PhD in

This course focuses on recent research in Natural Language Processing and on developing skills for performing research to advance the state of the art in Natural Language Processing.

This course focuses on recent topics in Natural Language Processing and on developing skills for performing research to advance the state of the art in Natural Language Processing. Specifically, this course will cover fundamentals of LLMs such as Transformers architecture, methods on training and evaluating LLMs via distributed training and efficiency methods, and application in multilinguality, translation and multimodality.

This course provides a comprehensive introduction to Speech Processing. It focuses on developing knowledge about the state of the art in a wide range of Speech Processing tasks, and readiness for performing research to advance the state of the art in these topics. Topics include speech production, speech signal analysis, automatic speech recognition, speech synthesis, neural speech recognition and synthesis, and recent topics in foundation models and speech processing.

This course introduces students to the diverse landscape of biological data, including its types and characteristics and explores the foundational principles of single-cell omics bioinformatics, encompassing key methodologies, tools, and computational workflows, with an emphasis on the development of foundation models for single cell omics data from a research perspective. Single cell omics technologies are a new and fast-growing family of biological assays that enables measuring the molecular contents of individual cells with very high resolution and is key to advancing precision medicine. The course covers essential bioinformatics aspects for working with single cell omics data.

This course provides a comprehensive introduction to Advanced computer vision techniques. The students will develop skills to critique the state-of-the-art computer vision research papers. The course aims at building foundation concepts for modern computer vision as well as developing expertise in several specialized areas of research in computer vision. The following topics will be covered in the course. (i) Deep learning for computer vision (ii) Recent developments in convolutional neural networks and transformers (iii) Advanced techniques in object detection and segmentation (iv) Advanced Vision applications such as medica image segmentation and Remote sensing change detection (v) Development of efficient computer vision architectures (vi) Human centric Vision and (vii) Introduction to Vision Language Models and Diffusion models.

The course exercises an in-depth coverage of special topics in 3D computer vision. The students will be able to critique the state-of-the-art methods on multi-view stereo, 3D reconstruction, 3D shape analysis, 3D deep learning and synthesis, students will have to implement papers to accomplish the following goals: (1) reproduce results reported in the papers, and (2) improve the performance of published peer-reviewed works. This course assumes that the students are familiar with the basic concepts of Computer vision, linear algebra and numerical methods.

This course introduces 3D geometry processing, an important field that intersects computer vision, computer graphics, and discrete geometry. This course will cover the mathematical foundations for studying 3D surfaces from a discrete differential geometric standpoint and present the full geometry processing pipeline: from 3D data capture, mesh smoothing, surface reconstruction, parameterization, registration, shape analysis (correspondence, symmetry, matching), data-driven synthesis, interactive manipulation, to 3D printing. This course will offer practical coding exercises to understand basic geometry processing algorithms and exciting project around data capture and geometry processing.

In the field of computer vision, models have typically been trained to perform well on a specific task or dataset by maximizing performance on a validation set. However, this approach only represents a small part of the types of scenarios that are of interest in real-world applications. In recent years, there has been growing interest in exploring different approaches to learning that can be applied in more diverse and dynamic environments. These approaches, which include lifelong learning, continual learning, meta-learning, transfer learning, multi-task learning, and out-of-distribution generalization, aim to enable models to be more robust, efficient, versatile, and well-behaved in non-stationary settings. This graduate course will focus on these emerging learning paradigms and how they can be applied to computer vision and multimodal learning tasks.

Vision and language encode complementary information and have long been studied together. With the advent of Large Language Models, vision-language models are now more popular than ever and represent one of the most active areas of modern computer vision. This course will cover learning methods and joint models for image and text modalities and will address a wide range of problems including vision-language pretraining, text-based image search, image and video captioning, visual question answering, visual dialog, text-to-image synthesis as well as vision-language navigation and manipulation.

Computer vision/machine learning systems are typically designed to operate under benign scenarios by trusted users. Recently, several studies have shown that computer vision systems have vulnerabilities, which can be exploited by adversaries to compromise the integrity and availability of such systems. These vulnerabilities include both inference-time evasion (adversarial) attacks and training-time poisoning (backdoor) attacks. Many techniques have also been proposed to counter these threats. This advanced graduate course will focus on analyzing these adversarial security threats and potential countermeasures.

