Abstract - The purpose of this paper is to describe a new neural framework for Artificially Intelligent systems, the Artificial Cognitive Neural Framework (ACNF), ...
The Artificial Prefrontal Cortex: Artificial Consciousness Dr. James A. Crowder, Shelli Friess, MA, NCC Raytheon Intelligence and Information Systems 16800 E. Centretech Parkway, Aurora, Colorado 80011
Abstract - The purpose of this paper is to describe a new neural framework for Artificially Intelligent systems, the Artificial Cognitive Neural Framework (ACNF), which allows for “conscious” software agents within the AI processing environment. “Conscious” software agents are autonomous agents that range in functionality and are situated in the processing environment. They sense the environment and act on it over time, in pursuit of their own agenda, based on their evolving constraints. As they evolve it is possible for them to change what they sense at a later time. These “conscious” agents are also “cognitive” agents, in that they are equipped with constructs for concept formation, consciousness, basic emotions, and short & long-term memories [14]. The long-term memories provide identification, recognition and categorization functions, as well as identification of feelings [21]. The short-term memories provide preconscious buffers as a workspace for internal activities. A transient episodic memory is also provided as a contentaddressable associative memory with a moderately fast decay rate. This provides the architectural framework for an Artificial Prefrontal Cortex (APC), which provides cognitive intelligence for AI processing systems and allows for rapid analysis, reasoning, and reporting capabilities. The APC facilitates information, intelligence, and memory integration and allows faster accommodation and delivery of knowledge and knowledge characteristics across the system [16].
1 Introduction The prefrontal cortex has long been suspected to play an important role in cognitive control, in the ability to orchestrate thought and action in accordance with internal goals. Cognitive control stems from the active maintenance of patterns of activity in the prefrontal cortex that represent goals and the means to achieve them [19]. They provide bias signals to other cognitive structures whose net effect is to guide the flow of activity along neural pathways that
establish the proper mappings between inputs, internal states, and outputs needed to perform a given task. The Prefrontal Cortex is integral to planning complex cognitive behaviors, personality expression, decision making and moderating correct social behavior [21]. The basic activity of this brain region is considered to be orchestration of thoughts and actions in accordance with internal goals [10]. Here we present structure within the architecture to provide an APC and provides the structure and context for artificial feelings and emotions within the overall AI system [17, 18]. Here we discuss the roles they can play in an intelligent software agent performing a practical, real world task. These agents would be actively involved in every instance of action selection, and at least potentially involved in each learning event [11]. The pervasive, central role that feelings and emotions would play in the control structure of these conscious software agents mimics the roles they play in human cognition, and, over time, may give rise to clarifying hypotheses about human decisionmaking and several forms of human learning [12, 13]. The functions carried out by the prefrontal cortex area can be described as executive functions. Executive functions relate to abilities to differentiate among conflicting thoughts, determine good and bad behavior, better and best, same and different, future consequences of current activities, working toward a defined goal, prediction of outcomes, expectation based on actions, and social "control" [8]. The prefrontal cortex is of significant importance when top-down processing is needed. Topdown processing by definition is when behavior is guided by internal states or intentions. All of these are driving toward the cognitive concept of “mindfulness:” •
Mindfulness: an awareness that lets us see things as they truly are without distortion or judgment, giving the most insightful explanation of how mindfulness can change not only our lives, but the very structure of our brains.
In order for AI systems to be truly autonomous, we must give them these “executive functions” abilities. One of the cognitive concepts that will be necessary for a truly autonomous system is the ability to top-down processing. To take an understanding of the mission or task at hand, from this define goals and prediction of outcomes, and utilize this knowledge to define the system behaviors needed to meet the mission or task goals. For an autonomous AI system, executive management and strategic knowledge would be essential. The executive management autonomous system processes would involve planning, monitoring, evaluating and revising the system’s own cognitive processes and products. Strategic knowledge would involve knowing what tasks or operations to perform (factual or declarative knowledge), knowing when and why to perform the tasks or operations (conditional or contextual knowledge) and knowing how to perform them (procedural or methodological knowledge). Both executive management and strategic knowledge capabilities are required for the system to autonomously self-regulate its own thinking and learning [2].
System [2]. The Mediator gathers information and facilitates communication between agents. Hence, each cognitive decision is handled by the Mediator (the Artificial Prefrontal Cortex) which takes information from perceptrons and from coalitions of perceptrons and updates the short-term, long-term and episodic memories or pedigree [9]. The information available in memory (what the system has learned) is continually broadcast to the conscious perceptrons that form the cognitive center of the system (i.e., they are responsible for the cognitive functionality of perception, consciousness, emotions, processing, etc.) [10].
