[Group Home]   [Research Statement]   [Projects]   [Publications]   [Members]  

Florida State AI Group

 
Florida State AI Group was founded by Dr. Xiuwen Liu at the Florida State University with the support from the Department and the University. The ultimate goal of our research is to understand the mechanisms underlying intelligence and build intelligent systems that are capable of performing intelligence tasks in real-time under typical unconstrained environments. The current goal of our group's research is to understand how deep learning models (including transformers)are able to generalize on large datasets using theory-guided explorations and develop pratical and effective solutions for engineering, cyber security, and quantum sicence.

Our Approach

Our research activities are centered around efficient representations and inference architectures for computer vision applications. As a concrete goal for the following few years, we want to build a system that is capable of detecting and recognizing 30,000 different objects in real-time. Here 30,000 is an estimate of the human's capacity for basic-level visual categorization given by Biederman [1]. In order to make progress toward our goal of building a machine system, our group recently has focused on the following problems. or real-world problems.

The main research areas of the Florida State AI Group at the Florida State University are broad areas of artificial intelligence, focusing on deep learning models and their applications. Our group believes that the traditional way of from-specific-to-generic methodology does not work because a system specially designed under certain assumptions will fail when those assumptions are not met. Also as shown by many perceptual phenomena, visual perception is nonlinear. This makes many of the tools very critical for linear systems not applicable.

It is widely accepted that a successful intelligent system must have a feedback loop between low-level modules, such as filtering, and high-level modules such as recognition. Our current research projects build on mechanisms of deep learning models, focusing their inherent limitations and how to use reasoning and knowledge to overcome the limitations.


Justifications and Philosophical Arguments

It has been a dream of many ambitious scientists to make a machine which can "see" robustly and flexibly in a natural environment as we human beings do. Toward realizing the dream, our approach is based on our belief that a feasible vision system can only be achieved by efficient approximations of representations and inference architectures as the computational complexity is a key requirement. There are several reasons that support our belief.
  1. The computational capacity of the universe is estimated to 10^{120} operations, known as the Landauer-Lloyd limit [2]. Given this, any algorithm that requires computation beyond this limit is a phantom.
  2. To certain extend, computer vision is not a theoretical problem. For example, one can give an optimal solution based on a look-up table; any intelligent agent can be modeled this way [3], which, of course, does not lead to any .
  3. Most effective computer vision techniques must be unique to vision applications; this is just another way of stating the well known "no free lunch theorem" (p. 456, [4]) and "ugly duckling theorem" (p. 461, [4]).

Reference

[1] I. Biederman, ``Recognition-by-Components: A theory of human image understanding,'' Psychological Review, vol. 94, pp. 115-147, 1987.
[2] S. Lloyd, "Computational capacity of the universe," Physical Review Letters, vol. 88, no. 23, pp. 237901-1 - 237901-4, 2002.
[3] S. Russell and P. Norvig, "Artificial intelligence: A modern approach," 2nd edition, Prentice Hall, 2003.
[4] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classification, John Wiley & Sons, 2001.

[Group Home]   [Research Statement]   [Projects]   [Publications]   [Members]  

Last modified oon June 1, 2024