A summary of recent courses offered is provided below.
Sparse representations are a foundational tool for modern signal processing and data analysis and have a wide range of applications, including denoising, compression, compressive sensing, machine learning etc. for a variety of signals including speech (audio), images, and video. The course will focus on foundations of sparse signal representations using multi-resolution analysis and wavelet theory (“a fixed basis/representation”). Additionally, as time permits, the course will also discuss data-learned signal representation commonly utilized in deep neural networks. We will also discuss emerging directions at the intersection of these domains, as well as the role of sparsity in the design of efficient deep neural networks. Emerging trends in sparsity for image analysis, that connect structured sparsity with ideas in machine learning and deep neural networks will also be introduced. The course will have a theoretical component as well as a hands-on project component where students will apply these techniques to a real-world image analysis problem.
This course will cover in depth concepts related to Probability, Random Variables, and Stochastic Processes, and will connect them to foundational elements in data science. It is expected that the students thoroughly review the material presented in class, and keep pace with self-study of related topics and by solving practice problems from the textbook. We will focus on developing foundational concepts in class, along with use of examples as needed. In addition to covering the theory behind these foundational concepts, we will link these concepts to applications in data science and signal processing (e.g. Subspace learning, Bayesian inference, Bayesian classification, clustering, denoising etc. - as time permits), and these will be tested through a project based on Python, where students will learn how to apply concepts learned in class on practical problems.
This is an advanced course in deep learning with a special emphasis on Image Analysis tasks. It covers recent advances in deep learning for challenging image analysis applications, including analysis of multi-channel, multi-sensor imagery and learning under challenging conditions, including with limited ground-truth data. The course will cover recent developments in deep learning for image analysis, such as:
Deep Learning with multi-channel imagery
Deep Learning with limited ground-truth
GANs for data augmentation and style/model transfer
Vision Transformers
Self-Supervised Learning
Transfer Learning
Multi-Task and Meta-Learning
The course will involve examples from real-world use-cases, including applications to GeoAI (geospatial imagery acquired from aircraft or satellites). There will be hands-on projects pertaining to the concepts discussed in this course, where students will deploy their understanding of the algorithms on real-world use-cases provided by the instructor. In addition to the lecture materials, students are also expected to read related papers and present them to the class.