Source-linked AI summary
A Structural Theory of Cognitive Representation and Problem Solving,Contexts, Invariance, and the Knowledge Space
Antal Jakovác, András Telcs
TL;DR
The paper addresses the lack of explicit representational structure for abstraction, invariance, and task-relevant regularities. It formalizes contexts and organizes concepts and typed procedures in a Knowledge Space; its examples show that reduced representations can simplify admissible solutions even for a fixed weak solver.
Problem
Many data-driven systems achieve strong predictive performance, but their learned representations often leave abstraction, invariance, and task-relevant structure implicit.
Source: §1 · p. 2Method
The paper formalizes contexts and uses coupled Concept and Procedure Graphs to organize concepts and typed operations, with simple solvers for problem-solving demonstrations.
Source: §1 · p. 2, §2.1 · p. 5, §6 · p. 19Results
In the nuisance-coordinate example, admissible solutions in the reduced representation use |S| variables, while the unreduced context requires accounting for irrelevant variation.
Source: §6.2 · p. 21Takeaways & Limitations
Contexts, invariance recognition, and representative selection provide a structural basis for suppressing irrelevant distinctions and exposing simpler task-relevant regularities.
Source: §7 · p. 22The examples are intentionally limited and use a fixed weak solver; typing only expresses compatibility of entities under operations.
Source: §5.1 · p. 16, §7 · p. 22
Abstract
from arXiv · showhide
Learning and problem solving depend critically on the structure of internal representations. While many modern data-driven artificial systems achieve strong predictive performance, their learned representations often lack explicit structure for expressing abstraction, invariance, and task-relevant regularities. We propose a minimal structural framework in which representational operations relevant to problem solving, such as context formation, invariance recognition, representative selection, abstraction, and procedural reuse, are made explicit. The central notion is that of a \emph{context}, formalized as a partition of a subset of an underlying state space, which fixes the distinctions, granularity, and form in which a problem can be posed. Within this setting, invariance recognition and representative selection are treated as fundamental representational operations. The framework is realized as a Knowledge Space composed of two coupled graph structures: a Concept Graph that hosts constructed and refined concepts, and a Procedure Graph that encodes typed operations over representations. Together, these structures provide a minimal cognitive-representational algebra for operating on representations without assuming sophisticated inference, learning, control, perception, or motor mechanisms. Using simple illustrative examples and a finite weak-solver demonstration, we show that appropriate representational organization can simplify the form and scope of admissible regularities, even when problem solving is carried out by a fixed and limited solver. The contribution of the paper is structural rather than algorithmic: it identifies representational prerequisites for abstraction, invariance, and procedural reuse in problem solving, and states explicit success and failure conditions for the weak-solver setting.
1 Introduction
The paper makes representational structure explicit through contexts, which specify meaningful distinctions and task granularity. It organizes concepts and typed procedures in a Knowledge Space and treats invariance and representative selection as operations for simplifying problem solving.
1 Introduction
- Contexts are finite partitions of state-space subsets that determine meaningful distinctions, task granularity, and the representational space for admissible solutions.They specify structural conditions under which problem solving is possible.Source: §1 · p. 2
- Invariance recognition and representative selection act on contexts to suppress irrelevant distinctions and expose simpler task-relevant laws.These are structural operations, not learning objectives or statistical effects of data or loss functions.Source: §1 · p. 2
- The Knowledge Space couples a Concept Graph for constructed and refined concepts with a Procedure Graph for operations, while keeping procedures weak and fixed.The Concept Graph is a reusable host rather than a fixed ontology; representational structure supplies the primary problem-solving power.Source: §1 · p. 2
- The paper specifies a representational core and explicit success and failure conditions, rather than a full learning, control, perception, motor, or causal architecture.Its finite weak-solver demonstration examines when representation reduces the burden on a fixed solver.Source: §1 · p. 3
2 Related Work and Positioning
The paper locates its contribution at the representational level, before particular learning algorithms or control architectures are specified. It formalizes contexts and representational operations while distinguishing this structural layer from cognitive architectures, predictor-level invariance methods, and existing schema or symbolic approaches.