This course focuses on building foundations and introducing recent advances in machine learning, and on developing skills for performing research to advance the state of the art in machine learning. This course builds upon basic concepts in machine learning and additionally assumes familiarity with fundamental concepts in optimization and math. The course covers foundations and advanced topics in probability, statistical machine learning, supervised and unsupervised learning, deep neural networks, optimization, reinforcement learning, and causality. Students will be engaged through coursework, assignments, and projects.

This course is designed to explore recent breakthroughs in machine learning and provide students with the necessary skills to conduct research and advance the field of machine learning. It will cover highly specialized topics related to large-scale optimization for real-world problems, including Large-Scale Training of Kernel Methods, Sparse Learning, Bilevel Optimization, Black Box Optimization, and Spiking Neural Networks. Prior knowledge of fundamental concepts in machine learning, optimization, and statistics is assumed.

The study of probabilistic and statistical inference deals with the process of drawing useful conclusions about data populations or scientific truths from uncertain and noisy data. This course will cover some highly specialized topics related to statistical inference and their application to real-world problems. The main topics covered in this course are latent variable learning, kernel methods and approximate probabilistic inference strategies. This course will provide an in-depth treatment to various learning techniques (likelihood, Bayesian and max-margin) and numerous practical complexities (missing data, observed and unobserved confounding, biases) for performing inference.

The course covers advanced topics in continuous optimization, such as stochastic gradient descent and its variants, methods that use more than first-order information, primal-dual methods, and methods for composite problems. Participants will read the current state-of-the-art relevant literature and prepare presentations to the other students. Participants will explore how the presented methods work for optimization problems that arise in various fields of Machine Learning and test them in real-world optimization formulations to get a deeper understanding of the challenges being discussed.

The course covers advanced topics in Reinforcement Learning (RL). Participants will read the current state-of-the-art relevant literature and prepare presentations to the other students. Participants will explore how the presented methods work in simplified computing environments to get a deeper understanding of the challenges that are being discussed. Topics discussed include exploration, imitation learning, hierarchical RL, multi agent RL in both competitive and collaborative setting. The course will also explore multitask and transfer learning in RL setting.

This is a graduate course in a new branch of machine learning: federated learning (FL). In FL, machine learning models are trained on mobile devices with an explicit effort to preserve the privacy of users’ data. FL combines supervised machine learning, privacy, distributed and edge computing, optimization, communication compression, and systems. This is a new and fast-growing field with few theoretical results and early production systems (e.g., Tensor Flow Federated and FedML). This course aims for students to become familiar with the field’s key developments and practices, namely optimization methods for FL and techniques to address communication bottlenecks, systems and data heterogeneities, client selection, robustness, fairness, personalization and privacy aspects of FL. The evaluation of the course heavily relies on students’ paper presentations and the final project selected by the student.

In the past decades, interesting advances were made in machine learning, philosophy, and statistics for tackling long-standing causality problems, including how to discover causal knowledge from observational data, known as causal discovery, and how to infer the effect of interventions. Furthermore, it has recently been shown that the causal perspective may facilitate understanding and solving various machine learning / artificial intelligence problems such as transfer learning, semi-supervised learning, out-of-distribution prediction, disentanglement, and adversarial vulnerability. This course is concerned with understanding causality, learning causality from observational data, and using causality to tackle a large class of learning problems. The course will include topics like graphical models, causal inference, causal discovery, and counterfactual reasoning. It will also discuss how we can learn causal representations, perform transfer learning, and understand deep generative models.

This course is an introduction to the core ideas and theories of statistical learning theory, and their uses in designing and analyzing machine learning systems. Statistical learning theory studies how to fit predictive models to training data, usually by solving an optimization problem, in such a way that the model will predict well, on average, on new data.

This course is an advanced course on algorithms for big data that involves the use of randomized methods, such as sketching and sampling, to provide dimensionality reduction. It also discussed topics such as Sub-space Embeddings, Low rank Approximation, L1 Regression, Data Streams. The course lies at the intersection of machine learning and statistics.

The course focuses on building foundations and introducing recent advances in dimensionality reduction and manifold learning, important topics in machine learning. This course builds upon fundamental concepts in machine learning and additionally assumes familiarity with concepts in optimization and mathematics. The course covers advanced topics in spectral, probabilistic, and neural network-based dimensionality reduction and manifold learning. Students will be engaged through course-work, assignments, and projects.

