Formation of structural attractors in neuromorphic systems

Yurii Parzhyn; Alexander Schwarzmann; Mykyta Lapin; Kostiantyn Bokhan

Source-linked AI summary

Formation of structural attractors in neuromorphic systems

Yurii Parzhyn, Alexander Schwarzmann, Mykyta Lapin, Kostiantyn Bokhan

arXiv:2609.06826v1Sep 6, 2026cs.AIcs.LGcs.NEq-bio.NC

TL;DR

The paper addresses structural problems in forming invariant class representations and proposes learning through structural attractors rather than optimization. Its formal and computational results report finite convergence and learning without optimization, while biological implementation remains only partly examined.

Problem
The theory identifies finding critical structural components and defining invariant parameters of the global structure as central problems.
Method
The approach represents class attractors as minimal invariant substructures within directed attributed graphs and forms them through structural reduction.
Results
All 13 concepts stabilized, with finite and varying iteration counts, and learning converged to attractors without optimization.
Takeaways & Limitations
Structural attractors can serve as minimal coverage sufficient to reconstruct the attractor, supporting a non-optimization learning mechanism.
Takeaways & Limitations
The extraction of primitives from raw images is not examined in detail for biological plausibility and is modeled using established methods.

Abstract

from arXiv · show

This paper examines the theory of Invariant Structural Learning (ISL), which proposes a non-optimization approach to concept formation. Learning is interpreted as convergence to structural attractors in a hypergraph space, rather than as the minimization of a global loss function. The paper presents the ISL model, including its mathematical formalization, computational verification, and a hypothetical neurobiological interpretation. The mathematical section introduces the formal apparatus of the structural reduction process and proves its finite convergence, the existence and uniqueness of class structural attractors, and the self-organization of attractor maps. The computational section demonstrates the feasibility of the proposed approach on classical image recognition tasks, utilizing the proposed learning mechanism without backpropagation and with extremely small training datasets. Finally, the neurobiological section formulates hypotheses regarding the possible implementation of structural attractors in dendritic trees, neural coding as a projection of internal attractor dynamics, and the development of neural architectures supporting the proposed learning concept. These hypotheses are discussed in the context of modern experimental data in the fields of dendritic computations, synaptic plasticity, and the structural organization of neural circuits. The proposed neurobiological mechanisms are presented as testable hypotheses rather than established biological facts. The results demonstrate the mathematical consistency and computational feasibility of the proposed model, while the neurobiological hypotheses outline potential directions for its experimental verification.

1 Augusta University, Augusta, USA

The paper proposes Invariant Structural Learning (ISL) as a biologically motivated alternative to optimization-based learning. It represents knowledge as stable invariant hypergraph structures formed through reduction and interprets these structures as concepts or attractors.

Motivation and proposal: ISL replaces error-driven optimization with structural self-organization in which stable invariant structures arise from hypergraph reduction.The approach is presented as avoiding BackProp, gradient descent, global error signals, and external objective functions.
Motivation and proposal: Structural attractors serve as endogenous concept representations formed through reduction rather than as externally specified learning targets.The paper identifies the attractor as both the result of learning and the criterion for recognition.
Neurobiological motivation: The proposed models are intended to reflect dendritic computation, local plasticity, co-located memory and computation, and distributed neural processing.The paper describes these biological correspondences as part of its neuromorphic design motivation.
Reported computational properties: The paper reports concept formation from extremely small datasets, without preliminary or meta-learning, using positive examples in a single pass.This claim is presented as a data-efficiency property of the proposed learning mechanism.
Reported computational properties: The proposed attractor topology is described as transparent enough to support mathematical interpretation of both correct classifications and errors.The paper frames this as a local explainability property of the model.
Scope and limitations: The neurobiological account is explicitly hypothetical and does not claim to directly describe biological brain mechanisms.The model is presented as a biologically motivated mathematical abstraction based on selected neurobiological observations.