Here we propose a model for an Artificial Prefrontal Cortex as part of an overall Artificial Cognitive Neural Framework (ACNF) required for real AI autonomous system control. We will begin with a discussion of the ACNF and then introduce a Hidden Markov Model of an Artificial Prefrontal Cortex, utilizing fuzzy possibilistic logic to drive the system between cognitive states. Figure 1– The Artificial Cognitive Neural Framework
2 The Artificial Cognitive Neural Framework As knowledge and cognitive context increases within the AI system, a formal framework for dealing with increasing and decreasing levels of cognitive granularity is necessary to learn, understand, and store the closeness of cognitive relationships [7]. We deal with these abilities utilizing a hybrid, fuzzy-neural processing system with genetic learning algorithms [1]. This processing system uses a modular artificial neural architecture. This architecture is based on a mixture of neural structures with Intelligent Software Agents that add flexibility and diversity to the overall system capabilities. In order to provide an artificially intelligent processing environment, we believe the system should possess the notion of artificial emotions that allow the processing environment to “react” in real-time as the systems environment changes and evolves. This hybrid fuzzy-neural processing framework is called the Artificial Cognitive Neural Framework (ACNF) [5]. Figure 1 illustrates the ACNF which has analogies to an AI blackboard system, except that it is greatly extended to allow for system-wide action selection. The three main subsystems within the architecture are the Mediator, the Memory System, and the Cognitive
The use of an ACNF for analysis, reasoning, and reporting provides the “Cognitive Intelligence” to allow the top-down executive processing required for real-time autonomous operations with AI systems [7]. The ACNF utilizes Fuzzy-Self Organizing Contextual Topical Maps that allow information fragments to be put together to form memories based on topical information. The Mediator, or Artificial Prefrontal Cortex, has the following properties: • • • • • •
Facilitates Information, Intelligence, and Memory Integration. Allows faster accommodation and delivery of cognitive knowledge and knowledge characteristics. Increases cognitive flexibility, allowing rapid adaptation to changing environments. Allows scalability of the overall AI system. Reliably provides cognitive information across domains, even after a software, network, or hardware failure. Provides a topical, information, cognitive knowledge hosting and management infrastructure that is highly distributed, yet manageable.
The Artificial Prefrontal Cortex is implemented or instantiated with Intelligent Information Software Agents. The logical flow is illustrated in Figure 2.
3 The Artificial Prefrontal Cortex The Artificial Prefrontal Cortex provides governance capabilities that enable definition and enforcement of cognitive policies governing the content and usage of the cognitive and topical maps by the Intelligent Software Agent framework across the AI enterprise. Has Coordination Is Represented By
Fulfills
Performs Has Permission
Prefrontal Cortex
SW Agent
Role
Activity
Resources Permits
Represents
Is Fulfilled By
Is Performed By Coordinates
Domain of Interest: Cognitive System Processes
Figure 2– Artificial Prefrontal Cortex Inference Flow In order to understand the cognitive interactions that must occur within an Artificial Prefrontal Cortex, a model was built to drive the Intelligent Software Agent framework that provides linkage between the major cognitive states within the cortex [4, 5]. Figure 3 illustrates this model. The cognitive processes represented are based on AI interpretations of Dr. Peter Levine’s Autonomic Nervous System States [20].
A
B
A
B
A
EVID
Represents the possibility of B, given A (Po) Represents the possibility of B, given the Possibility of A with confidence bound (Po2) Represents “what’s the logical causality of the evidence/observation given A happened.”
