|
The thing seemed miraculous. It became smarter all the time, figuring out more information from less evidence. It seemed constantly to get better at knowing what to do in ever more complex circumstances …. Science fiction? No! People do this, as our kindergarten through Ph.D. schooling process attests.
As I’ve been pointing out in my last several installments of this column, human intelligence can be a great model for decision-support systems. In this column, I explore some interesting new methods for solving business problems, modeled after some of the techniques the human brain apparently uses to gain its amazing capabilities: using a hierarchical and distributed architecture with information specialists implementing adaptive learning laws.
First, we simply have to trash that antiquated idea of the homunculus, the little person who sits behind your eyes, running the show. Rather, a better analogy for the brain, one that directly links the brain model to information systems, is a group of many information-processing specialists, each with specific tasks and skills.
Second, for the same reason we take marketing 101 before 606, we need artificial neural networks that work hierarchically. The ability to program networks in parts is a big advance — much easier than doing it all at once, which in many cases is just intractable. After all, we don’t learn all about life all at once.
Multiplicity of Mind
Long ago (in 1950), researchers at the University of Chicago completed an amazing study. Doing a factor analysis by hand (computers were very scarce in those days!), they found that human intelligence was made of multiple components; a given skill might make use of several underlying capabilities. Each individual could score a very different IQ on each. For example, there seemed to be a certain capacity underlying arithmetic and spelling skills, and a different one for poetry and higher mathematics. The first might be thought of as a knack for details, the latter, the ability to think with images. Howard Gardener’s now famous book Frames of Mind — The Theory of Multiple Intelligences (Basic Books, 1985), which is just now influencing educational methods, is really a sequel to this study.
From there, the research bloomed. New techniques and technologies have helped reveal that the brain comprises all sorts of specialized units, each with specialized competencies. Each sense, for example, requires many specialized information units to function.
As UCLA professor of psychology Randy Gallistel states: “There are many learning mechanisms or modules. Gould and Marler have called these different mechanisms “instincts to learn.” Each such mechanism has a structure that enables it to compute certain facts about the world. This specialization of computational structure renders the mechanism hopelessly ill suited for computing other facts about the world. For [other facts], one needs other mechanisms, with a different computational structure.”
When the origin of an information stream is complex and multilayered, such as with speech or market prices (see Decision Support, Feb. 9, 2000), you would naturally expect that a wide range of specialized modules would be required to unpack content — in particular, modules sensitive to both current internal and external contexts.
As Harvard psychology professor Alfonso Caramazza points out specifically about language processing: “Crucial among these is the idea that language ability is the result of the activity of many processing mechanisms operating over many differing forms of knowledge. Thus for example, even the simple task of reading aloud a word is thought to implicate a number of distinct types of representations (visual, graphemic, lexical-orthographic, articulatory) and associated processing mechanisms. There are no single brain centers for reading, writing, or comprehension. There are only networks of highly specific mechanisms dedicated to the individual operations that comprise a complex task.” (Both Gallistel and Caramazza are quoted from Conversations in the Cognitive Neurosciences, MIT Press, 1996, ed. Michael S. Gazzaniga.)
But this is not the end of the story. There are important ways in which the parts are integrated.
Seeing Is Believing?
In Figure 1, most people see squares where there are none. This is a rarely problematic by-product of the highly useful way the specialized layers of visual processing interact. There is a layer each for recognizing dots, lines, and planes, with a great many more links feeding back from the higher-level dimensional recognizers to the lower ones than the other way around. With this system, expectations of how the whole might look influence what we see in the parts. Without this kind of processing, humans could not possibly see whole objects when parts are covered up, or put together fragmented or sparse information and yet see or understand a whole. An optical illusion is a very low price to pay.
FIGURE 1 This optical illusion of squares exploits the exceedingly useful top-down guessing nature of the visual faculty. Amid complex and
ambiguous information, intelligence requires constant extrapolation from limited data, even at the expense of sometimes being wrong.
Stephen Grossberg, professor of psychology at Boston University, graciously outlined for me some of the many applications:
Boeing, for example, used this approach in a CAD system to inventory airplane parts. In designing a new plane, you can quickly check if a proposed part is already in the inventory of previously designed parts. Boeing has something like 16 million airplane parts in its inventory, each defined by a million-dimensional vector. It is hard to search.
To solve this problem, they developed a learning, search, and recognition engine based on the human vision model I explained earlier. It can do fast incremental learning that preserves stability in response to arbitrarily large, nonstationary databases, can rapidly select the globally best matching classification to an input from its memory, and give similarity and confidence measures about its choice.
The vision work has been applied by a number of groups to problems in synthetic aperture radar, multi-spectral infrared night vision, laser radars, the design of vision chips, and more. These approaches hold tremendous potential for organizing unstructured text bases, but that will have to be the topic for a future discussion.
Motivated by all this interesting research, I tried out an approach highly analogous to the aforementioned solutions, but with a business problem in mind — one that comes up again and again: How do you deal usefully with a large number of sparsely populated dimensions potentially containing many erroneous values, as might occur in a large survey?
Many OLAP techniques have been implemented to efficiently store and access this kind of data, but it is often unclear which analytic methods would offer the most value in understanding the data.
This kind of problem arises, for example, when questionnaire respondents just do not answer many items. The problem is similar to the one people’s minds conquer by preconsciously filling in missing parts of a whole. Using this capacity, we successfully traverse and live in this complex world with only partial access to essential information.
We see only what is in front of us, cannot read a person’s mind, often do not know the history of a situation, and cannot pay attention to everything that is around us, yet we generally feel like we have things reasonably under control. A great deal of data that e-commerce now generates is similarly spotty, with companies recording cryptic clickstream data and customers opting to leave some form fields blank.
By constructing a hierarchy of neural networks, a net of nets, I was able to quite capably fill in a lot of missing information in a meaningful way. With this technique, I could predict the lifestyles of customers much more readily than imaginable with most statistical and data mining techniques. It seems that, by educating parts of the network first, developing specialists for given parts of the predictive process, and then pulling it all together in the right way, you can see information in the data other methods cannot show you.
Next time, I will delve into the details of my lab results.
Barry Grushkin (bgrushkin@dsslab.com) is the senior lab researcher at the DSS Lab (www.dsslab.com) in Cambridge, Mass.
RESOURCESGrossberg, Stephen: www.cns-web.bu.edu/Profiles/GrossbergLaser radar (LADAR): www.spie.org/web/courses/edcatalog/aero3.html#Statistical Synthetic aperture radar (SAR): www.asf.alaska.edu Vision chips — giving sight to the blind: web.mit.edu/newsoffice/nr/1995/39420.html |
|
|
|
| |||||||||||||||||||||||||||||||
