SAN DIEGO — A software development firm says it has solved one of the hardest problems in artificial intelligence, successfully extracting hierarchical categories from streams of sensory data. And for $50,000, HNC Software Inc. will tell you just how it accomplished that feat.

The company's software, called Cortronics, uses neural networks to model fundamental operations that a person's brain calls on to handle those same tasks.

"We believe that Cortronics' associative-memory neural network technology could be the most powerful and promising approach to artificial intelligence ever discovered," said Robert Hecht-Nielsen, a co-founder of HNC, which provides high-end analytic and decision-management software.

The Cortronics technology was developed under a $3.3 million research contract jointly funded by the Defense Advanced Research Projects Agency and HNC, under the supervision of the Office of Naval Research.

The system, HNC said, can be applied to diverse problems such as extracting a voice stream from a noisy background or identifying camouflaged vehicles on a battlefield.

The brain implementing such an architecture memorizes synchronicity among its various sensor inputs and scores each new sensory experience for similarity to all previous memories. Even at the gigahertz speeds of current-day serial microprocessors, implementing such system-wide associative memories on any usable scale has been difficult. Without the higher-level software that implements the time-based associations, practical applications have not materialized.

HNC said it has solved this software problem, opening the door to genuine machine intelligence. It's now demonstrating its associative-memory neural networks on a ring of 30 Pentiums that can solve intractable AI problems like the classic "cocktail party problem." In that problem, a listener in a room filled with conversation must extract the voice stream of one speaker. Cortronics does this by modeling the "sparse" neural networks of the brain, with only a few connections to nearby neurons.

HNC is offering Cortronics to interested engineers in a three-week seminar priced at $50,000 a seat. It includes hands-on experience and the source code to the 30-Pentium ring running the "brain" operating system.

Engineers will still need to understand the theory behind the problem-solving system to apply it to real-world systems. "What we really hope to gain by sharing our technology with other companies is to find partners for future AI application development," said Hecht-Nielsen.

He predicted that applications developed for the brain OS will be able to instill genuine machine intelligence into next-generation "conversational" applications such as talking automatic teller machines or fully automatic customer-service "personalities" that outperform even a knowledgeable human being in answering ad hoc queries. HNC said it will assist participating engineers in creating their own AI applications during the technology course.

"The cocktail party problem is just a prelude to the kind of AI applications we think are now possible," said Hecht-Nielsen. "We envision all kinds of automated conversational customer services."

Neurons mimicked

Under the hood of Cortronics' solution to the classic "sparse" coding problem is a feature-attractor architecture in which a neural network self-organizes a huge but fixed set of tokens into a universal representation of the data set. Each token in the universal set is associated with only a few other tokens, mimicking the sparse connections among the billions of neurons in the human brain. Every entity in the database then becomes a string of tokens and their associations.

By reinforcing the associations between both spatially and temporally "contiguous" information, the neural net reinforces the connections between often-appearing contiguities in its data stream. As a result, a higher-order amalgamation of often-synchronized features emerges from the topology of the network — namely, "objects" become defined as globs of often-appearing-together tokens.

"The principle is that the components of holistic objects, such as a human face, which is composed of eyes, nose, mouth and so forth, always appear together, so their interconnections in the neural network are reinforced." This lets Cortronics "isolate and excise objects from their background," said Hecht-Nielsen.

This ability to "detect" and "segment" images into objects has been one of the persistent problems of machine vision research. Real-world scenes are confused by deformed and distorted viewpoints, and partial occlusion by nearby objects. Cortronics solves this detect-and-segment problem, HNC said, with its spontaneously appearing holistic objects, which a separate management level of the brain operating system tracks by "paying attention" to significant portions of occluded objects, thereby verifying or falsifying their presence in real-time.

The second type of memory association is association by similarity, and here the Cortronics technology lays claim to genuine machine intelligence. HNC said Cortronics automatically recognizes higher-level objects by logging the similarities of their component parts to form an ascending hierarchy of related objects. With a face, for instance, the Cortronics hierarchical abstractor self-organizes a higher-level face object with the component parts. Likewise, given enough examples it will also self-organize separate categories for male vs. female faces.

The utility of these self-organizing hierarchical abstractions is made plain when an application seeks to identify specific types of objects from real-time inputs. Ordinarily, input data streams contain a mishmash of irrelevant information that often confuse attempts at automatic recognition.

'Find the males'

HNC said Cortronics uses its hierarchy of abstract objects to succeed. Thus, if instructed to "find the males" in a scene, it will tentatively activate all instances of the desired objects and their component parts and will then recognize and "expect" other, related features to appear when the scene changes.

"This ability to make objects of interest pop out from a cluttered background was another holy grail of artificial intelligence," said Hecht-Nielsen, an adjunct professor at the University of California, San Diego, specializing in neural networks and computing. "It's how my students at UCSD helped solve the cocktail party problem. In essence, Cortronics forms expectations about what it should recognize next, enabling it to track a single voice among many by recognizing which sound must have followed from earlier ones."

In the cocktail party problem, where five people are speaking simultaneously, Cortronics creates, on the fly, a list of "next utterances" that could likely follow from the current one, then tracks whichever next word, from five separate voices, matches. This strategy requires that the "first" word be manually flagged as belonging to the voice of interest. Once that's done, Cortronics pays attention to just the selected voice.

"This is the way people's brains work too. To prime the pump with the first word, listeners will move their head slightly closer [to the speaker] to increase the signal-to-noise ratio, but once the first word is recognized, they follow that voice based on their expectations of what next words could logically follow from the current ones," said Hecht-Nielsen.

The Cortronics Technology Course will be held March 25 to April 12 in San Diego.