Figure 3– Artificial Prefrontal Cortex Affected State Model Detecting cognitive process information within the ACNF begins with sensors that capture information about the system’s physical or cognitive state or behavior. The information is gathered and interpreted by the cognitive perceptrons similar to how humans utilize cues to perceive cognitive states or emotions in others. The Artificial Prefrontal Cortex (APC) provides the possibilistic inferences for the system to transfer between cognitive states. The APC shown in Figure 3 illustrates only three
cognitive states for display purposes only, but it is straight forward to include more states. The idea is that the APC makes it possible to transition between cognitive states at any instant, and transition between these states with certain possibilistics. These possibilistic parameters evolve over time, driven by the learning algorithms and how they affect both normal and emotional memories. These cognitive state transition conditional possibilistics provide the APC with the ability to make executive-level plans and move between cognitive states, each of which has its own set of priorities, goals, and motivations. The driving requirement for the APC is to create a truly autonomous AI system that can be used in a variety of applications like UAVs, intelligence processing systems, cyber monitoring and security systems, etc. In order to accomplish these tasks, the APC must have the following capabilities processes and acted on by the human prefrontal cortex: Cue Familiarity: cue familiarity is the ability of the system to evaluate its ability to answer a question before trying to answer it [21]. In cue familiarity, the question (cue) and not the actual memory (target) becomes crucial for making cognitive judgments. This implies that judgments regarding cognitive processing and decisions would be based on an the system’s level of familiarity with the information provided in the cue. This executive-level, top-down cognitive judgment requires APC abilities to allow the AI system to judge whether they know the answer to a question, i.e., is the system familiar with the topic or mission, allowing the system to to judge that they do not know the answer to a question which presents new or unfamiliar terms or conditions. Cognitive Accessibility: cognitive accessibility suggests that the system’s memory will be more accurate when the ease of cognitive processing (accessibility) is correlated with memory behavior (emotional memory). This implies that the quality of information retrieval depends on the system’s density of knowledge on the topic or subject (or individual elements of informational content about a topic), since the individual elements of topical information differ in strength. The speed of access is tied to both density of knowledge and emotional memory responses to the information. Cognitive Competition: cognitive competition can be described as three principles: •
The AI cognitive processing system (the brain) is activated by a variety of inputs (sensors). There is textual, audio, and visual (picture and video) information that compete for cognitive processing access.
•
Competition occurs within the multiple cognitive processing subsystems and is integrated by the Intelligent Software Agents between the various cognitive processing subsystems.
•
Competition can be assessed utilizing top-down neural priming within the APC, based on the relevant characteristics of the object at hand.
Cognitive Interaction: This combines cue familiarity and cognitive accessibility. In cognitive interaction, once cue familiarity fails to provide enough information to make cognitive inferences, cognitive accessibility in employed access extended memories and may employ stored emotional memory cues to access the required information to make the required cognitive inferences. This may result in slower response time that with cue familiarity alone.
memory. The resulting eigenspaces determine topics that are compared within the contextual FSOM to look for “closeness” of topics to be used in cognitive processing algorithms to determine the cognitive state that will be used to make inferences about the question or task being posed. The eigenspaces are estimated under a variety of emotional memory conditions and their dependencies on external inputs and cognitive factors determined. Eigen Trajectories are then characterized, capturing the dynamic aspects of relationships between topical closeness and the information and memories available.
4 Artificial Prefrontal Cortex Processing In order to provide the APC with capabilities described above, processing constructs must be in place to allow cognitive inferences to be made, based on the information received, inferences and decisions learned, and an overall sense of priorities, goals, and needs. The following describes processing constructs that allow a viable APC to be constructed. 4.1 The Fuzzy, Self-Organizing, Contextual Topical Map The Fuzzy, Self-Organizing, Contextual Topical Map (FSOCTM) is a general cognitive method for analyzing, visualizing, and providing inferences for complex, mulitdimensional sensory information (textual, auditory, and visual). The FSOCTM is actually built on two, separate, Fuzzy, Self-Organizing topical Maps (FSOM). The first is a semantic FSOM that organizes the information semantically into categories, or topics, based on the derived eigenspaces of features within the information. Figure 4 illustrates an FSOM with information and topical “closeness” search hits designated. The larger hexagons denote topical sources that best fit the search criterion. The isograms denote how close the hits are to a particular cognitive information topic. The FSOM information and topical closeness map has several important attributes: •
Image processing algorithms can be utilized to analyze the output of the FSOM
•
Searches use contextual information to find cognitive links to relevant memories and information available.
•
The FSOM is self-maintained and automatically locates input from relevant Intelligent Software Agents and operates unsupervised.