Training the largest Machine Learning (ML) programs requires petaFLOPs (1015) to exaFLOPs (1018) of computing operations, as well as multiple terabytes (1012) of hardware accelerator memory. Accordingly, 100s to 1000s of these accelerators are needed to satisfy both the computing and memory requirements of the large-scale ML. This course covers systems architecture design, communication strategies and algorithmic modifications required to execute ML training in a parallel and distributed fashion across many network-connected hardware accelerators. In the first part of the course, students will learn a comprehensive set of principles, representations, and performance metrics for parallelizing ML programs and learning algorithms, as well as learn how to compare and evaluate different parallel ML strategies composed out of basic parallel ML “aspects”. In the second part of the course, students will apply these skills to read and critique peer-reviewed literature on parallel and distributed ML systems.

This course will provide students with a deep dive into key issues related to trustworthy and responsible machine learning. Students will learn about adversarial/poisoning and privacy/confidentiality attacks against machine learning systems and defense mechanisms to mitigate them. The course will approach adversarial machine learning through an optimization and game-theoretic framework. The emphasis on privacy-preserving computation for data science and machine learning would be through a formal mathematical notion of privacy called differential privacy. The course will also cover other ethical issues in machine learning such as fairness, safe unbiased and responsible content generation, watermarking for content authentication and ownership verification (provenance), and attacks such as deep fakes, their detection and verification of robustness of relevant defenses (multimedia forensics).

This comprehensive PhD-level course explores the intricacies of modern machine learning, with a specific focus on TinyML, efficient machine learning and deep learning. Through an integration of lectures, readings, and practical labs, students will be exposed to the evolution of TinyML from its traditional roots to the deep learning era. This course will introduce efficient AI computing strategies that facilitate robust deep learning applications on devices with limited resources. We will explore various techniques including model compression, pruning, quantization, neural architecture search, as well as strategies for distributed training, such as data and model parallelism, gradient compression, and methods for adapting models directly on devices. Additionally, the course will focus on specific acceleration approaches tailored for large language models and diffusion models. Participants will gain practical experience in implementing large language models on standard laptops.

The course introduces students to the topic of natural language generation and related emerging topics in the field. The course prepares them to perform research that advances the state of the art in this research area.

This course focuses on recent topics in Natural Language Processing and on developing skills for performing research to advance the state of the art in Natural Language Processing.

This course explores the cutting-edge techniques and methodologies in the field of speech processing. The course covers advanced topics such as Automatic Speech Recognition, Language Modeling and Decoding, Speech Synthesis, Speaker Identification, Speech Diarization, Paralinguistic analysis, Speech Translation & Summarization, Multilinguality and low-resource languages and Spoken Dialog Systems. Students will delve into modern models and frameworks for the different speech tasks. The course emphasizes both theoretical understanding and practical implementation, fostering skills necessary for innovative research and development in speech technologies.

The course introduces students to advanced topics in natural language processing concerning the robustness and trustworthiness of language models (LMs), specifically world knowledge in LMs, safety and inclusivity of LMs, and inner workings of LMs. The course prepares them to perform research that advances the state of the art in these research areas.

PhD thesis research exposes students to cutting-edge and unsolved research problems, where they are required to propose new solutions and significantly contribute towards the body of knowledge. Students pursue an independent research study, under the guidance of a supervisory panel, for a period of 3 to 4 years. PhD thesis research helps to train graduates to become leaders in their chosen area of research through partly-supervised study, eventually transforming them into researchers who can work independently or interdependently to carry out cutting-edge research.

This course will prepare students to produce professional-quality research and solve a practical research challenge in an organization based on an innovative, sustainable, and entrepreneurial research topic. This course will provide exposure to a variety of special topics, research integrity, ethics, organizational challenges, and needs related to various disciplines. Students will design and implement a research project suitable for conference presentation or journal submission relevant to their field of interest, in addition to peer-reviewing a paper. The instructor, and guest lecturers, as appropriate, will present topics necessary to develop well-rounded researchers, innovators, and entrepreneurs in the AI disciplines.

PhD Internship (up to four months)