Key Considerations for Artificial Cognitive Systems
Key Considerations for
Artificial Cognitive Systems (Representations and Architectures)
Author: Susanne Lomatch
In the process of completing a more comprehensive future primer on artificial cognitive systems (ACS), I present here some key considerations for cognitive system representations and associated architectures.
Key considerations for cognitive representations and associated
architectures include:
• Recursive, recurrent and nested
hierarchically distributed structure
• Cognitive neurological function and functional structure is manifested operationally as a mixture of non-nested and nested hierarchies at both the neurocellular and global levels [Feinberg, 2009]
• In nested hierarchies, constituent parts not only contribute to higher levels, nested parts at any level comprise those higher levels
• Non-nested hierarchies provide centralization and convergence at levels
• Both non-nested and nested hierarchies contribute to emergent properties and downward constraints (bottom-up and top-down contributions from lower-level interactions and higher-level constraints)
• Recursive properties are manifest in language, speech and other higher-order abstract cognitive processes
• Represent long-range dependencies, constituent structure, semantic context, etc.
• Recurrent connectivity between multimodal cortical areas likely accounts for dynamic recursive functionality
• Physiological and functional structure likely differ, spatially and temporally – functional structure is subject to dynamics at shorter time scales, while physiological structure remains relatively stable [Sporns, 2011]
• Collective, coordinated or synergistic
structure
• This includes self-organization, but with constraints and convergence properties that promote synergistic modes that enable the system to act as a single coherent unit: emergence with constraint and convergence
• There may also be collective emergent structure from dynamic interactions, i.e., phase locking or synchronization dynamics between functional neural clusters, see [Edelman, 2000] and [Sporns, 2011], and spatiotemporal dynamics below
• Hetero-associative and “unconventional” memories
• Hetero-associative memories are generally ignored, with a focus on auto-associative and non-associative memories
• “Unconventional” memories such as a proposed “value-category memory” by [Edelman, 2000] that is driven partly by biochemical changes/signaling
• Cross modal processing
• Key for the fusion of knowledge/information from disparate functional processing areas, which we know to occur at various levels of hierarchical processing
• Recent studies indicate that cross modal or multimodal processing may occur at stages earlier than originally thought, and may form an integral part of both simple and complex sensory processing [Ghazanfar, 2006]
• Spatiotemporal dynamics
• Dynamics between and across hierarchical levels – this may motivate certain approaches such as topological representations
• Reciprocal connections involving forward and backward dynamics may be asymmetric (stimulatory vs. modulatory), among other topological configurations
• Recurrent connections ensure binding dynamically occurs across activities of widely spaced neuronal groups in different brain maps (Edelman and Tononi call this “reentry” [Edelman, 2000])
• Plasticity (both neuronal and synaptic)
• This consideration is a significant one, and linked to memory and learning
• Practical application of this to artificial systems will require coordinated breakthroughs in neurobiology, bioengineering and materials science, perhaps with some help from molecular biology (very recent breakthroughs in synthetic genetic polymers may be one lead, especially if repeatable replication and control through selectivity could be achieved, without the need for natural “seeds”)
• Ability of biological elements or systems that are structurally different to perform the same function or yield the same output; contrasted with redundancy, which occurs when the same function is performed by identical elements, degeneracy, which involves structurally different elements, may yield the same or different functions depending on the context in which it is expressed [Edelman, 2000 & 2001]
• Promotes functional “homeostasis” of cognitive processes, and likely works in conjunction with convergence and constraints within a hierarchical network
• Integration and Discrimination
• The brain is essentially an "information integration and discrimination” processor, able to integrate and select among a vast number of states (many from low-level sensory processing) and to further integrate those states as part of the discrimination process, forming higher-level discriminant states; this integration/discrimination process depends on many of the factors outlined above, notably spatiotemporal dynamics within the structure outlined, as well as constraints from value learning and, arguably, causal reasoning (to what extent the latter is employed is an open-ended research question)
• Classical multisensory areas such as the superior temporal sulcus (STS) are regions where signals from multiple sensory systems converge and are integrated; stimulus integration of a brain region such as STS does not depend solely on convergent neural connectivity and intrinsic flow of information; integration can be greatly enhanced by coordinated movement of the body [Sporns, 2011]; this view also couples with the view that environmental interaction greatly affects the integration/discrimination process – it is “integral” to its functioning and development
• Embedded reasoning and learning modes
• Again, another significant consideration, given the failure of many symbolic or connectionist/distributed graphical representations to enable tractable learning and reasoning or inferencing: many learning/reasoning approaches for more general representations are intractable or NP-hard problems
• Practical representations and architectural models that implement learning procedures and embedded reasoning in a natural, tractable manner are required, and may involve completely new approaches
• Training/learning involving a simulated “value system” of neuromodulator signaling is an interesting, biologically-inspired approach ([Edelman, 2000] and [Friston, 1994]) that may lead to greater tractability through constraint; this type of learning is fundamentally reinforcement learning, but it could be integrated with unsupervised or semi-supervised learning modalities
• Embedded sensory simulator and predictor
