Technology Today

2011 Issue 2

Cognitive Computing Advances Help Computers Work More Like the Human Brain, Improve Information Gathering

Cognitive Computing Advances Help Computers Work More Like the Human Brain, Improve Information Gathering

Raytheon's recent advances in cognitive computing (software/hardware that mimics human mental processes of perception, memory, intelligence and consciousness) have shown that computers have the potential to do as well as humans at assimilating and analyzing information. This includes dealing with information across multiple information sources (e.g., newspapers, magazines, Web, chat, television, radio, telephone, recordings) and information types (e.g., hard copy or electronic). These advances have resulted in the development of an artificial cognitive neural framework that uses metacognitive (knowledge about knowing) and metamemory (knowledge about memory) processing concepts similar to those that the human brain uses to process, categorize and link information. Current applications include Integrated System Health Management and Adaptive Learning Systems for command, control, communications, computers, intelligence, surveillance and reconnaissance (C4ISR).

Figure 1
Artificial Cognition

To understand the world we live in, we gather information via our senses and process that information in ways that enable us to reason about what we perceive. The accuracy of our perception is determined by the accuracy of our information (e.g., credibility, applicability, comprehensiveness, representativeness and context) and how we apply reasoning to that data through the lens of our beliefs and assumptions.

This is also true when information analysts mine diverse information sources for clues and insights. First, the information is collected from diverse sources, each with its own context and error source. Because of the diverse error sources and contexts, ambiguity is introduced into the correlation and inference processes applied to the combined information. This ambiguity can make it difficult for information analysts to use such data to find related events and infer likely outcomes. Moreover, an analyst's personal variables (e.g., health, focus, experience, learning, emotions, beliefs and assumptions) may also affect the analysis process.

To assist the information analyst, Raytheon has developed a processing architecture that employs intelligent information software agents (ISAs) that operate within an artificial cognitive neural framework (ACNF), as shown in Figure 1. ISAs are active, persistent software components that perceive, reason, act and communicate with each other. The ACNF is a hybrid computing architecture that uses genetic learning algorithms, neural networks and fuzzy classification algorithms that allow diverse information sources and events to be associated and correlated, enabling the ACNF to make observations, process information, make inferences and recommend decisions.

The ACNF uses continuously recombinant neural fiber networks that map complex memory and learning patterns as the ACNF evolves or adapts to its environment. Continuously recombinant neural fiber networks are neural networks whose internal topology adapts as the system learns and evolves. The entire system functions and communicates via the ISAs that mimic human capabilities (analysis, reasoning, learning, and reporting; i.e., cognitive intelligence). These capabilities can process diverse information types and sources and translate them into actionable information within the ACNF framework. Specifically, this cognitive intelligence can answer questions and explain situations.

One category of ISAs within the ACNF takes the form of cognitive perceptrons, a type of ISA used to carry cognitive information throughout the ACNF, which can learn from experiences and can be used to predict future states (prognostics). Cognitive perceptron ISAs analyze sensor data to detect complex states and diagnose problems. They can interface with other autonomic software agents within the system to form software agent coalitions. They can also reason using domain-specific application objects and can have autonomous behaviors and goals. The ACNF framework provides memory categories similar to those of human memory systems, including sensory memories, short-term (working) memories and long-term memories. These memory categories retain not only knowledge, but also the contextual, temporal and spatial information relevant to the sensory data and information processed and retained by the ACNF processes.

Other types of ISAs furnish the artificial cognition and the artificial prefrontal cortex1 (mediator) processes that allow the ACNF to function autonomously (without supervision). These ISAs are autonomous software agents that create, in essence, an information agent ecosystem, comprehending its external and internal environment and acting on it over time, in pursuit of its own agenda and goals, so as to affect what it comprehends in the future. The three main subsystems within the ACNF are:

  • The mediator (the artificial prefrontal cortex), shown in green: This takes information processed through the artificial cognition processes from the cognitive perceptrons, forming coalitions of perceptrons that are used to update the short term, long-term and episodic memories.
  • The memory system, shown in tan: The memory system consists of sensory, short-term and long-term memories. The memory system continually broadcasts the information that is available within the ACNF memories (what the system has learned and "knows") to the conscious perceptrons that form the cognitive center of the system.
  • The cognitive system, shown in blue: This provides the cognitive, learning, emotion and consciousness structures, which are responsible for the cognitive functionality of perception, consciousness, emotions, information processing and other cognitive functions within the ACNF.
Figure 2
Cognitive Perceptrons

Four types of cognitive perceptrons operate within the ACNF (Figure 2):

  • The data steward perceptron generates and maintains the metadata required to find and extract information from heterogeneous information sources.
  • The advisor perceptron generates and maintains the topical/subject information required to find information within the memory system relative to the current problem (or topic) being processed.
  • The reasoner perceptron analyzes questions and relevant source information to provide answers to analysts and to develop cognitive rules within the ACNF for future cognitive processes (adapt and evolve the system as it learns).
  • The analyst perceptron uses inferences (patterns of thinking) to direct questionand-answer generation and to create situational analysis, based on the current and learned information about a current topic, with integrated explanations based on all available information.
Conclusions

Our research has shown that the cognitive perceptrons, in combination with the ACNF, provide an architecture, framework, and processes that facilitate cognition, learning, memories and information processing in an autonomous fashion, similar to human reasoning, and can greatly assist information analysts. Future work includes providing knowledge density (a measure of how much the system knows/has learned about a particular topic/subject) and analytical competency (a measure of the system's ability to reason or analyze information pertaining to a particular topic/subject).These processes will enable the system to assess its own capabilities and knowledge gaps.

1In humans, the prefrontal cortex carries out executive functions that provide the ability to differentiate among conflicting thoughts, determine correct actions and determine future consequences of current activities. It also provides our metacognitive and metamemory capabilities.

Dr. James A. Crowder

Top of Page