The Enigma of SimilaritySimilarity judgments lie at the very heart of good decision-support capabilitiesContinued from Page 1 In contrast, the image identification software developed at Los Alamos National Laboratories emphasizes finding the right input data transformations — optimizing the senses. It combines a simple head algorithm — discriminate analysis — with a sophisticated genetic-programming-based, image-processing algorithm. In genetic programming, populations of programs mutate and breed, with the most successful at solving a target problem surviving. In the Los Alamos system, a whole range of image filters and operations are swapped in and out and combined. Given any object, this approach can rapidly learn the key features that can systematically pick out that object (for example, the infrared reflectance of a given boat being watched or the ultraviolet absorption and texture signature of ice clouds or opium poppy fields). Genalytics Inc.'s GA3 software can help companies determine which customers will respond to an offer by applying genetic programming techniques to CRM and marketing applications. Its head algorithm is logistic regression done genetically. In this product, the sense search involves looking at statistical and linear transformations of variables. Together, using large clusters of CPUs, Genalytics searches for the best statistical model to predict some issue of interest to marketers, such as responsiveness to a type of advertising. MD Online's (MDOL's) Image Match software is a visual search engine that stands to revolutionize the medical image diagnosis process. The idea was developed by Michael E. Leventon, a founder of the company, in his Ph.D. thesis at Massachusetts Institute of Technology. Because Image Match knows, in spatially context-dependent terms, the prototype features likely needed for correct matching or differentiation, it is an example of the knower computational principle. A doctor can take any X ray or patient scan and instantly find images with closely similar pathologies that have already been diagnosed. (See Figure 2.) To accomplish this process, raw data is transformed to produce brightness, texture, and contrast feature vectors for many small windows over the image. Using training and validation samples from the database, these feature vectors are differentially weighted over the surface of the image to optimize search match precision. Because each pathology has specific areas in which it's likely to occur, the important issues for identifying any given pathology or differentiating it from other pathologies are preloaded into the relevant regions. Emphasizing pathological regions automatically during searches provides doctors with the information needed to efficiently find correct matches. MIT Lincoln Laboratory's image-mining system, a remote multisensor 3D-image fusion technique developed by Allan Waxman and William Streilein, can keep track of objects independent of light, time of day, and weather changes by knowing when things that look similar are really the same. The system first performs the image transformations identical to the multilayered hierarchical pattern enhancement mechanisms used in the cortex of humans and other animals. The results are stunning. With this system, you can see just as well at night with only a few stars visible as you can in bright daylight. The head on this system includes knower capabilities. A real-time learning neural net system called fuzzy ART MAP is used to produce a real-time object feature learning system. Unlike the Los Alamos system, it can learn and differentiate multiple objects in real time. By marking the system with a pen, you can train it to recognize an item in a few seconds. (See Figure 3.) It can then automatically track an item, such as a ship picking up arms in Iran, and monitor it unaided until it reaches the Red Sea. Next TimeThe capacity to recognize object similarity is clearly not trivial. As I will show in my next column, what you've learned from these principles can help you answer some deep philosophical questions, as well as show how to create the next breakthrough decision-support capabilities. Barry Grushkin [BLG23@Cornell.edu] is chairman and CTO of The Machine Intelligence Development Co., a group specializing in sophisticated data mining and business process optimization. RESOURCESGenalytics Inc.: www.genalytics.com Los Alamos National Laboratories: www.lanl.gov MD Online (MDOL): www.mdol.com MDG Ltd.: aiisreal@math.tau.ac.il MIT Lincoln Laboratory: www.ll.mit.edu
|
 |
Most Popular This Week
IE Weekly Newsletter
Subscribe to the newsletter
| |
| ||||||||||||||||||||||||||||||||