2 Related Work and Positioning
- The framework treats representation as a prior structural question, distinct from specifying learning algorithms or control architectures, following a representational-level emphasis associated with Newell and Simon and Marr.The related-work discussion frames representational choice as relevant to problem-solving tractability, independent of search strategy.Source: §2 · p. 3
2.1 Cognitive architectures, schemas, and procedural memory
- The Knowledge Space is a representational substrate, not a complete cognitive architecture with a control cycle, perception-action loop, or behavioral predictions.It is presented as a structural layer that may support architectural mechanisms, rather than replace them.Source: §2.1 · p. 3
- Classical schema theories describe organized structures, frames, and scripts [18, 19, 20]; here, contexts formally fix task distinctions and granularity, while schemas are stable representational and procedural patterns reconstructed from features, relations, procedures, and solution traces.Source: §2.1 · p. 4
- Procedure Graph nodes represent typed operations over representations, which can be reused, composed, and generalized from successful traces rather than serving as production rules in a complete interpreter.The Concept Graph and Procedure Graph form a structural long-term store, while task codes and active traces are transient working structures; the memory analogy is not a claim of psychological completeness.Source: §2.1 · pp. 4–5
- Unlike predictor- and loss-level invariance methods or prescribed equivariant architectures [26] [28], this framework treats invariance as a task-relative equivalence over contexts.It does not explicitly model causality, instead identifying minimal structural primitives described as prerequisites for later causal reasoning.Source: §2.1 · pp. 5–6
3 Foundations
The foundations treat world states as semantic scaffolding and contexts as the primary representational units. Contexts partition object classes, can be refined or represented in several ways, and support features that induce further partitions.
3.1 Universe, States and Time
- The formal foundations assume a complete state determines future states, yielding a deterministic, memoryless description of time evolution, while treating the Universe and its evolution as theoretical constructs.Time labels spatial instances and has an additive structure.Source: §3.1 · p. 7
3.2 Object classes
- Object classes are subsets of the state space that collect states differing only in irrelevant details, providing the basic objects for an agent's descriptions.A context then partitions an underlying class to express finer distinctions, such as dog breeds within dogs.Source: §3.2 · p. 8, §3.3 · p. 8
3.3 Context; 3.3.1 Contexts instead of states
- A context is a finite partition of an object class; its cells determine what can be distinguished and frame the task or problem posed.Contexts, rather than states, are treated as primary, and may be refined as needed.Source: §3.3 · p. 8, §3.3.1 · p. 9
3.3.2 Refinement and induced maps
- Refinement makes each finer cell fall within a unique coarser cell, and common refinement combines contexts by intersecting their cells.The common refinement is the finest context that refines both input contexts.Source: §3.3.2 · pp. 9–10
3.4 Characterization of contexts; 3.4.1 Direct representation; 3.4.2 Global coordinates; 3.4.3 Local coordinates
- Contexts can be represented directly, by global coordinates whose common refinement recovers the context, or by local refinements organized as a tree with prefix descriptions.Direct cell indexing is formally complete but infeasible in high dimensions; local-coordinate descriptions typically pair a value with the tested property.Source: §3.4.1 · p. 10, §3.4.2 · p. 10, §3.4.3 · p. 11
3.5 Features
- A feature is constant on each context cell and induces a partition by grouping states with the same feature value.Features extend to finer contexts by constancy on sub-cells; relevance is related to rough-set consistency, not claimed as a new formal notion.Source: §3.5 · p. 12
4 The Knowledge Space as Two Coupled Graphs (2G)
The Knowledge Space is a minimal structural substrate composed of a Concept Graph and a Procedure Graph. The graphs organize reusable concepts and typed representational operations, including equivalence and generalization, without modeling a complete control or planning architecture.
