Artificial Memory Reconstruction James A. Crowder Raytheon Intelligence and Information Systems 16800 E. Centretech Parkway, Aurora, Colorado 80011
[email protected]
Abstract. Any realistic artificial cognitive processing system must be able to store and recall (reconstruct) complex sequential patterns. Such a system must contain both episodic and semantic memories and demonstrate these memory properties within a cognitive domain of a large number of spatio-temporal memory patterns (episodes), given only a simple example or representation of such a pattern (a partial memory). The cognitive system must be able to construct memories from information fragments without significant interference, as well as exhibit similaritybased category generalization described as semantic memory properties in humans. Presented here is the theory and architecture for an artificial cognitive system that provides information fragment storage and memory reconstruction utilizing methodologies similar to human memory reconstruction processes. We will describe the overall architecture, called Intelligent information Software Agents (ISAs) to facilitate Artificial Consciousness (ISAAC) which includes the cognitive and memory cycle required for memory storage and retrieval (reconstruction). Keywords: Artificial Intelligence, Episodic Memory, Memory Reconstruction.
1
Introduction
Current research into information theory for artificial cognitive systems has led to the creation of architectures and algorithms that allow information to be disassembled into its separable information fragments and stored with knowledge relativity threads that provide metadata as to the context of the information fragment to subjects or topics the artificially intelligent system understands [Crowder 2010, Crowder and Carbone 2011, Crowder and Friess 2011&2012a]. Memories involve the acquisition, categorization, classification and storage of information. The purpose of memory is to provide the ability to recall (reconstruct) information and knowledge as well as events, based on information fragments that have been sense, experienced, learned, and stored. In order to provide an autonomous, artificially intelligent system with sensing, analyzing, processing, and learning capabilities similar to humans, we must also redefine the process of information storage and retrieval (reconstruction) to provide architectures and methodologies similar to humans. Here we describe a new cognitive architecture for artificial intelligence controlled devices that incorporates artificial neural memory
types, similar to humans (including procedural memory [Crowder, Taylor, and Raskin 2012]), that provides memory creation and recall (reconstruction) similar to human brain processes. This cognitive architecture will provide a scalable framework that provides episodic memory creation as the entity experiences, and, over time, develops procedural memory “scripts” that allow the entity to repeat tasks it has “learned” how to accomplish. The system will utilize a temporal-calculus driven spatial map to store spatio-temporal information, as well as the procedural memories. This provides the system with the cognitive abilities to approach task selection and reason through both experiential and spatial learning. This is accomplished though the use of Intelligent information Software Agents (ISAs) that provide the cognitive capabilities of perception, internal state reasoning, behavior selection, learning, memory, and motor control, in order to control the overall behavior of the entity, taking into account its own internal state.
2
The Artificial Cognitive Neural Framework
Here we describe the ISAAC Artificial Cognitive Neural Framework (ACNF) that provides ISAAC with the cognitive and processing capabilities to organize information semantically into meaningful fuzzy concepts and information fragments that create cognitive hypotheses as part of its topology [Zadeh 2004]. This approach addresses the problems of autonomous information processing by accepting that the system must purposefully communicate concepts fuzzily within its processing system, often inconsistently, in order to adapt to a changing real-world, real-time environment. Additionally, we describe a processing framework that allows the system to deal with real-time information environments, including heterogeneous types of fuzzy, noisy, and obfuscated data from a variety of sources with the objective of improving actionable decisions using Recombinant Knowledge Assimilation (RNA) processing [Crowder and Carbone 2001b, 2011c] integrated within an ACNF to recombine and assimilate knowledge based upon human cognitive processes. The cognitive processes are formulated and embedded in a neural network of genetic algorithms and stochastic decision making with the goal of recombinantly minimizing ambiguity and maximizing clarity while simultaneously achieving a desired result [Crowder 2010b]..
a. The ISAAC ACNF Architecture The ACNF (see Figure 1) is a hybrid computing architecture that utilizes genetic, neural-network, fuzzy, and complex system components, that allow integration of diverse information sources, associated events, and multiple learning and memory systems to make observations, process information, make inferences, and decisions. Within the ACNF, Continuously Recombinant Neural Fiber Networks are utilized to map complex memory and learning patterns as the system learns and adapts [Bishop 1995]. The entire system “lives” and communicates via Intelligent information Software Agents (ISAs) that mimic human reasoning by understanding how to create and develop hypotheses [Crowder 2003b, Crowder 2010a, 2011a].
Figure 1 – The Artificial Cognitive Neural Framework Architecture
This architecture provides a collection of constraints, building blocks, design elements, and rules for composing the cognitive aspects. Figure 1 illustrates the ACNF architecture. The three main subsystems within the ACNF are: 1.
2.
