In this work, we propose ModelPred, a framework that helps to understand the impact of changes in training data on a trained model. This is critical for building trust in various stages of a machine learning pipeline: from cleaning poor-quality samples and tracking important ones to be collected during data preparation, to calibrating uncertainty of model prediction, to interpreting why certain behaviors of a model emerge during deployment. Specifically, ModelPred learns a parameterized function that takes a dataset S as the input and predicts the model obtained by training on S. Our work differs from the recent work of Datamodels as we aim for predicting the trained model parameters directly instead of the trained model behaviors. We demonstrate that a neural network-based set function class is capable of learning the complex relationships between the training data and model parameters. We introduce novel global and local regularization techniques to prevent overfitting and we rigorously characterize the expressive power of neural networks (NN) in approximating the end-to-end training process. Through extensive empirical investigations, we show that ModelPred enables a variety of applications that boost the interpretability and accountability of machine learning (ML), such as data valuation, data selection, memorization quantification, and model calibration. / Amazon-Virginia Tech Initiative in Efficient and Robust Machine Learning / Master of Science / Also published as Zeng, Y., Wang, J. T., Chen, S., Just, H. A., Jin, R., & Jia, R. (2023, February). ModelPred: A Framework for Predicting Trained Model from Training Data. In 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML) (pp. 432-449). IEEE. https://doi.org/10.1109/SaTML54575.2023.00037 / With the prevalence of large and complicated Artificial Intelligence (AI) models, it is important to build trust in the various stages of a machine learning model pipeline, from cleaning poor-quality samples and tracking important ones to be collected during the training data preparation, to calibrating uncertainty of model prediction during the inference stage, to interpreting why certain behaviors of a model emerge during deployment. In this work, we propose ModelPred, a framework that helps to understand the impact of changes in training data on a trained model. To achieve this, ModelPred learns a parameterized function that takes a dataset S as the input and predicts the model obtained by training on S, thus learning the impact from data on the model efficiently. Our work differs from the recent work of Datamodels [28] as we aim for predicting the trained model parameters directly instead of the trained model behaviors. We demonstrate that a neural network-based set function class is capable of learning the complex relationships between the training data and model parameters. We introduce novel global and local regularization techniques to enhance the generalizability and prevent overfitting. We also rigorously characterize the expressive power of neural networks (NN) in approximating the end-to-end training process. Through extensive empirical investigations, we show that ModelPred enables a variety of applications that boost the interpretability and accountability of machine learning (ML), such as data valuation, data selection, memorization quantification, and model calibration. This greatly enhances the trustworthy of machine learning models.

Au cours de la thèse, nous nous sommes intéressés aux problèmes d'empilement de wafers. Ces problèmes apparaissent lors de la fabrication de processeurs en 3 dimensions. Au cours du processus de fabrication, les puces électroniques doivent être empilées les unes sur les autres. Jusqu'à peu, ces dernières, une fois gravées sur des plaques de silicium appelées wafers, étaient découpées, puis triées afin d'écarter les puces défectueuses et enfin assemblées les unes entre elles.Cependant empiler les wafers plutôt que les puces présente de nombreux avantages techniques et financiers. Naturellement, étant impossible d'écarter les puces défectueuses sans découper la plaque de silice, le problème de la superposition d'une puce viable avec une puce défectueuse se pose. Une pile de puces, étant considérées comme défectueuse si elle contient ne serait-ce qu'une puce défectueuse, la superposition non réfléchie des wafers entre eux mènerait à un rendement désastreux.Afin de générer un nombre minimum de piles défectueuses, une "cartographie" de chaque wafer candidat à la superposition est réalisée lors d'une phase de test, permettant de situer les puces défectueuses sur le wafer. Une fois cette cartographie réalisée, l'objectif est de sélectionner les wafers qui seront assemblés ensembles de manière à faire correspondre les défauts de chacun des wafers.Ce problème peut être modélisé à l'aide d'un problème d'affectation multidimensionnelle. Chaque wafer est représenté par un vecteur comportant autant de composantes que de puces sur le wafer qu'il