4 The Knowledge Space as Two Coupled Graphs (2G)
- The Knowledge Space couples a Concept Graph of conceptual entities and relations with a Procedure Graph of typed operations linked to conceptual domains and ranges.Together, they provide a minimal substrate for defining basic cognitive primitives.Source: §4 · pp. 12–13
4.1 Concept Graph (CG)
- The Concept Graph records reusable distinctions, including classes, context cells, and derived constructs, using equivalence and generalization relations.Contexts provide the semantic scaffolding for forming and refining concepts but are not themselves graph objects.Source: §4.1 · p. 13
4.2 Procedure Graph (PG)
- Procedure Graph nodes are possibly partial typed functions over concept-structured inputs and outputs; equivalence identifies procedures differing only by input/output relabeling and inducing isomorphic domain partitions.The equivalence condition is expressed by bijections making the procedure diagram commute.Source: §4.2 · p. 14, Figure 1 · p. 15
- Procedures can include representative selection, abstraction, derived-relation construction, and simple inference or transformation; procedural generalization reflects collapsed distinctions and coarser domain partitions.The graph organizes procedures structurally, not by planning, control, or execution order.Source: §4.2 · pp. 15–16
- The framework localizes invariance and representative selection as representation-level operations and evaluates their effects with simple solvers, without introducing explicit goals, executive control, or long-horizon planning.Transient task-specific graph copies and transfer of successful solutions to permanent memory are described, while detailed memory organization remains unmodeled.Source: §4.2 · p. 16
5 A Primitive Cognitive Algebra
The paper defines a minimal cognitive algebra with formal assumptions and a small set of primitives for operating on representations. These primitives support abstraction, generalization, extension, and changes of representation.
5.1 Formal assumptions
- Types label compatibility for concept nodes and procedure interfaces; they are structural labels, not semantic domains.The assumptions also permit finite typed collections, function composition and identity, and predicates over well-typed arguments.Source: §5.1 · pp. 16–17
- Equality is an equivalence relation within each type, but procedures are not assumed to preserve equality.This qualification matters for representative selection.Source: §5.1 · p. 17
5.2 Core Cognitive Schemas and Algorithmic Support
- Abstraction combines invariance-based equivalence with representative selection, while generalization identifies shared distinctions and extension adds concepts or relations when existing distinctions are insufficient.Change of representation can restructure the representational space through new coordinates, redefined contexts, or alternative invariant representations.Source: §5.2 · p. 18
- The minimal, fixed primitives operate on Knowledge Space structures and support direct procedure use, bounded composition, and bounded generation followed by testing.They include graph search, representative selection, abstraction, fixed grouping, set operations and decisions, and Boolean operations.Source: §5.2 · pp. 18–19
6 Problem Solving with Simple Solvers on 2G
The section describes problem solving as representational operations followed by evaluation from a fixed, weak solver. Its examples show how invariance and representative selection can simplify the admissible solution forms without changing that solver.
6 Problem Solving with Simple Solvers on 2G
- A problem is identifying a valid, compact rule or transformation consistent with constraints; a fixed weak solver evaluates admissible representations, not long-horizon plans.Inputs are structured, transformed through cognitive primitives, then evaluated under explicit constraints; any complexity reduction is attributed to representation and context.Source: §6 · pp. 19–20
6.1 Example 1: Permutation invariance
- For duplicate detection, permutation invariance makes reordered sequences equivalent, and sorting is one representative choice that changes which solutions the context can express.A multiset encoding could serve the same representative-selection role; the solver itself is unchanged.Source: §6.1 · pp. 20–21
6.2 Example 2: Invariance under nuisance coordinates
- When only coordinates in S affect the output, inputs agreeing on S are equivalent, and the reduced representation admits solutions over |S| variables.The task specification does not identify S; in the refined context, irrelevant variation is structurally excluded, and the example is not a new reduct algorithm.Source: §6.2 · pp. 21–22
7 Discussion and Conclusion
The paper formalizes contexts as a foundation for representation, abstraction, and problem formulation, then places concepts and procedures in a Knowledge Space. Its structural operations are demonstrated with a fixed, weak solver rather than presented as a new algorithm.
