vertical integration of computational architectures - the mediator problem
TRANSCRIPT
Vertical Integration of Computational Architectures: The mediator problem
© Yehor Churilov 2016
Statements
1. The problem of architectural mediatorBetween the varieties of states of consciousness at sensory level and at level of higher cognitive functions there is a major gap that can not be effectively bridged only by interaction between the two architectural tiers (CA, SA) and thus requires an introduction of a conceptually independent architectural layer.
2. The problem of the integration platform heightIncrease of a cognitive ability demands increasing a height of a platform
at which the architectures are fully integrated
A bit of theory
Symbolic Approach (SA)• operates with clearly separated
information objects - symbols• operations: formal logical
inference, symbols produce symbols• derive from language or
correlates well with it• time- and agent-independent
Connectionist Approach (CA)• knowledge is a distributed
network state• operations: self-organizing
networks, dynamics of global system states• well correlated with the sensory
apparatus and the associated tasks• embodied, time- and agent-
dependent
Now there are two dominant approaches to the representation of knowledge and the construction of AI:1) symbolic or cognitivist vs 2) emergent or connectionist*
Reasons and incentives of hybridization in artificial systems
Incentives
1. Compensate for the current or the principal disadvantages of individual approaches
2. Reduce total complexity3. “Code Reuse" - a reduction of costs of
theoretic and implementation models4. Biological alignment5. Manage a variety of symbolic
representations wider beyond the limit that a purely symbolic architecture practically allows
6. Better understand cognitive system than a purely connectionist architecture allows
7. To achieve functional integrity, the need for which is understood consciously or intuitively
Reasons
1. Initially non-holistic approach to the study and realization of cognitive systems (cognitive science vs neuroscience)
2. Unavoidable limitations in knowledge and understanding of brain and mind make available only partial representations and implementations
3. Uneven development of different directions of research and technology
4. Primordial domination of logisism and symbolic approach
5. Shortage of computing power and/or institutional capacity to implement
Hybrid systems classification*a) Unified hybrid systems – purely connectionist systems
implementing symbolic models. Early experiments with localist representations; “radical connectionism” [Dorffner 1996]
b) Transformational hybrid systems – ANN and symbolic processor are separate, some data transport is in-between. “Function-replacing” [Goonatilake and Khebbal 1995].Federated approach, accoording to ISO 14258
c) Modular hybrid systems – the different architectures may be interleaved, with feedback channels
* Согласно McGarry K. et al - Hybrid Neural Systems. From Simple Coupling to Fully Integrated Neural Networks, 1999.
All cognitive functions of a biological agent are integrated. There’re no «hybrids» *
* This determination depends on boundaries defined for a system. For the purposes of this presentation I introduce the opposition between "non-hybrid" natural and "hybrid" artificial systems.
Vertical integration in the case of cognitive systems
Top level
Middle level
Bottom level
"Vertical" axis is an axis where levels of abstraction of states are distinguished, from a wide variety of signals or action at the bottom to a small variety of necessary results at the top.
"Horizontal" axis is an axis of combining components of the same operational scale.
A number of levels is rather pragmatic choice, depending on the need and ability to manage a signal transformation process. Each level represents a grouping of operations of certain order those have similar life cycle, operate on data of a similar scale, and so forth.
Vertically integrated system is a holistic implementation of certain cognitive function, “from top to bottom“. In modern hybrid system such disposition exists as a fact, but is not conceptually determined.
North
South
SymbolicArchitectures
Mediators?
ConnectionistArchitectures
Symbolic Architectures operate on more generalized states: symbols taken from language or formal systems.
Neural networks can deal with a wide varieties, similar to sensory apparatus of animals and humans.
Hybrid systems combine these approaches in some way. (Enterprise Architecture has not a a notion of «hybrid», there’re south-bound architectures, north-bound architectures, enterprise bus. Business domain is elaborated better?)
"Mediators" are indicated here, but in the modern science and engineering they are not specifically conceptualized. In fact, state-of-the-art is the two-tier architecture. See below.
Vertical integration in the case of cognitive systems
Higher representations and functions
Primary perception
Complete model of a vertically integrated architecture
SAMediators?CA
The direction of decreasing diversity
“Pure” CA systems CA machine human
“Pure” SA systems SA machinehuman
The gap
features conceptsobjectsraw data
Having expanded the boundaries of cognitive system, we get a holistic architectural invariant with individual parts possibly delegated, differently in one or another implementation.
