introduisons les condensats de Bose-Einstein ainsi que les excitons-polaritons de cavité.La première partie des résultats originaux discute des phénomènes topologiques à une particule dans des réseaux en nid d'abeilles. Cela permet de comparer deux modèles théoriques qui mènent à l'effet Hall quantique anormal pour les électrons et les photons dû à la présence d'un couplage spin-orbite et d'un champ Zeeman. Nous étudions aussi l'effet Hall quantique de vallée photonique à l'interface entre deux réseaux de cavités avec potentiels alternés opposés.Dans une seconde partie, nous discutons de nouveaux effets qui émergent due à la présence d'un fluide quantique interagissant décrit par l’équation de Gross-Pitaevskii dans ces systèmes. Premièrement, il est montré que les interactions spin anisotropes donnent lieu à des transitions topologiques gouvernées par la densité de particules pour les excitations élémentaires d’un condensat spineur d’exciton-polaritons.Ensuite, nous montrons que les tourbillons quantifiés d'un condensat scalaire dans un système avec effet Hall quantique de vallée, manifestent une propagation chirale le long de l'interface contrairement aux paquets d'ondes linéaires. La direction de propagation de ces derniers est donnée par leur sens de rotation donnant lieu à un transport de pseudospin de vallée protégé topologiquement, analogue à l’effet Hall quantique de spin.Enfin, revenant aux effets géométriques linéaires, nous nous sommes concentrés sur l’effet Hall anormal. Dans ce contexte, nous présentons une correction non-adiabatique aux équations semi-classiques décrivant le mouvement d’un paquet d’ondes qui s’exprime en termes du tenseur géométrique quantique. Nous proposons un protocole expérimental pour mesurer cette quantité dans des systèmes photonique radiatifs. / This thesis is dedicated to the description of both single-particle and bosonic quantum fluid Physics in topological systems. After introductory chapters on these subjects, I first discuss single-particle topological phenomena in honeycomb lattices. This allows to compare two theoretical models leading to quantum anomalous Hall effect for electrons and photons and to discuss the photonic quantum valley Hall effect at the interface between opposite staggered cavity lattices.In a second part, I present some phenomena which emerge due to the interplay of the linear topological effects with the presence of interacting bosonic quantum fluid described by mean-field Gross-Pitaevskii equation. First, I show that the spin-anisotropic interactions lead to density-driven topological transitions for elementary excitations of a condensate loaded in the polariton quantum anomalous Hall model (thermal equilibrium and out-of-equilibrium quasi-resonant excitation configurations). Then, I show that the vortex excitations of a scalar condensate in a quantum valley Hall system, contrary to linear wavepackets, can exhibit a robust chiral propagation along the interface, with direction given by their winding in real space, leading to an analog of quantum spin Hall effect for these non-linear excitations. Finally, coming back to linear geometrical effects, I will focus on the anomalous Hall effect exhibited by an accelerated wavepacket in a two-band system. In this context, I present a non-adiabatic correction to the known semiclassical equations of motion which can be expressed in terms of the quantum geometric tensor elements. We also propose a protocol to directly measure the tensor components in radiative photonic systems.

Since the synthesis of Mn12, which can be regarded as the birth of the class of magnetic molecules, many different molecules of various sizes and structures have been produced. The magnetic nature of these molecules originates from a number of paramagnetic ions, whose unpaired electrons form collective angular momenta, referred to as spins. The interaction between these spins can often be described in the Heisenberg model. In this work, we use the rotational band model to predict the energy spectrum of the giant Keplerate {Mo72Fe30}. Based on the approximate energy spectrum, we simulate the cross-section for inelastic neutron scattering, and the results are compared to experimental data. The successful application of our approach substantiates the validity of the rotational band model. Furthermore, magnetic molecules can serve as an example for studying general questions of quantum mechanics. Since chemistry now allows the preparation of magnetic molecules with various spin quantum numbers, this class of materials can be utilized for studying the relations between classical and quantum regime. Due to the correspondence principle, a quantum spin system can be described exactly by classical physics for an infinitely large spin quantum number s. However, the question remains for which quantum numbers s a classical calculation yields a reasonable approximation. Our approach in this work is to develop a converging scheme that adds systematic quantum corrections to the classical density of states for Heisenberg spin systems. To this end, we establish a correspondence of the classical density of states and the quantum spectrum by means of spin-coherent states. The algorithm presented here allows the analysis of how the classical limit is approached, which gives general criteria for the similarity of the classical density of states to the quantum spectrum.

