When: Wednesday, October 18, 2017
Where: PGH 563
Time: 11:00 AM – 12:30 PM

Multi-task Learning for Commercial Brain Computer Interfaces

Speaker: George Panagopoulos

Conventional machine learning algorithms applied in Brain Computer Interfaces either adapt completely on each specific subject, or do not discriminate between subjects at all. This work examines how a middle solution based on multi-task learning outperforms conventional pooled approaches and produces consistent among subjects pattern extraction that agrees with the field's literature.

Bio:

George Panagopoulos is a second-year PhD student, advised by Prof. Ioannis Pavlidis. His current research interest revolve around sequential machine learning models, multi-task learning and mining bio-signal data.

An Optimization-oriented Framework for Feature Extraction in Flow Visualization

Speaker: Lieyu Shi

Visualizing integral curves of 3D flow datasets is very challenging due to visual cluttering and data complexity. Unsupervised clustering techniques are a good choice to reduce visual occlusion while preserving physically important features. However, current clustering techniques applied in flow visualization is either unpredictable for final feature extraction or intrinsically demanding in terms of time and memory consumption, which are not scalable to large-scale flow datasets. To solve this problem, we propose an optimization-oriented clustering framework based on the conventional k-means clustering, which is incorporated with our linear and differentiable metric for measuring the similarity (or distance) between pairs of integral curves. Our framework aims to effectively extract vortical structures of interest, and to achieve faster convergence than conventional clustering techniques. Besides, our optimization framework is extensible and versatile for other types of feature extraction determined by the design of the proper objective function.

Bio:

Lieyu Shi is a fourth-year Ph.D. student working with Dr. Guoning Chen in physical-based fluid simulation, visualization and analysis.

Update 10/17/17 3:22pm: Soumyottam Chatterjee is unable to present as scheduled. Lieyu Shi will present instead.

The Complexity of Leader Election: A Chasm at Diameter Two

Speaker: Soumyottam Chatterjee

Leader election is one of the fundamental problems in distributed computing. In its implicit version, only the leader must know who is the elected leader. This work focuses on studying the message complexity of leader election in synchronous distributed networks, in particular, in networks of diameter two. For graphs of diameter two, the complexity was not known. In this work, we settle this complexity by showing a tight bound of Θ̃(n) on the message complexity of leader election in diameter-two networks.

Bio:

Soumyottam Chatterjee is a 3rd year PhD student working with Dr. Gopal Pandurangan. He is interested in the design and analysis of algorithms, especially in the domain of distributed computing.