The high-level topical spaces are compared, within the APC to identifiable “eigenmoods” within the emotional
Figure 4– The Fuzzy, Self-Organizing Topical Map The high-level topical spaces are compared, within the APC to identifiable “eigenmoods” within the emotional memory. The resulting eigenspaces determine topics that are compared within the contextual FSOM to look for “closeness” of topics to be used in cognitive processing algorithms to determine the cognitive state that will be used to make inferences about the question or task being posed. The eigenspaces are estimated under a variety of emotional memory conditions and their dependencies on external inputs and cognitive factors determined. Eigen Trajectories are then characterized, capturing the dynamic aspects of relationships between topical closeness and the information and memories available. Once the FSOM is created, the resultant topical eigenspaces are mapped to the larger FSOCTM to show cognitive influences and ties to larger cognitive processes and other memory information, as depicted in Figure 5. The value of superimposing the FSOCTM onto the SOM is that it defines it defines the cognitive information domain’s ontology, and enables the use of a Topic Map Query Language (TMQL) within the APC. The Topic Map enables end APC to rapidly search information conceptually. It also enables sophisticated dialectic searches to be performed for them.
the reasoning that renders the cognitive intelligence lead plausible and enables the possibility to be measured and cognitive inferences made within the APC.
Associations, by type Is associated with Is influenced by
The approach to cognitive intelligence inferencing within the APC is threefold:
Figure 5– Superimposing the FSOCTM onto the FSOM
•
First the FSOCTM is investigated to semantically organize the diverse information collected and retrieved from memory.
•
The map produced by the FSOM is utilized to enhance the APCs comprehension about the situations under analysis.
•
Third, as the APC traverses the map to find related and relevant events, the results are used to create cognitive clues that are stored in the emotional memory for use under similar
5 The Dialectic Search (DS) The Dialectic Search uses the Toulmin Argument Structure to find and relate information and memories that develops a larger argument, cognitive inference. The Dialectic Search Argument (DSA), illustrated in Figure 6, has four components: •
Information and Memories: both in support of and rebutting the argument or hypothesis under analysis by the APC.
•
Warrant and Backing: explaining and validating the hypotheis.
•
Claim: defining the hypothesis itself
•
Fuzzy Inference: relating information/memories to the hypothesis.
the circumstances.
The Dialectic Search serves two purposes: •
First, it provides an effective basis for mimicking human reason.
Figure 6– The Dialectic Search Structure
•
Second, it provides a means to glean relevant information from the Topic Map and transform it into actionable cognitive intelligence.
This approach mimics human intelligence, learning from Intelligent Software Agents using knowledge ontology to define particular knowledge domains (topics), having experts (intelligent information software agents) to cartographically label the FSOM to capture the meaning of the integrated information thus capturing the knowledge of each cognitive inference. The APC processing environment has three processing levels, illustrated in Figure 7 [6].
These two purposes work together to provide an intelligent system that captures the capability of the human reasoning to sort through diverse information and find clues (based on Cue Familiarity discussed above). This approach is considered dialectic in that it does not depend on deductive or inductive logic, though these may be included as part of the warrant. Instead, the Dialectic Search depends on non-analytic inferences to find new possibilities based upon warrant examples. The Dialectic is dialectic because its reasoning is based upon what is plausible; the Dialectic Search is a hypothesis fabricated from bits of information fragments (memories) put together utilizing the topical maps and eigenspaces [3]. Once the available information has been assimilated by the Dialectic Search, information that fits the support and rebuttal requirements is used to instantiate a new claim or hypothesis. This claim is then used to invoke one or more new Dialectic Searches. The developing lattice forms
•
The first will identify patterns of behavior that have been seen (or behavior similar in a “fuzzy” relational way) before.
•
The second is an expanded pattern recognition that involves pattern discovery algorithms that augment patterns that are similar to known patterns but need additional information to describe the pattern divergences.
•
The third is a full up pattern discovery paradigm to make sense of information that has not been previously described (how does the system find things it didn’t know it was looking for.
Data/Information: Easy Known Quasi-Familiar Unknown Weird
3.
(E) (K) (F) (U) (W)
E, K, F, U, W
4. Resource Constraints
Level 1 All Known Solutions (exact, with error bounds)
L, I, J, K: Lumped System Resources F, U, W
∑ L(I ) = 1, I = 1,2,3 ∑ I (i) = L(1), i = 1,K, N ∑ J ( j ) = L(2), j = 1,K, N ∑ K (k ) = L(3), k = 1,K, N
All Known Solutions with Variations
Pattern Recognition
New Solutions & Variations of Known Solutions
5.
1
Level 2
2 3
Learning/Evolutionary Paradigms
New Solutions
L(1)min ≤ L(1) ≤ L(1)max
Expanded Pattern Recognition & Pattern Variation
L(2)min ≤ L(2) ≤ L(2)max L(3)min ≤ L(3) ≤ L(3)max
U, W
6.
