Part I: Representation Mediated By Neural-like Computation
Information Embedding And Iconic Pattern Completion
The model postulates that the pattern representation aspect of “Access-Consciousness” function involves in-time universal approximation, by way of pattern completion.
There are distinct advantages to iconic representation:
- It is implementable as a locally self-organizing and error pruning computational scheme (i.e. amenable to distributed processing of distributed information,
minimizing the overall communications flux)
- It is generally information conserving, supporting in-system regeneration and re-evaluation
Generative predictive processing theories tend to emphasize the potential for information compression and distribution, which applies also to the proposed model. This particular perspective is motivated by the potential for concurrent, reciprocally superimposed, pattern factorization and regeneration.
A notional sensory perimeter is a useful conceptualization.
An agent that (from within) can project upon it’s sensory perimeter patterns to match environmental projections, is in possession of a complete model of the projecting sources.
Stable representations of that which would project upon the agent’s perimeter, predicting the dynamical trajectory of such projections, implement the complete informational set available to the agent – inclusive of hierarchies of factored causes and their combining relations (e.g. as learned by Bayesian inference. Well known limitations e.g. heuristics to disambiguate and produce 3D world model constructs, can be pragmatically accommodated in the proposed regime).
Accordingly, the learning phase of neural-like (computational) pattern processing may involve the casting of an isomorphic, invertible, embedding transform. Cohesive and contextual spatiotemporal relations may be extracted and embedded in a slowly conducting plastic medium, reduced (by weighted directed interference) to morphing configurations of information depleted, near isotropic, activity.
Information extraction, or “factorization”, is likely implemented in a progressively subtractive manner.
An inverse enacting (biasing, filtering) operation would facilitate computational recognition – reconstitution of iconic world model components matching projection patterns upon an agent’s sensory perimeter (imagination may be facilitated by tonic afferent sensory noise).
Thus an active representation is a bidirectional process that involves co-sustaining aspects of patterned and isotropic activity.
Active pattern representations will have an information-depleted aspect that is co-active with an iconic, literal, representation of projections on the sensory perimeter. (The information depleted aspect consists of a compressed point of activity or a population of units exhibiting synchronous oscillations or stochastic tending-to-isotropic activity, over a spatiotemporal configuration).
In “representational space” the two aspects should tie with the same contextual relations.
While pattern phase matching is smeared, the relations supporting pattern composition are conserved (to the degree that they can be predicted). This would be especially apparent with accommodation for representational invariance (e.g. convolutional models).
A Virtual Focal Point Suggestive Of An ‘I’ Primitive
Assuming phenomenality is (somehow) associated with network activity then by this model the correlates of the basic experience that permeates what we identify as Self (a tonic feeling grounding diachronic unity) could be associated with the hypothesized morphing configurations of near-isotropic activity.
Effective Representational Continuity
The ability to maintain on-going adaptation of internally generated representations to optimally match dynamical patterns of sensory input relies on the accommodation of singularities (and extreme non-linearity) in the continuous universal patterning relational transform (the biological prevalence of opponent center-surround receptive fields conforms with this proposition).
Sensory Resolution Limits Constrain Representational Capacities
Patterns consist of spatiotemporal structure and dynamics. Behaving agents rely on selected sensory causes with behaviorally relevant spatial and temporal pattern resolution (grain) and horizons.
Representational capacity is constrained on both measures: Sensitivity (lower bound of dynamic range, signal to noise ratio) constrains the predictive horizon. Crucially, a modeled sensory projection is cut off from the fine details of it’s interacting cause at the sensory interface’s resolution limit.