Cette thèse par article comprend quatre articles qui contribuent au domaine de l'apprentissage profond, en particulier à l'accélération de l’apprentissage par le biais de réseaux à faible précision et à l'application de réseaux de neurones profonds au traitement du langage naturel. Dans le premier article, nous étudions un schéma d’entraînement de réseau de neurones qui élimine la plupart des multiplications en virgule flottante. Cette approche consiste à binariser ou à ternariser les poids dans la propagation en avant et à quantifier les états cachés dans la propagation arrière, ce qui convertit les multiplications en changements de signe et en décalages binaires. Les résultats expérimentaux sur des jeux de données de petite à moyenne taille montrent que cette approche produit des performances encore meilleures que l’approche standard de descente de gradient stochastique, ouvrant la voie à un entraînement des réseaux de neurones rapide et efficace au niveau du matériel. Dans le deuxième article, nous avons proposé un mécanisme structuré d’auto-attention d’enchâssement de phrases qui extrait des représentations interprétables de phrases sous forme matricielle. Nous démontrons des améliorations dans 3 tâches différentes: le profilage de l'auteur, la classification des sentiments et l'implication textuelle. Les résultats expérimentaux montrent que notre modèle génère un gain en performance significatif par rapport aux autres méthodes d’enchâssement de phrases dans les 3 tâches. Dans le troisième article, nous proposons un modèle hiérarchique avec graphe de calcul dynamique, pour les données séquentielles, qui apprend à construire un arbre lors de la lecture de la séquence. Le modèle apprend à créer des connexions de saut adaptatives, ce qui facilitent l'apprentissage des dépendances à long terme en construisant des cellules récurrentes de manière récursive. L’entraînement du réseau peut être fait soit par entraînement supervisée en donnant des structures d’arbres dorés, soit par apprentissage par renforcement. Nous proposons des expériences préliminaires dans 3 tâches différentes: une nouvelle tâche d'évaluation de l'expression mathématique (MEE), une tâche bien connue de la logique propositionnelle et des tâches de modélisation du langage. Les résultats expérimentaux montrent le potentiel de l'approche proposée. Dans le quatrième article, nous proposons une nouvelle méthode d’analyse par circonscription utilisant les réseaux de neurones. Le modèle prédit la structure de l'arbre d'analyse en prédisant un scalaire à valeur réelle, soit la distance syntaxique, pour chaque position de division dans la phrase d'entrée. L'ordre des valeurs relatives de ces distances syntaxiques détermine ensuite la structure de l'arbre d'analyse en spécifiant l'ordre dans lequel les points de division seront sélectionnés, en partitionnant l'entrée de manière récursive et descendante. L’approche proposée obtient une performance compétitive sur le jeu de données Penn Treebank et réalise l’état de l’art sur le jeu de données Chinese Treebank. / This thesis by article consists of four articles which contribute to the field of deep learning, specifically in the acceleration of training through low-precision networks, and the application of deep neural networks on natural language processing. In the first article, we investigate a neural network training scheme that eliminates most of the floating-point multiplications. This approach consists of binarizing or ternarizing the weights in the forward propagation and quantizing the hidden states in the backward propagation, which converts multiplications to sign changes and binary shifts. Experimental results on datasets from small to medium size show that this approach result in even better performance than standard stochastic gradient descent training, paving the way to fast, hardware-friendly training of neural networks. In the second article, we proposed a structured self-attentive sentence embedding that extracts interpretable sentence representations in matrix form. We demonstrate improvements on 3 different tasks: author profiling, sentiment classification and textual entailment. Experimental results show that our model yields a significant performance gain compared to other sentence embedding methods in all of the 3 tasks. In the third article, we propose a hierarchical model with dynamical computation graph for sequential data that learns to construct a tree while reading the sequence. The model learns to create adaptive skip-connections that ease the learning of long-term dependencies through constructing recurrent cells in a recursive manner. The training of the network can either be supervised training by giving golden tree structures, or through reinforcement learning. We provide preliminary experiments in 3 different tasks: a novel Math Expression Evaluation (MEE) task, a well-known propositional logic task, and language modelling tasks. Experimental results show the potential of the proposed approach. In the fourth article, we propose a novel constituency parsing method with neural networks. The model predicts the parse tree structure by predicting a real valued scalar, named syntactic distance, for each split position in the input sentence. The order of the relative values of these syntactic distances then determine the parse tree structure by specifying the order in which the split points will be selected, recursively partitioning the input, in a top-down fashion. Our proposed approach was demonstrated with competitive performance on Penn Treebank dataset, and the state-of-the-art performance on Chinese Treebank dataset.