représente. Une composante égale à zéro matérialise une puce défectueuse tandis qu'un un matérialise une puce viable. Chaque lot de wafers est représenté par un lot de vecteurs. Formellement, une instance d'empilement de wafers est représenté par m ensembles de n vecteurs binaires p-dimensionnels. L'objectif est alors de réaliser n m-uplets disjoints contenant exactement un vecteur par ensemble. Ces m-uplets représenteront les piles. Chaque m-uplet peut être représenté par un vecteur binaire p-dimensionnels, chaque composante étant calculée en réalisant le ET binaire des composantes correspondantes des vecteurs qui composent le m-uplet. Autrement dit, une composante du vecteur représentant le m-uplet est égale à un si et seulement si tous les vecteurs ont cette composante égale à un. Et donc une pile de puces est viables si toutes les puces qui la composent sont viables. L'objectif est alors de minimiser le nombre de zéros ou de maximiser le nombre de un.La thèse comporte deux grandes parties. Une partie théorique abordant la complexité des différentes versions du problèmes en fonction de certains paramètres tels que m, n, p ou encore le nombre maximum de zéros par vecteurs. Nous montrons entre autre que ces problèmes peuvent être utilisés pour modéliser des problèmes plus classiques tels que Maximum Clique, Minimum Vertex Cover ou encore k-Dimensional Matching, permettant de prouver un certain nombre de résultats négatifs que ce soit d'un point de vue de la complexité classique, l'approximabilité ou la complexité paramétrée. Nous fournissons également des résultats positifs pour des cas particuliers du problème.Dans un second temps, nous nous intéressons à la résolution pratique du problème en fournissant et comparant un certain nombre de formulations en Programmation Linéaire en Nombres Entiers. Mais nous nous intéressons également aux performances en pratique de certaines heuristiques à garantie de performances détaillées dans la partie théorique. / In this thesis, we focused in the Wafer-to-Wafer integration problems. These problems come from IC manufacturing. During the production of three-dimensional processors, dies have to be superimposed. Until recent, the dies were engraved on a silicon disk called wafer, then were cut, tested and sorted to suppress faulty dies and lastly superimposed one to each other.However superimposing wafers instead of dies presents several technical and financial advantages. Since faulty dies can only be dismissed when cutting the wafer, superimpose two wafers can lead to superimpose a faulty die with a viable one. In this case, the resulting stack of dies is considered as faulty. It follows that a bad assignment between the wafers can lead to a disastrous yield.In order to minimize the number of faulty dies stacks, a "failure map" of each wafer is generated during a test phase. This map gives location of the faulty dies on the wafers. The objective is then to take advantage of this map to define an assignment of the wafers to each other in order to match as many failures as possible.This problem can be modelized with Multidimensional Assignment problems. Each wafer can be seen as a vector with as many dimensions as the number of dies engraved on it. A coordinate set to zero marks a faulty die while a coordinate set to one indicates a viable one. Each seat of wafers is represented by a set of vector. Formally, an instance of a Wafer-to-Wafer integration problem is represented by m sets of n p-dimensional vectors. The objective is then to partition the vectors into n disjoint m-tuples, each tuple containing exactly one vector per set. An m-tuple represents a stack of wafers. Every m-tuple can be represented by a p-dimensional vector. Each coordinate is computed by performing the bitwise AND between the corresponding coordinates of the vectors that compose the m-tuple. In other words, a coordinate of the representative vector is equal to one if and only if this coordinate is equal to one in every vector composing the tuple. It follows that a dies stack is viable if and only if all the dies composing the stack are viable. The objective is then to maximize the overall number of ones of to minimize the overall number of zeros.The first part of the thesis is a theoretical one. We study the complexity of the considered versions of the problem with regards to natural parameters such as m, n, p or the number of zeros per vector. We show that these problems can encode more classical problems such as Maximum Clique, Minimum Vertex Cover or k-Dimensional Matching. This leads to several negative results from computational complexity, approximability or even parameterized complexity point of view. We also provide several positive results for some specific cases of the problem.In a second part, we focus on the practical solving of the problem. We provide and compare several Integer Linear Programming formulations. We also focus on performances of some approximation algorithms that we detailed in the theoretical part.

