About AMP Lab        Projects        Downloads        Publications        People        Links

Should Pattern Recognizer's be more human?

The ability of a entity, be it beast or machine, to isolate, categorize and distinguish objects (i.e. patterns) from within the continuum that it exists and/or observes in is at the very heart of what we subjectively refer to as perceptual intelligence. Pattern recognition is central to many aspects of practical computer science and engineering such as biometrics, speech recognition and perceptual intelligence in general. 

 

An open problem that is still being actively researched by many in the pattern recognition community is the inability of classifiers to deal with train/test mismatches which is intimately related to the concept of context. Context is defined as the collection of situations or parameters that meet the assumptions of the trained classifier. The term context is used to describe the concept that a classifier's knowledge (i.e. ability to make confident decisions) is restricted by the context it has been trained under. The difference in context between the train and test sets is referred to as a train/test mismatch. The measure of train/test mismatch is not the physical difference between the train and test observations sets but a measure of how generalized the knowledge gained from the train set is, with reference to the unknown test set.

 

Graphical depiction of train/test (a) match (b) mismatch.

The importance of context and ability to generalise in the creation of a truly intelligent classifier can best be expressed in a paragraph by Dreyfus and Dreyfus (1990) pertaining to the current shortcomings in classifiers compared to truly intelligent entities,

The problem here is that the designer has determined, by means of the architecture of the net (classifier), that certain possible generalizations will never be found. All this is well and good for toy problems in which there is no question of what constitutes a reasonable generalization, but in real-world situations a large part of human intelligence consists in generalizing in ways that are appropriate to a context. If the designer restricts the net (classifier) to a predefined class of appropriate responses, the net (classifier) will not have the common sense that would enable it to adapt to other contexts, as a truly human intelligence would.

In the long term we are actively pursuing improvements to pattern recognition so as to minimize this effect. Currently our research is focused on, but not restricted to, classifier adaptation  and  combination theory.

Related papers:-

• S. Lucey and T. Chen, "Improved speaker verification through probabilistic subspace adaptation," presented at Eurospeech, pp. 2021-2024, Geneva, Switzerland, 2003. [similar technical report]
• S. Lucey and S. Sridharan, "A theoretical framework for independent classier combination," presented at International Conference on Pattern Recognition (ICPR), Quebec City, Canada, 2002. [similar technical report]
• S. Lucey, "Audio-visual speech processing," Ph.D. thesis, in School of Electrical & Electronic Systems Engineering. Brisbane: Queensland University of Technology, 2002, pp. 243. [thesis]

(This page is still under construction!!!)

Top of Research Interests