modes
• Capturing anticipatory processes
• (The brain as a “reality model simulator and predictor”)
• See memory prediction framework(s) and generalizations ([Hawkins, 2009] and [Friston, 2005-10])
• Hybrid representation approaches may be
required…
• Recursive, recurrent and nested hierarchically distributed representations (see above)
• Nonlinear dynamical representations
• Complexity representations
• Convolutional representations
• Topological representations
• Stochastic, indeterministic and deterministic representations
• Continuous and distributed discrete representations
• Sparse encoding and dense encoding schemes
• (Hybrid models have been found to be increasingly useful for practical implementation and application in machine vision and NLP/ASR – the “inference” is that the same might be true for cognitive systems)
• “Neurocognitive networks” – are biologically inspired, particularly by processing in the thalamocortical and corticocortical system, nominally aimed at the neural circuit/network scale, e.g. [Mesulam, 1998], [McIntosh, 2000], [Bressler, 2006-10], [Izhikevich, 2008], [Sporns, 2011] – some of the efforts within this thrust can be generalized to more algorithmic approaches that would essentially be similar to the hierarchical temporal memory model [Hawkins, 2009] discussed in State of AI/Part 3, or more complex generalizations of it (specifically see [Friston, 2005-10] and [Grossberg, 2007]); for more unconventional approaches and insights see [Tononi, 2004] and [Granger, 2006]
• Radically new approaches may be required…
• Perhaps anything that doesn’t remind one of traditional computational representation models, and this includes almost all approaches out there to date (remember, the brain is not a von Neumann machine, no matter how hard we try to model it as one; but knowledge representations are most certainly integral to the foundation of any architecture, even if they are “unconventional”)
• Approaches that reconcile the continuum/discretized and global/nonlocal/local observables, hierarchical structure and spatiotemporal dynamics
• (No, I am not falling into the Penrose trap here – I am simply advocating that we widen the field to integrate key observables that have been ignored, and this may not require an exotic quantum computing paradigm advocated by Kurzweil – the emergent but constrained and convergent dynamic classical properties of the biological brain system have much to answer for here – readers searching for some interesting data might take a look at a few “brain state-space” maps created from invasive BMI data measuring neuronal ensemble phenomenology from widely distributed points across the brain, showing a high degree of global dynamical functional integration, see [Nicolelis 2011] and refs. therein – this data supports the concepts proposed by [Edelman, 2000] of dynamic integration, reentry and degeneracy, and further, the concept of a “value-category memory” that is modulated by diffusely projecting neurotransmitter “value” systems originating from subcortical structures or “specific nuclei” – see for example, the ‘dynamic core hypothesis,’ measures of complexity of functional clusters therein, the connections and interactions of the thalamocortical dynamic core to cortical appendages, and the interactions with the neuromodulatory value systems of subcortical specific nuclei [Edelman, 2000])
Note to Readers: Those looking for more detail on knowledge representations and learning may want to review the Primer on Knowledge Representations & Acquisition and the Primer on Machine Learning, and further the Primer on NLP for how both are used in natural language processing (NLP) and automatic speech recognition (ASR). (I am also working on a primer on machine vision, which will also include such detail for that application, and plan a future primer on cognitive systems and architectures.)
(Disclaimer: This primer is meant to inform. I encourage readers who find factual errors or deficits to contact me (click on contact link below). I also welcome constructive and friendly comments, suggestions and dialogue.)
References
and Endnotes:
[Feinberg, 2009] “From Axons to Identity: Neurological Explorations of the Nature of the Self,” T.E. Feinberg, Norton & Co., 2009.
[Edelman, 2000] “A Universe of Consciousness,” G.M. Edelman and G. Tononi, Perseus Books, 2000.
[Friston, 1994] “Value-Dependent Selection in the Brain: Simulation in a Synthetic Neural Model,” K.J. Friston et al., Neuroscience, vol. 59, 1994.
[Edelman, 2001] “Degeneracy and complexity in biological systems,” G.M. Edelman and J.A. Gally, Proceedings of the National Academy of Sciences, vol. 98, 2001.
[Mesulam, 1998] “From Sensation to Cognition,” M.M. Mesulam, Brain, vol. 121, 1998.
[McIntosh, 2000] “Towards a Network Theory of Cognition,” A.R. McIntosh, Neural Networks, vol. 13, 2000.
[Tononi, 2004] “An Information Integration Theory of Consciousness,” G. Tononi, BMC Neuroscience, vol. 5, 2004.
[Bressler, 2006-10] “Operational Principles of Neurocognitive Networks,” S.L. Bressler and E. Tognoli, International Journal of Psychophysiology, vol. 60, 2006. “Large-Scale Brain Networks in Cognition: Emerging Methods and Principles,” S. Bressler and V. Menon, Trends in Cognitive Sciences, vol. 14, 2010.
[Granger, 2006] “Engines of the Brain: The Computational Instruction Set of Human Cognition,” R. Granger, AI Magazine, vol. 27, 2006.
[Ghazanfar, 2006] “Is the Neocortex Essentially Multisensory?” A.A. Ghazanfar and C.E. Schroeder, Trends in Cognitive Sciences, vol.10, 2006.
[Grossberg, 2007] “Towards a Unified Theory of Neocortex: Laminar Cortical Circuits for Vision and Cognition,” S. Grossberg, Prog. Brain Res., vol. 165, 2007.
[Friston, 2005-10] “A Theory of Cortical Responses,” K. Friston, Phil. Trans. R. Soc. B, vol. 360, 2005. “Hierarchical Models in the Brain,” K. Friston, PLoS Computational Biology, vol. 4, 2008. “The Free-Energy Principle: A Unified Brain Theory?,” K. Friston, Nature Reviews Neuroscience, vol. 11, 2010.
[Hawkins, 2009] “Towards a Mathematical Theory of Cortical Micro-circuits,” D. George and J. Hawkins, PLoS Computational Biology, vol. 5, 2009.
[Izhikevich, 2008] “Large-Scale Model of Mammalian Thalamocortical Systems,” E.M. Izhikevich and G.M. Edelman, Proc. Natl. Acad. Sci., vol. 105, 2008.
[Sporns, 2011] “Networks of the Brain,”
Olaf Sporns, MIT Press, 2011.
[Nicolelis, 2011] “Beyond Boundaries: The New Neuroscience of Connecting Brains with Machines - and How it will Change Our Lives,” Miguel Nicolelis, Henry Holt & Co., 2011.