Adaptive Exploitation of Non-Commutative Multimodal Information Structure
- ARMY MURI grant: W911NF-15-1-0479
- ARO Program Manager: Dr. Hamid Krim
- Prof. Yoram Bresler, University of Illinois Urbana Champaign
- Prof. Todd Coleman, University of California San Diego
- Prof. Alfred O. Hero, University of Michigan
- Prof. Raj R. Nadakuditi, University of Michigan
- Prof. Robert D. Nowak, University of Wisconsin
- Prof. Maxim Raginsky, University of Illinois Urbana Champaign
- Prof. Silvio Savarese, Stanford University
- Prof. Vahid Tarokh, Duke University
A complete roster and web page links may be found at People.
Non-commutativity is intrinsic to emerging complex sensing and processing systems. The performance of a distributed sensor network depends on the ordering or partial ordering of the sequence of information sharing actions taken across the network. Multiuser brain-computer interfaces provide directed channels of neural communication from human to machine and between humans. Understanding human activities from video requires differentiating between different ordered or partially ordered sequences of gestures and actions. Hierarchical models for signals, images, and videos, such as convolutional networks and HDP representations, are directed graphs that may link multiple layers of data types including real-valued scalars, vectors, and matrices; categorical and ordinal variables (ranks); graphs and hypergraphs; and symbolic categories with complicated spatial and temporal co-dependencies. Extracting useful information from such complex data sources is always constrained by limited resources, such as energy or computation power, limited sensing opportunity (i.e., transient visibility), and limited transmission capacity. The powerful toolboxes of convex optimization, information theory, and statistical machine learning have enabled breakthroughs in computer vision, speech recognition, and database indexing, for which algorithms can be designed offine from large amounts of training data. A grand challenge is to move beyond offine static designs to online, sequential, and adaptive designs of data collection and analysis. To do so, we must grapple with the increased design complexity associated with adaptive procedures. Principal among the reasons for increased complexity is non-commutativity of the sequence of sensing actions: the optimal action depends critically on the result of previous actions. The objective of our proposed effort is to address this challenge by developing new tools for the next generation of adaptive online sensing and processing systems.
publications: publications resulting from this program.
The 4th annual review of the MURI will take place on Oct 21 2020.
The meeting will be virtual and will take place in the Zoom room https://umich.zoom.us/j/97652740954.
The high level agenda is below. All times are EDT.
10:00 - Introduction (Bresler, Hero and Krim).
10:05 - MURI overview of accomplishments (Al and Yoram)
10:30 - Technical talks
12:00 - East coast lunch break
12:30 - Technical talks
14:00 - West coast lunch break
14:30 - Technical talks
16:00 - Wrapup (Al and Yoram)
16:15 - Govt Caucus
17:30 - Adjourn
Detailed Technical Agenda
10 AM: Overview of MURI accomplishments (Al and Yoram)
Topic 1: Non-commutative information: representation of uncertainty
10:30 AM: Overview of area (Vahid, Max and Al)
- UM grad student <Bob Malinas> Space-Time Adaptive Detection at Low Sample Support Highlight slide
- Duke grad student <Mohammadreza Soltani> model-free information Highlight slide
- Duke post-doc <Khalil El-Khalil > Fisher-autoencoder Highlight Slide
- UIUC co-PI <Max Raginsky> Highlight slide Slides Recorded presentation
Topic 2: Fusion: multimodal learning in high dimensions
12:30 AM: Overview of area (Yoram, Raj and Todd)
- UM grad student <Wayne Wang> SyGlasso graphical models and clustering Highlight slide
- Stanford post-doc <Roberto Martin-Martin> JRDB - the JackRabbot Dataset and Benchmark Highlight slide
- UIUC Co-PI <Yoram Bresler> Robustness in deep learning-based image reconstruction Highlight slide
- UM grad student <Rishi Sonthalia> Non-commutative imputation of missing data with deep nets Highlight slide
- UCSD grad student <Anjulie Agrusa> regularized Wasserstein clustering Highlight slide
Topic 3: Non-commutative information gathering and decision making
2:30 AM: Overview of area (Rob and Silvio)
- Stanford post-doc <Claudia D’Arpino> reinforcement learning Highlight slide
- UM grad student <Neo Charalambides> coded computing and crowdsourced optimization Highlight slide
- UW grad student <Blake Mason> bandits with connections to coded computing & deep net training Highlight slide
- UW post-doc <Ardhendu Tripathy> representer theorem Highlight slide