Artificial Intelligence and Machine Learning: Part 5: Complexity

In this series I am writing several blog articles about Artificial Intelligence. This week’s article, the last in the series, introduces the CORE toolset that can be used by the Government under an existing license to meet most of its AI/ML needs.

In 1995 Bevilacqua Research Corporation was awarded a Small Business Innovative Research (SBIR) contract by the Army Threat Systems Management Office. The purpose of the contract was to develop an AI/ML approach that could be used to simplify and improve the performance of threat systems models being used in distributed simulation environments. Because at the time we were one of the few companies actively working in the field of AI/ML we decided to collect the tools we had developed to create our algorithms into a toolkit that could be used by our DoD customers. This collection of tools which we called the CORE toolset was delivered to the Government in 1995. Since that time the toolset has been used successfully to add AI/ML capabilities to many applications across the services. These include:

  • Advanced human behavioral models
  • Threat System Models
  • Intelligent MTI trackers
  • Intelligent Battle Damage Assessment tools
  • Lossless, Ultra-High Compression Communications
  • Autonomy
  • Decision Support

The foundational work has been ongoing at BRC for more than twenty-six years with the development of this innovative toolset. AI/ML models built using the tools mimic the operation of the current understanding of how the human brain itself is organized to collect, store and add knowledge elements to prior knowledge in long-term memory. Figure 1 shows a quote from the DOD Techlink program after they assessed the current BRC machine learning tools last year (Michael Wallner, personal communications 6, 2017).

Figure 1: Recent DOD Techlink Office Assessment of the Existing BRC CORE/ML tools

The knowledge storage mechanism within CORE is based upon the use of conceptual graphs that mimic the connectionist view of schema development within human long-term memory. Developed by cognitive psychologists in the 1950’s (Schmidt, 1975) this mature knowledge storage mechanism is completely non-brittle allowing knowledge elements to be added (learning) or subtracted (forgetfulness) without affecting current knowledge. Once a knowledge base has been established the understanding gained from that knowledge from a specific observer’s point of view must be defined. For example a bridge over a river means “river-crossing” to a military analyst but the same knowledge elements mean “dangerous hazard” or “blockage” to a water skier. This has always been the biggest problem for adaptable knowledge systems…i.e. how do you tell a knowledge base what a newly added knowledge element means in terms of the understanding of the current observer? BRC has solved this problem through the use of a physics-based approach that uses adaptive resonance theory to pre-define element understanding in “knowledge clusters” that are related in multiply dimensioned “prior knowledge space”. By looking at where in this space a new knowledge element falls, the model itself can determine how and where a new knowledge element should be connected in the existing conceptual graph-based schema. This can be done without human intervention or supervision. Figure 2 shows an example of how an end-to-end Machine Learning architecture built using CORE would look.

Figure 2: Example of A Full Machine Learning Architecture Built Using the CORE/ML Tools

One of the exciting things about the use of the ART neural network is that it decomposes knowledge into a set of vectors that can be represented in n-dimensional knowledge space. This is essentially a “gisting” process that creates clusters of “similar” knowledge elements. The inclusion of a new knowledge element within an existing cluster tells the system that the new information is dimensionally “similar” to other information in that cluster. The location of inputs to the model within each cluster indicates to what degree it is similar to other cluster information. This closely approximates what we know of how new information is stored in long-term memory (Baddeley, 2000) in the human brain (Figure 3).

Figure 3: Clustering in the Art Neural Network

Attempts to mimic the memory structure of the human brain are not new. One of the best known of these is John Anderson’s ACT-R model (Anderson, 1996). Although this model provides a good approximation of our current understanding of how the human memory process works, it is essentially a scientific research tool and therefore is not ‘friendly’ enough to be of use to non-researchers. One of the reasons that we have spent 26 years studying conceptual graphs was to improve their ease of use and practical application to everyday problems at low-cost.

Our particular approach uses conceptual graphs to create large classes of finite knowledge states that can easily be validated because they bound the knowledge into separable states. In fact the number of states that can be created goes as 2N where N is the number of nodes in the graph being used. For this reason we often refer to this as an “infinite-state machine”. The addition of adaptive resonance theory, a purely physics-based approach, to automatically identify where in knowledge space a new knowledge state should reside based upon prior knowledge is clearly both a novel and unique approach to a third wave Artificial Intelligence architecture.

In summary the CORE ML tools allow one to build AI and ML algorithms that provide the the following:

  • Automated Knowledge Acquisition from Subject Matter Experts
  • Based upon an ISO Open Standard for Knowledge Interchange
  • Does not Require Programming or AI Expertise
  • Proven on many DoD programs over a 25-year period
  • Provides the Capability to validate Results
  • Government-Licensed-No Proprietary code
  • Automatic generation of code for insertion into applications
  • Allows rapid generation of algorithms using an easy-to-use graphical interface
  • Runs on a PC platform using windows-no Special hardware or software required
  • Award-winning Tools Recognized by and used by all services
  • TRL LEVEL 9—CORE applications currently deployed in DoD Software
  • BRC has a team of programmers, physicists, AI/ML experts and Cognitive Psychologists to support the toolset use if needed
  • Provides the ability to compress AI/ML knowledge by 90% or more in a secure-lossless format

For more information about the CORE toolset or to schedule a demonstration contact myself or David Vidler, BRC Business development at (256) 882-6229.

Andy Bevilacqua, Ph.D.
Cognitive Optical Psychophysicist

2 thoughts on “Artificial Intelligence and Machine Learning: Part 5: Complexity

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s