3.
The Cognitive System: this consists of the Artificial Cognition, Learning Algorithms, Artificial Neural Emotions, Artificial Consciousness, and Cognitive Perceptrons that make up the consciousness structures. These are responsible for the cognitive functionality of perception, consciousness, emotions, informational processing, and other cognitive functions within the ACNF. The Mediator (the Artificial Prefrontal Cortex): the Mediator takes information from the ISAs, processed through the Artificial Cognition processes, and forms coalitions of perceptrons that are used to update the short-term and long-term, and episodic memories. The Memory System: the Memory System consists of the Memories and Memory Integration function and takes information that is available within the ACNF memories (what the system has learned and ‘knows’) and continually broadcast it to the conscious perceptrons that form the cognitive center of the system; it also integrates these into current short term memory to provide Integrated Knowledge (“World Data”) to the Cognitive Perceptrons to analyze incoming sensory information
3
Constructivist Memory Theory
In order to design, develop, and implement ISAAC to be truly autonomous, it must be provided with dynamic memory abilities [Crowder 2010a]. Memories are typically classified into three different types: Sensory, Short-Term, and Long-Term. Each memory type has several instantiations, dealing with different types of information. Here we will explore each type of memory system and its implications to ISAAC. We begin our discussion of memory processing with a look at the relationships between the three main types of memories. Figure 1 illustrates an AIS Memory Upper Ontology, similar to human memory systems, describing these relationships (Eichenbaum 2002). ISAAC’s cognitive processes are based on Constructivist Learning, in which the ISAAC cognitive learning processes are a building (or construction) process in which the ISAAC’s cognitive system builds internal illustrations of its learned knowledgebase, based on its experiences and personal interpretation (fuzzy inferences and conceptual ontology [Raskin & Taylor 2010a and Taylor & Raskin 2011a]) of its experiences. ISAAC’s Knowledge Representation and Knowledge Relativity Threads [Crowder and Carbone, 2011c], within ISAAC’s cognitive system memories are continually open to modification, and the structures and linkages formed within ISAAC’s short-term, long-term, and emotional memories [Crowder and Friess, 2010b], along with its Knowledge Relativity Threads (KRT) [Crowder and Carbone 2011c], form the bases for which knowledge structures would be created and attached to ISAAC’s memories, which are stored as Binary Information Fragments [Crowder and Friess 2012b]. This drives us to a memory processing, encoding, and retrieval (reconstruction) based on Knowledge Binary Information Fragments [Crowder and Carbone 2012]. This Artificial Memory processing procedure happens within ISAAC’s Sensory and Short-Term Memory systems and contains the following steps (see Figure 2): •
•
Information Fragment Selection: this involves filtering the incoming information from ISAAC’s Artificial Preconscious Buffers into separable information fragments and then determining which information fragments are relevant to be further processed, stored, and acted on by the cognitive processes of ISAAC as a whole. Once information fragments are created from the incoming sensory information, they are analyzed and encoded with initial topical information, as well as Metadata attributes that allow the cognitive processes to organize and integrate the incoming information fragments into ISAAC’s overall Long-Term Memory system. The Information Fragment encoding creates a small, Information Fragment Cognitive Map that will be used for the organization and integration functions. Information Fragment Organization: these processes within the Artificial Cognition framework create additional attributes within the Information Fragment Cognitive Map that allow it to be organized for integration into the overall ISAAC Long-Term Memory framework. These attributes have to do with how the information will be represented in Long-Term Memory and
•
determine how these memory fragments will be used to construct new memories, or recall, memories later by as needed by ISAAC, using Knowledge Relativity Thread representation to capture the context of the Information Fragment and each of its qualitative relationships to other fragments and/or bundles of fragments already created. Information Fragment Integration: Once the Information Fragments within the Short-Term Memory have been KRT encoded, they are compared, associated, and attached to larger, Topical Cognitive Maps that represent relevant subject or topics within ISAAC’s LTM system. Once these Information Fragment Cognitive Maps have been integrated, processed, and reasoned about, including emotional triggers or emotional memory information, they are sent on to both the Long-Term Memory (LTM) system, as well as ISAAC’s Artificial Prefrontal Cortex to determine if actions are required.
Figure 2 – ISAAC’s Binary Information Fragment Encoding
LTM information fragments are not stored in databases or as files, but encoded and stored as a triple helix of continuously recombinant binary neural fiber threads that represent: •
The Binary Information Fragment (BIF) object along with the BIF Binary Attribute Objects (BAOs).
• •
The BIF Recombinant Knowledge Assimilation (RNA) Binary Relativity Objects. The Binary Security Encryption Threads.
