CISE Seminar: February 28, 2020 – Yasaman Khazaeni, IBM Research AI

Machine learning and deep learning have proven successful in solving a wide range of specialized tasks, often outperforming humans. Modern AI systems, however, are expected to excel at many tasks simultaneously, imitating humans acquiring diverse skills throughout their lifetimes. Classical “end-to-end” learning paradigms fail when presented with multiple sources of data and are asked to learn many tasks at once, even though powerful pre-trained systems for each individual task are available in abundance. We propose to consider a landscape of modular AI systems that can be used as building blocks to assemble a larger AI system, a technique that we refer to as model fusion. AI modules built for different tasks, modalities, and sensors need not be re-trained for each new machine learning problem, but rather can be synthesized into larger systems; a focus on simpler, isolated tasks makes these modular pieces easier to train and test. Beyond simplifying AI for application-oriented users, modular AI enables new perspectives on a broad range of challenges, including federated learning, inference from smaller datasets, and distributed optimization. This talk covers our recent results on fusing models learned from sequestered and heterogeneous datasets. We formalize this task as an instance of a matching problem and develop a general Bayesian nonparametric model fusion framework. We present a corresponding algorithm for learning shared global latent structures by identifying matches among local model parameterizations and discuss several extensions. Our proposed framework is model-independent and is applicable to a wide range of model types. We demonstrate applications ranging from federated learning of neural networks to motion capture analysis via fusion of the hidden Markov models.

Yasaman Khazaeni is currently a research staff member and manager at IBM. She has been working on the application of multi-agent systems in conversation AI focusing on creating control and orchestration algorithms for multi-domain conversational AI in customer care and digital business automation. Her team also focuses on AI planning and Federated Learning research. Yasaman holds an undergraduate degree in both electrical engineering and petroleum engineering from Sharif University of Technology in Iran. In 2009, she continued her graduate studies in petroleum engineering at West Virginia University where she earned her master’s degree in reservoir engineering focusing on the application of Artificial Intelligence in reservoir modeling. She received her PhD in Systems Engineering from Boston University where she focused on multi-agent systems control and optimization with a specific interest in event-driven systems. Her research interests are in the area of mathematical modeling, optimization, operations research, machine learning and statistics.

Faculty Host: Ayse Coskun
Student Host: Mahroo Bahreinian