No description available.

Combinatorial optimization include such tasks as finding the quickest route to work, scheduling jobs to specialists, and placing bus stops so as to minimize commuter times. We consider problems where one is given a collection of constraints with the objective of finding an assignment satisfying as many constraints as possible, also known as Constraint Satisfaction Problems (CSPs). Most CSPs are NP-hard to solve optimally and we turn to approximations - a solution is said to be a factor-c approximation if its satisfies at least c times the optimal number of constraints. This thesis presents new results on the approximation limits of CSPs in various settings. In ordering CSPs, one is given constraints which specify the relative order of items, and the objective is order the items so as to satisfy as many constraints as possible. We give improved approximation hardness results for two classical problems: it is NP-hard to approximate Maximum Acyclic Subgraph with a factor better than 14/15 and Maximum Betweenness with a factor better than 1/2. We present ordering problems which are NP-hard to approximate better than random assignments, and that there are ordering problems arbitrarily hard to approximate. Next, Gaussian elimination can efficiently find exact solutions for satisfiable collections of so-called parity constraints. We show that whenever constraints accept at least one assignment in addition to a parity, then the problem is NP-hard to approximate better than random assignments. Finally, we study the uselessness property which basically states that if one is given a collection where almost all constraints are simultaneously satisfiable and one is permitted to relax the constraints to accept or reject additional assignments, then it is still NP-hard to find solutions noticeably better than random assignments. We consider the setting where all variables appear unnegated and provide the first examples of non-trivially useless predicates assuming only P != NP. / Kombinatoriska optimering inkluderar naturliga uppgifter såsom att hitta den snabbaste vägen till sitt arbetet, att schemalägga jobb hos specialister, eller att placera hållplatser för att minimera resenärers restid.Vi begränsar vi oss till de problem i vilket man ges en samling vilkor på variablermed målet att hitta en tilldelning av variablerna uppfyllandes så många som möjligt av vilkoren;så kallade Vilkorsuppfyllningsproblem (eng: Constraint Satisfaction Problems, CSPs).De flesta CSPs är NP-svåra att lösa optimalt och vi beaktar istället approximationer. Specifikt kallas, för 0 &lt;= c &lt;= 1, en lösning för en faktor-c approximation om antalet vilkor uppfyllda av lösningen är minst cgånger det största antalet uppfyllda av någon läsning. Denna avhandling består av tre artiklar som presenterar nya resultat begränsande hurpass väl man kan approximera CSPs i diverse situationer.För paritetsvilkor är en samling konsistenta vilkor enkla att lösa genom Gausselimination. Vi visar att för samtliga vilkor som uppfylls av en paritet och åtminstonde en ytterliggare tilldelning så är det inte bara NP-svårt at lösa utan till och med att ge en icke-trivial approximation.Oanvändarbarhet är en stark svårighetsegenskap som i princip säger att det är NP-svårt att ge icke-triviala approximationer även när man gemensamt för alla vilkor får ändra vad som uppfyller dem eller inte. Vi ger de första exemplen på icke-trivialt oanvändbara vilkor utan negationer betingat endast på P != NP.Vi visar på approximerbarhet för diverse ordningsvilorsproblem. I dessa ges man vilkor på hur objekt ska vara ordnade relativt varandra och målet är att hitta en ordning som uppfyller så många av vilkoren som möjligt. Vi ger bättre svårighetsresultat för de två mest välkända ordningsproblem, visar att det finns problem där det är NP-svårt att approximera bättre än triviala strategier, och att det finns ordningsproblem där godtyckligt dåliga approximationsgarantier är NP-svåra. / <p>NADA är en delad institution mellan SU och KTH där senare nu kallar den CSC.</p> / ApproxNP