Built into the RNA Binary Relativity Objects are Binary Memory Reconstruction Objects, based on the type and source of BIF, that allow memories to be constructed for recall purposes. There are several types of Binary Memory Reconstruction Objects, they are: • • • • •
4
Spectral Eigenvectors that allow memory reconstruction using Implicit and Biographical LTM BIFs Polynomial Eigenvectors that allow memory reconstruction using Episodic LTM BIFs Socio-Synthetic Autonomic Nervous System Arousal State Vectors that allow memory reconstruction using Emotional LTM BIFs Temporal Confluence and Spatial Resonance coefficients that allow memory reconstruction using Spatio-Temporal Episodic LTM BIFs Knowledge Relativity and Contextual Gravitation coefficients that allow memory reconstruction using Semantic LTM BIFs
Artificial Memory Reconstruction
Given that ISAAC’s memories are not stored as database files, but as Binary Information Fragments with Knowledge Relativity Threads that provide contextual and meta-information, memory recall, similar to humans, is a process of reconstructing the memory, based on topical maps that map the topic or subject to be “remembered” into the Conceptual Ontology and Information Base to find those information fragments that are relative to the topics(s) to be pulled, associated, integrated, and presented as a memory to ISAAC’s conscious processes. The process outlined below illustrates the memory reconstruction process (see Figure 3). 1.
For each memory reconstruction node from LTM, compute the total contextual relativity weights from the current set of Binary Information Fragments associated with the memory reconstruction node. These contain Knowledge Relativity Threads related to the concepts involved it the memory reconstruction:
߰݅, = ݐΣ∈ ݆ݔΓ ݆ܹ݅ ݐ 2.
Normalize the values from each Topical Map that relate to a concept within the Conceptual Ontology, related to LTM Binary Information Fragments. i.e., we find the maximum value above and divide by all the individual values by the greater of the maximum and the F-matrix threshold F Θ t . This is done to ensure that the memory feed-forward signals are not amplified in subsequent normalization steps to avoid catastrophic interference in the memory reconstruction process:
Ψ݅,= ݐ 3.
߰݅,ݐ ݉ܽݔ൫݉ܽݔ൫݆߰ , ݐ൯, ܮΘ ݐ൯
Analogous to step 1 – only for Short Term Memories:
߶݅, = ݐΣ݆ =Δ ݅−1 , > ݐ0 4.
Analogous to step 2 – only for Short Term Memories:
Φ݅,= ݐ 5.
߶݅,ݐ
݉ܽݔ൫݉ܽ ܱܥ∈ ݆ݔ൫߶݆ , ݐ൯, ܵ Θ ݐ൯
The memory reconstruction process is an iterative process (assessing whether the reconstruction process meets the conscious memory’s requirements and constraints). Here, the Knowledge Relativity Threads from LTM, and STM are put through filters to produce a generalization gradient for their relativity to the memory to be constructed, based on the Semantic, Self-Organizing Topical Maps, utilizing adapted gravitational theory [Crowder and Carbone 2011b]. This gives them membership values between 0 and 1:
u v ߯݅, = ݐΨi,t Φj,t
6.
In step 6, we look for those memory fragments that are most relevant within the context of the Conceptual Ontology and Information Base:
Φ݅,= ݐ 7.
߶݅,ݐ
݉ܽݔ൫݉ܽ ܱܥ∈ ݆ݔ൫߶݆ , ݐ൯, ܵΘ ݐ൯
Once we have all of the Binary Information Fragments from LTM and STM that are relevant, given the Topical Maps that may be relevant, we look for those Topical Maps that make the most sense, given the strength of their membership functions for each set of combined ST and LT memories. Again, this is based on ISAAC’s current Conceptual Ontology (CO):
ߨ݇, ݆ܺ ܱܥ∈ ݆ݔܽ݉ = ݐ, ݐ, 1 ≤ k ≤ Q
8.
This is in the process in case of an “unexpected” result, based on cues from procedural, emotional, and spatio-temporal memory creation. This triggers if the actual combined memory is unexpected, given the current temporal context of ISAAC’s past experience and clues. This may be an inference for a new concept not currently in ISAAC’s Conceptual Ontology. In this case, we get a compressive non-linearity that maps the required information into a new concept within the Conceptual Ontology:
ܳ
= ݐܩΣ݇=1 9.
ߨ݇,ݐ
ൗܳ
Finally, we simply choose the winners of the nodes and provide the constructed memory out to ISAAC’s Cognitive Perceptrons for use by ISAAC’s conscious processes and ISAs:
݅,= ݐ
݂൫ܺ݅, ݐ, ݐܩ൯ Σ݆ ∈݂ ܱܥ൫ܺ݅, ݐ, ݐܩ൯
Figure 3 – ISAAC’s Memory Reconstruction Architecture
5
Conclusions and Discussion
What we have described is a memory encoding and reconstruction process that will allow artificially intelligent life forms to process, encode, store, and retrieve information (memories) similar to human memory processing. Much work is needed to verify and validate these processes and research will continue over the next year to build simplistic versions of this system for initial testing.