1. Base Spaces and Structures.

ISL represents inputs as hypergraphs built from structural elements, spatial connectivity, and parameter relations. Exogenous primitive detection and measurement define the representation on which invariant selection and reduction operate.

Input and representation spaces: The input space includes objects such as MNIST images, geometric shapes, contours, and two- or three-dimensional models.These objects are mapped into a space of hypergraph representations.
Input and representation spaces: Each hypergraph contains structural elements, spatial edges, and parametric hyperedges describing interrelationships among measured parameters.The hypergraph forms a parametric superstructure over a base spatial graph.
Input and representation spaces: Primitive detection extracts structural elements, while measurement functions assign parameter values to those elements.The paper treats primitive-detector construction as outside the main goal of the work.
Segmentation and quantization: Exogenously specified coordinate systems and measurement scales partition parameter spaces into generalized segments used during learning.The paper identifies this exogeneity as a modeling cost.
Invariant selection: The selection operator retains elements occurring in all class examples, producing the anchor set used to construct the minimal attractor structure.Anchors may instead be selected with a frequency threshold when training examples incompletely express the hidden attractor.
Reduction properties: The anchor set is uniquely determined by the accumulated examples under the stated invariance criterion.The resulting reduced structure is therefore uniquely determined under the model’s conditions.
Reduction properties: The reduction process terminates after finitely many steps at a minimal non-increasing attractor containing no further removable elements.The finite termination argument uses a decreasing quantity that reaches zero.

4. Learning Dynamics (accumulative, order-invariant).

ISL learns by repeatedly merging a new example into the accumulated hypergraph and reducing the result. Under consistent anchor matching, the final class configuration is invariant to presentation order up to isomorphism.

Accumulative learning: A new positive example is incorporated by aligning anchors, taking a cumulative union, and transferring its edges into the accumulated hypergraph.Identified anchors are glued together while remaining elements are added.
Accumulative learning: The learning update is 𝒢_t+1 = R(𝒢_t⊕𝒢′), where union expands the structure and R performs reduction.Learning is formulated as iteration of an operator over hypergraphs.
Order invariance: Associativity and commutativity hold up to structural isomorphism when anchor mappings are consistent and the selection rule is fixed.The equivalence relation preserves segmented parameters and local minimal structures around anchors.
Order invariance: For a finite class whose examples share the same hidden hypergraph up to isomorphism, presentation order changes intermediate states but not the resulting configuration.Different orders lead to the same accumulated configuration 𝒢* up to isomorphism.
Class-specific learning: The method trains individual class-detector neurons on positive examples by forming an invariant hypergraph for one class rather than separating classes with statistical hypersurfaces.The paper connects this architecture to modeled processing in a pyramidal neuron’s dendritic tree.
Class-specific learning: The system evolves toward a stable state from measurements of its structural components without an external optimality criterion.This is described as an evolutionary process rather than direct answer computation.

5. Principle of Invariant Reduction.

Invariant reduction defines learning as an order-induced process that removes structural redundancy while preserving true invariants. Its fixed points are minimal structural attractors representing class concepts.

Reduction principle: The reduction operator decreases structural redundancy and preserves invariant elements until reaching a minimal fixed structure.The invariance functional selects elements with frequency f(e) = 1.
Reduction principle: ISL creates a non-differential, order-induced gradient-like dynamic instead of relying on an external loss function.The direction of reduction is induced by the partial order and segmentation hierarchy.
Convergence: Invariant sets may gain or lose elements during refinement, but their sequence converges to a class-determined stable core E*.The paper distinguishes this behavior from strict monotonicity of individual invariant sets.
Structural attractor: The structural attractor is the fixed point of reduction and a minimal carrier of the invariant structure.It is also characterized as the result of learning and a canonical representation for class objects.
Convergence: For any class representation, reduction reaches the class attractor in a finite number of steps and stabilizes there.The fixed-point property makes the attractor the endpoint of class-specific reduction.
Structural attractor: Learning reduces diverse object representations to one structural normal form, with the concept defined as the internal model of a class.The concept is treated as an endogenous ontological object in the proposed framework.
Neurobiological interpretation: The neurobiological interpretation maps attractors to stable activation patterns, anchors to stable synaptic configurations, and reduction to selection within dendritic trees.The authors explicitly qualify this mapping as model-based rather than a direct biological assertion.

