Giraldo, Luis Gonzalo Sanchez

Computational Learning Intelligence and Perception Laboratory (CLIPLab):  Research Overview and Computational Needs.

PI: Luis Gonzalo Sanchez Giraldo

PhD Students: Oscar Skean, Jhoan Keider Hoyos-Osorio

MSc Students: Santiago Posso-Murillo, Nicholas Lanning

Current Undergraduates: Phillip Chung

Research Overview:

Our goal is to develop algorithms that allow computers to directly extract and integrate multiple sources of information from raw sensory data. To this end, our current research efforts center around the question of how to measure the information content in data without making strong assumptions about its distribution. To make headway, rather than finding new estimators of conventional information theoretic quantities such as Shannon's entropy and mutual information, we are exploring new quantities that preserve or mimic some of the properties of these conventional quantities, but that are better suited for empirical estimation. Such measures have the potential to move forward the fields of machine learning and statistical signal processing providing a principled objectives for unsupervised learning.

Project1: Measures of Information via Representation Learning

In machine learning, the use of conventional information theoretic quantities has been limited by a lack of robust and scalable estimators. Our view is that the fundamental problem is not the estimators, but the quantities we are trying to estimate. Here, we propose novel information theoretic quantities, with similar properties as the traditional quantities, that break through this barrier. This work will develop the theoretical foundations of alternative definitions of entropy and mutual information to address a

fundamental problem in Machine Learning: building objective functions that can handle and

integrate data from multiple sources with minimal supervision.

Personnel:

Jhoan Keider Hoyos-Osorio

Oscar Skean

Nathan Jacobs

Luis Gonzalo Sanchez Giraldo

Computational Methods:

Deep Neural Networks,  Eigendecompostion of Matrices, Mathematical Optimization, Automatic Differentiation.

Software:

Python, Numpy, Pytorch, JAX, MAGMA, Intel MKL.

Software availability: all software is freely available.

Project2: Nonuniform Sampling using Information Theoretic Measures

Current state of the art methods for image classification are usually restricted to limited resolutions. This project studies different strategies to efficiently sub-sample images in a non-uniform manner such that areas that are deemed as more informative for subsequent classification task are preserved with higher resolutions. This problem is great importance in medical image analysis where only specific areas of the image contribute to the decision made by an expert.

Personnel:

Santiago Posso-Murillo

Nicholas Lanning

Luis Gonzalo Sanchez Giraldo

Computational Methods:

Deep Neural Networks,  Eigendecompostion of Matrices, Mathematical Optimization, Automatic Differentiation, Computer Vision.

Software:

Python, Numpy, Pytorch, JAX, MAGMA, Intel MKL, OpenCV.

Software availability: all software is freely available.

Grants:

Luis Gonzalo Sanchez Giraldo (PI), Nathan Jacobs (Co-I) FA9550-21-1-0227 (DEPSCOR) MEASURES OF INFORMATION VIA REPRESENTATION LEARNING. DoD Aug-01-2021 to Jul-31-2024. $598,000