References: 1.
2. 3.
4. 5.
6.
7.
8.
9.
10.
11.
12.
13. 14.
Bishop, C.M. 1995. Neural Networks for Pattern Recognition. Oxford University Press. Crowder, J. A. 1996. Reusable Launch Vehicle Automated Mission Planning Concepts. NASA Report X33/RLV-MP-PHII-1996-005, Lockheed Martin, Littleton, CO. Crowder, J. A. 2002. Machine Learning: Intuition (Concept Learning) in Hybrid Neural Systems, NSA Technical Paper- CON-SP-0014-2002-06 Fort Meade, MD.
Crowder, J. A. 2010a. The Continuously Recombinant Genetic, Neural Fiber Network. Proceedings of the AIAA Infotech@Aerospace-2010, Atlanta, GA. Crowder, J. A. 2010b. Anti-Terrorism Learning Advisory System (ATLAS): Operative Intelligent Information Agents for Intelligence Processing. Proceedings of the AIAA Infotech@Aerospace-2010, Atlanta, GA. Crowder, J. A., 2010c. Flexible Object Architectures for Hybrid Neural Processing Systems. Proceedings of the 11th International Conference on Artificial Intelligence, Las Vegas, NV. Crowder, J. A., and Carbone, J. 2011a. Recombinant Knowledge Relativity Threads for Contextual Knowledge Storage. Proceedings of the 12th International Conference on Artificial Intelligence, Las Vegas, NV. Crowder, J. A., and Carbone, J. 2011b. Transdisciplinary Synthesis and Cognition Frameworks. Proceedings of the Society for Design and Process Science Conference 2011, Jeju Island, South Korea. Crowder, J. A., and Carbone, J. 2012. Knowledge Binary Information Fragment Encoding for AI System Memories. Proceedings of the 13th Annual International Conference on Artificial Intelligence, Las Vegas, NV. Crowder, J. and Friess S. 2012a. Artificial Psychology: The Psychology of AI. Proceedings of the 3rd International Multi-Conference on Complexity, Informatics, and Cybernetics, Orlando, FL. Crowder, J. and Friess S. 2012b. Artificial Neural Sensory/Short-Term/LongTerm/Emotional Memory Integration for Autonomous Systems. Processings of the AIAA Space 2012 Conference, Pasadena, CA. Crowder, J. 2012a. Cognitive System Management: The Polymorphic, Evolutionary, Neural Learning and Processing Environment (PENLPE). Proceedings for the AIAA Infotech@Aerospace 2012 Conference, Garden Grove, CA. Crowder, J. 2012b. The Artificial Cognitive Neural Framework. Proceedings for the AIAA Infotech@Aerospace 2012 Conference, Garden Grove, CA. Crowder, J. 2012c. Cognitive System Management: The Polymorphic, Evolving, Neural Learning and Processing Environment. Proceedings for the AIAA Infotech@Aerospace 2012 Conference, Garden Grove, CA.
15. Crowder, J., Raskin, V., and Taylor, J. 2012. Autonomous Creation and Detection of Procedural Memory Scripts. Proceedings of the 13th Annual International Conference on Artificial Intelligence, Las Vegas, NV. 16. Raskin, V., Taylor, J. M., & Hempelmann, C. F. 2010. Ontological semantic technology for detecting insider threat and social engineering. New Security Paradigms Workshop, Concord, MA. 17. Rosenblatt F. 1962. Principles of Neurodynamics. Spartan Books. 18. Scally, L., Bonato M., and Crowder, J. 2011. Learning agents for Autonomous Space Asset Management. Proceedings of the Advanced Maui Optical and Space Surveillance Technologies Conference, Maui, HI. 19. Taylor, J. M., & Raskin , V. 2011. Understanding the unknown: Unattested input processing in natural language, FUZZ-IEEE Conference, Taipei, Taiwan. 20. Hong, J. and Landay, J. 2001. An infrastructure approach to context-aware computing. Human-Computer Interaction, vol. 16, pp. 287-303. 21. Howard, N. and Qusaibaty, A. 2004. Network-centric information policy. Proceedings of the Second International Conference on Informatics and Systems. 22. Ejigu, D., Scuturici, M., and Brunie, L. 2008. Hybrid approach to collaborative context-aware service platform for pervasive computing. Journal of Computers, vol. 3, pp. 40. 23. Zadeh, L. 2004. A note of Web Intelligence, World Knowledge, and Fuzzy Logic. Data and Knowledge Engineering, vol. 50, pp. 291-304.
