Eldorado Overview

We are using Alchemy/Goal Mind to conduct internal research and development in the domains of Information Extraction, Smart Environments and Robotics. This research consists of a three level approach:

• A Unified Cognitive Theory (now called IMPISH)
• Cognitive Models based on this IMPISH theory
• Multi-agent Environments based on these models

IMPISH Theory

Hebb suggested the idea of brain areas in the 1940's. Almost all current neuroscience and cognitive modeling research is based on and supports the idea that the brain's neocortex contains areas of processing that support distinct cognitive functions such as language understanding, visual perception, motor skills, etc. Both traditional symbolic AI and connectionism accept the resulting ability of explanatory cognitive models to 'divide and conquer' brain (or mind) functions into smaller subsets of cognition based on this idea of brain areas, but both approaches have had difficulty in extending simple models into more complex unified models of cognition. Connectionism cannot easily lead to general cognition due to the computational complexity of large ANNs. The lack of a unifying principle form modeling often causes symbolic approaches to develop ad hoc sub-solutions that do not lead to more general models of cognition at the depth needed to address real-world problems.

We have developed a possible solution to these problems by creating a unified cognitive theory based on both a brain and mind view. The brain view controls the communication method used between sub-models (called components in the Alchemy/Goal Mind environment). The mind view controls how the components work internally.

Our research in cognitive modeling has spanned over 15 years. During this time, we have developed a series of three unified cognition theories. Our first approach predates our existing Alchemy/Goal Mind products. This AMEBA theory was used to explore the basic brain/mind model still used within our modeling environment. However, research with the AMEBA theory demonstrated that the resulting models were too sequential in their processing methods. The Interlaced Micro-Pattern (or IMP) theory was developed to address a more parallel modeling approach that better represents the MPP capabilities of the brain. The Interlaced Micro-Pattern with Integrated Sub-symbolic Heuristics (or IMPISH) theory was later developed to fix internal activation problems with the IMP theory.

Both the AMEBA and IMP theory were based of the following logic:

1. If the function of a brain area can be represented using the standard computational model of Input-Process-Output (or IPO),
2. Then, if we can emulate the input and output stimuli of a brain area,
3. Then, the internal process method is not important unless the method does not allow us to do some necessary conversion of input to output in some required timeframe.
4. If any problem with the processing time can be 'fixed' either by changing the processing method or the scale of the process,
5. Then, the internal inference method used in a model's component can be viewed as a valid explanatory representation of a given brain area as long as it provides the correct output with the correct time delay to all other components in the model.

Using this logic, given that the basic structure of a model correctly represents its related brain areas, any non-explanatory action within a model can be modified to create an explanatory action by either changing the component's P (processing) or dividing the component's IPO into smaller IPOs supported by a new set of components all tied together (by a SRN) to look identical to the original component. The communication and process distribution support of Alchemy (based on the AMEBA theory) makes this IPO redistribution seamless. The communication architecture supported by Goal Mind's SRN component (also based on the AMEBA theory) further limits the effect of this redistribution on the model's other components by restricting the broadcast range of any new inter-component communication (or stimuli) need by these new components. However, by using a symbolic approach to inference within a component, there is a limit to the how many times the processing element within a component can be divided due to the related size of its knowledge representation (and that KR's overhead).

The IMP theory proposed a way to cut through this knowledge representation limit by using micro-patterns instead of patterns and interlacing these patterns to form larger (virtual) patterns over multiple components within a model. These micro-patterns also allowed a fuzziness within the resulting knowledge which allowed a more parallel approach to complex cognitive abilities like language processing which needs to address at some level the syntax, semantics and conceptual aspects of both language understanding and generation, as well as the relationships between different levels of structure like morphemes, words, phrases, and discourse.

The Hebbian brain view has always had a problem with explaining how some output of an area does not seem to be related to any of its input. While the IMP theory worked well in our LEAP cognitive model, it failed in our early work with the SILK model for just this reason. The IMPISH theory was developed to correct this problem by adding the ability of sub-symbolic activations within the pattern network. Without going into all of the details, our method of doing this is similar to the sub-symbolic method used in the ACT-R theory to activate the facts in the factbase that a rule can 'see' at any given time. However, IMPISH based models function in the completely opposite direction to ACT-R models with the 'rules' creating the activations as a side-effect of using a fact that has been activated by an interlaced micro pattern. (There is also quite a bit of difference in the whole inference approach used within Alchemy/Goal Mind and ACT-R so this comparison is really only helpful if you have a detailed understanding of ACT-R.)

Cognitive Models

We are currently exploring five cognitive models:

• LEAP - Language Extraction from Arbitary Prose
• MOVE - Motive Orientation to a Visual Environment
• REAP - Reference-augmented language Extraction from Arbitary Prose
• SILK - Speaker Identification using Language Knowledge
• SPOT - Spatial-concept Perception and Object Tracking

Agent Environments

We are currently exploring three cognitive-based agent environments:

• MIST - Multi-agent Intelligent Search Tool
• SAGE - Smart-environment Adaptive-agent Generation Environment
• SALT - Smart-environment Adaptive-agent Langauge and Tasking