Hybrid systems SA machineCA machine smth
Theses on a holistic architecture1. Necessity for embodiment. In order to approach a human level of cognitive
ability an artificial agent requires to be embodied in an environment with a effective variety, comparable with that of environment of human society
2. Domination of connectionism. The need for computationally intensive work with the highest level of variety of states leads to the fact that the connectionist approach will prevail. Still, ANN is a special case of connectionist computational model
3. Irremovability of symbolic approach. Human thinks in symbols, thus for human-machine interaction to continue, symbolic architecture can not be removed from the stack. BCI could help only partially, because rational thinking still needs compact symbolic stimuli
4. Statement on the scale gap. Symbolic architectures tend to operate with a variety of O(mind), whereas connectionist architectures tend to the limit of O(brain). Obviously, O(mind) << O(brain), and this indicates the presence of a scale gap.
5. The need for architectural mediator. Increasing cognitive ability of vertically integrated architectures requires the introduction of at least one more conceptually independent level to manage variety, which could bridge the scale gap.
To formulation of the problems ofvertical integration
Mind is understood in terms of logical positivism, as a language processor
Brain is understood in terms of physicalism as 4D-macroobject consisting of objectively identified spatial elements and physical states
Concepts and the denoting language elements appear as a result of repeated and purposeful generalization of a large number of perceptual microstates - the construction of macrostates from microstates.And also through abstraction – purposeful elimination of irrelevant parts of the spectrum.
As a result, the signal variety decreases by many orders of magnitude, which is expressed by the formula O(mind) << O(brain)
Independent development of symbolic or connectionist architectures targeted to reduce the scale gap will required to introduce additional constructs that do not fit well into either paradigm. Their consistent introduction will cause difficulties and conceptual compromises; avoidance of the introduction will stall the development.
Statement of the scale gap. Symbolic architectures tend to operate with a variety of O(mind), whereas connectionist architectures tend to the limit of O(brain). Obviously, O(mind) << O(brain), and this indicates the presence of a scale gap.
1.The problem of architectural mediator
Between the varieties of states of consciousness at sensory level and at level of higher cognitive functions there is a major gap that can not be effectively bridged only by interaction between the two architectural tiers (CA, SA) and thus requires an introduction of a conceptually independent architectural layer.
SA
CA
Architectural Mediator
Enterprise Architectures: parallels
Images: http://www.simcrest.com/blog/BlogEntry/35
Evolution
Challenges1. Increase of amount of initial data, of means and methods of
presentation, of processing, of transportation, of the number of manufacturers ...
2. Increasing complexity of interconnectivity3. Manageability problems...
“Business Logic Tier” is infrastructure element, not visible “from outside” the system. Data are below, solutions are above. But it bears more and more of functional burden.
In the stack of holistic cognitive system, this item is not conceptually identified at the moment.
What is “integration”?Integration (in terms of ISO 14258) is a form of interaction between systems, when they exchange entity instances from a single executable model or conceptual scheme, without additional interface transformation and effort. Integrity means that a common computing platform is built up to a certain level in the stack. For modern hybrid systems this is the level of computational environment or programming language in which they are implemented.
System 1 System 2Common Data Model
h
2. The problem of the integration platform heightIncrease of a cognitive ability demands increasing a height of a platform at which the architectures are fully integrated
SA
CA
Architectural mediator
CA
Symbolic computations
Architectural mediator
Connectionistrepresentations
Representationalmediator
Symbolicrepresentations
?
Transposition of modular
integration to
representationalMod
ular
inte
grati
on
Approach to a solution
The two phases
1. Solution to the architecture mediator problem through the introduction of conceptually independent presentation layer. The main element of this is multiscale attentional patterns, binding primary generalizations of connectionist representations with highly-generalized categories and ontological elements of symbolic representations.
2. Facilitating of platform integration problems through unification of computing on a metaconnectionist basis. "Radical connectionism“ expanded: the use of highly scalable distributed representations for the full range of operations. But, perhaps, not ANN, but networks with elements of higher order, where more complex computing organelles will act as members.
Thank you!Yehor Churilov, Minsk