All Known Solutions with difficult variations
Pattern and Hypothesis Discovery – Computationally Level 3 Pattern Discovery Expensive
Major Hypothesis Tesing
7. Figure 7– The APC Intelligent Software Agent Processing Levels
8.
6 Conclusions and Discussion The need to mimic human intelligence demands a polymorphic architecture that is capable of both hard and soft computing. The APC with the FSOCTM, soft computing, and utilizing the ACNF framework provides the structure that allows the APC to evolve and grows as it learns more about its environment. There is also a need to process streams of diverse information to provide terse vectors for FSOM and cognitive mapping [15]. This is accomplished through the use of the genetically evolving ACNF processing network.
9.
10.
11. 12.
Reason for using a FSOM approach is to ensure the results can be readily understood by the APC. The FSOM performs a critical role by collapsing multiple dimensions in information onto 2-dimensional space – a form that be more easily computed and understood by the APC, especially when it has been enhanced to include emotional memory information. As more information is acquired, it is mapped into an already understood structure within the ACNFs structure [5]. We believe the work outlined in this paper creates the foundation for further research into the creation of an Artificial Prefrontal Cortex, which we feel is essential to an actual learning, thinking, reasoning, and autonomous AI system.
19.
7 References
20.
1.
21.
2.
Crowder, J. A., “Adaptive Learning Algorithms for Functional Constraints on an Evolving Neural Network System.” NSA Technical Paper CON_0013_2002_003 (2002). Crowder, J. A., “Machine Learning: Intuition (Concept) Learning in Hybrid Genetic/Fuzzy/Neural Systems.” NSA Technical Paper CON_0013_2003_009 (2003).
13. 14. 15. 16. 17. 18.
Crowder, J. A., “Operative Information Software Agents (OISA) for Intelligence Processing.” AIAA Aerospace@Infotech 2010 Conference. Crowder, J.A., Friess, S., “Artificial Neural Diagnostics and Prognostics: Self-Soothing in Cognitive Systems.” International Conference on Artificial Intelligence, ICAI’10 (July 2010). Crowder, J. A., Friess, S., “Artificial Neural Emotions and Emotional Memory.” International Conference on Artificial Intelligence, ICAI’10 (July 2010). Crowder, J. A., “Flexible Object Architectures for Hybrid Neural Processing Systems.” International Conference on Artificial Intelligence, ICAI’10 (July 2010). Crowder, J. A., Carbone, J, “The Great Migration: Information to Knowledge using Cognition-Based Frameworks.” Springer Science, New York (2011) DeYoung, C. G., Hirsh, J. B., Shane, M. S., Papademetris, X., Rajeevan, N., and Gray, J. R. (2010). Testing predictions from personality neuroscience. Psychological Science, 21(6):820-828. Marsella, S., and Gratch J., “A Step Towards Irrationality: Using Emotion to Change Belief.” 1st International Joint Conference on Autonomous Agents and Multi-Agent Systems, Bologna, Italy (July 2002). Miller EK, Freedman DJ, Wallis JD (August 2002). "The prefrontal cortex: categories, concepts and cognition". Philos. Trans. R. Soc. Lond., B, Biol. Sci. 357 (1424): 1123–36. Newell, A., “Unified Theories of Cognition.” Cambridge MA: Harvard University Press (2003). Damasio A (1994) Descarte's error: Emotion, reason, and the human brain. New York: Gosset/Putnam. Davis M, Whalen PJ (2001) The amygdala: vigilance and emotion. Mol Psychiatry 6:13-34. Eichenbaum H (2002) The cognitive neuroscience of memory. New York: Oxford University Press. Kosko, G., “Fuzzy Cognitive Maps,” International Journal of Man-Machine Studies, 24: 65-75. LaBar KS and Cabeza (2006) Cognitive neuroscience of emotional memory. Nat Rev Neurosci 7: 54-64. LeDoux JE (1996) The Emotional Brain. New York: Simon and Schuster. LeDoux JE (2000) Emotion circuits in the brain. Annu Rev Neurosci 23:155-184. LeDoux JE (2002) Synaptic Self: How our brains become who we are. New York: Viking. Levine, P., “Walking the Tiger: Healing Trauma.” North Atlantic Books, Berkeley, CA (1997). Yang Y, Raine A (November 2009). "Prefrontal structural and functional brain imaging findings in antisocial, violent, and psychopathic individuals: a meta-analysis". Psychiatry Res 174 (2): 81–8. doi:10.1016/j.pscychresns.2009.03.012. PMID 19833485.