7 Discussion and Conclusion
- Invariance recognition and representative selection suppress irrelevant distinctions and expose simpler task-relevant regularities.The paper treats them as procedures acting on structured conceptual content within its cognitive algebra.Source: §7 · p. 22
- The weak solver starts from an explicit K0, applies and composes procedures, generates bounded candidate extensions, and stores only successful procedures and traces.This provides an in-paper account of procedural reuse and representational extension with explicit success conditions.Source: §7 · p. 22
- Contexts are partitions of state-space subsets that provide a uniform way to express relevance, refinement, and task-specific structure.They are the foundational unit of distinction underlying representation, abstraction, and problem formulation.Source: §7 · p. 22
- The Knowledge Space couples a Concept Graph and a Procedure Graph, with the Concept Graph hosting concepts that are constructed, refined, and reused.It is a single representational scaffold, not a fixed ontology.Source: §7 · pp. 22–23
- Abstraction and generalization are not proposed as novel algorithms but organize simplification examples and supply primitives for a finite weak-solver demonstration.The focus is a testable representational mechanism, not a complete learning or control architecture.Source: §7 · p. 23
8 Additional Contextual Examples
The examples show how context choice determines admissible distinctions and how contexts can be organized through uniform coordinates or selective refinement.
8.1 Example 3: Context choice and representational form
- In Cpos, sequences differ by symbol identity at each position; Cmult groups sequences by symbol multiplicities and is a coarsening of Cpos.The two contexts express structure through position and ordering versus counts or set membership.Source: §8.1 · p. 24
- Context choice fixes which distinctions are admissible, independently of the underlying objects.This is the example’s central contrast between position-sensitive and multiplicity-based representations.Source: §8.1 · p. 24
8.2 Example 4: Local versus uniform context construction
- Uniform coordinates fix a representational interface from the start, whereas local refinement introduces distinctions selectively and preserves their refinement history.The two constructions can yield the same partition while organizing distinctions and procedures differently.Source: §8.2 · p. 24
A.1 Task code and feature coordinates
The task format gives the weak solver objects as segmented feature tuples, while leaving feature relevance unspecified.
A.1 Task code and feature coordinates
- The solver receives objects as finite tuples of feature values and does not receive images or learn segmentation.Segmentation is supplied as part of the input representation.Source: §A.1 · p. 26
- The solver can read and compare coordinate values but does not know their external semantic interpretation.Coordinates could correspond externally to attributes such as shape, colour, size, or position.Source: §A.1 · p. 26
- A task contains positive examples, negative examples, and candidates, but the solver is not told which feature coordinates are relevant.The task source may choose relevant coordinates; segmentation is given, relevance is not.Source: §A.1 · p. 26
A.2 Initial Knowledge Space
The weak-solver demonstration specifies how bounded procedures operate on a growing Knowledge Space, and when those procedures can or cannot solve a task. It shows a fixed solver using representation and stored procedures rather than hidden rules or an oracle.
A.2 Initial Knowledge Space
- K0 contains primitive operations, while the solver uses anonymous coordinate tests and has no hidden rule labels or oracle.Source: §A.2 · p. 27
A.3 Separating coordinate tests
- Acceptable coordinate subsets preserve agreement among positives and separate negatives; ambiguous candidate choice without resolving feedback leaves a task underdetermined.Source: §A.3 · p. 28
A.4 Weak-solver algorithm
- The algorithm tests direct application, bounded composition, then bounded generation; these stages respectively use existing items, recombine known items, or extend the Knowledge Space.Source: §A.4 · pp. 29–30
A.5 Knowledge Space update
- Successful compositions are stored as procedures, while successful generated items enter the Knowledge Space and repetition may stabilize patterns into concepts or reusable procedures.Source: §A.5 · p. 30
A.6 Curriculum
- The curriculum advances from direct equality and inequality through classification, composition, and bounded generation.Source: §A.6 · pp. 30–31
A.7 Worked trace
- The worked trace grows the Knowledge Space through successful solution traces, using primitives, stored procedures, and bounded generation without hidden rules or an oracle.Source: §A.7 · p. 33
A.8 Success and failure conditions
- The solver succeeds when a bounded admissible separator selects a unique candidate; stored procedures can also solve later matching tasks directly.Source: §A.8 · p. 33
- It fails when no admissible item lies within declared bounds or candidate choices remain indistinguishable without feedback.Source: §A.8 · p. 34