7. Convergence Theorem for the Reduction Operator.

The reduction operator repeatedly simplifies a finite hypergraph while decreasing a discrete complexity functional, so the process reaches a fixed point in finitely many steps. Under fixed class conditions and deterministic closure, the resulting class structural attractor is unique within that class, while separability depends on full invariant signatures.

Convergence mechanism: For any finite initial hypergraph, the reduction sequence stabilizes in a finite number of steps at a fixed point of R.The proof uses discrete complexity values bounded below by zero and the impossibility of an infinite strictly decreasing sequence.
Convergence mechanism: The complexity functional C is non-increasing during reduction and strictly decreasing outside fixed points, making the process dissipative toward a minimal fixed structure.C measures structural and parametric complexity, and its decrease supports class-separation decisions during Winner-Take-All competition.
Class attractor uniqueness: Under fixed invariance criteria, anchor sets, and deterministic closure, the full reduction operator has a unique fixed point for a given accumulated class configuration.Local reduction stabilizes anchors, after which deterministic closure produces a unique structure.
Class attractor uniqueness: The uniqueness theorem is conditional within a class and does not assert global uniqueness across the entire space of possible structures.Different classes may assign different meanings to the same structural configuration, so biological systems require contextual attractor distinguishability.
Attractor separability: Local anchors and global parametric invariants form the full invariant signature Ψ_C=(A_C, Φ_C), which provides the basis for separating structural attractors.The paper states that distinct classes must be distinguishable by their full invariant signatures.

10. Attractor as a "Fuzzy" Prototype of a Class.

The class attractor is a unique structural object that permits a continuum of parametrically varying realizations rather than requiring exact example matches. Its variability is formed from observed class scatter, while reconstruction can require only a small identifying subset that covers the attractor's anchors and connections.

Fuzzy prototype: The attractor is unique as a structural object but generates a continuum of parametric realizations, with interval widths reflecting intra-class variability.Each class example realizes the attractor, and augmentation expands the realization set without changing the structural attractor.
Fuzzy prototype: The attractor functions as a formal prototype: it captures the class's structural invariant, while [A_C] represents permissible parametric proximity.This formulation allows approximate realizations rather than exact matches to a single template.
Endogenous variability: Parametric intervals are formed endogenously during cumulative reduction from the observed scatter of each anchor parameter, without an external criterion.The attractor jointly encodes topology and intra-class variability, distinguishing ISL from methods with exogenously defined or learned metrics.
Scope boundary: The model's fully endogenous variability remains limited because primitive detectors, coordinate systems, and scales are exogenous assumptions.Endogenizing coordinate systems and primitives is identified as future research.
Few-shot reconstruction: An identifying set uniquely reconstructs the attractor, whereas a sufficient covering set may cover all attractor elements without guaranteeing uniqueness.Every identifying set is sufficient, but the converse is generally false.
Few-shot reconstruction: Attractor reconstruction depends on coverage and connectivity rather than sample count, so 2-3 examples might suffice while 100 examples might be insufficient.A finite minimal set exists when the attractor has finitely many anchors and connections.

12. The Role of Augmentation in Learning.

Augmentation constructs the parametric structure around structural examples, while learning derives a metric through segmentation and invariant matching. The resulting dynamics form finite, self-organizing attractor maps under stated structural and distinguishability conditions.

Augmentation and metric formation: Augmentation builds the parametric structure or metric rather than merely increasing sample size to reduce overfitting.The paper contrasts this role with classical data augmentation.
Augmentation and metric formation: Structural examples determine topology, whereas augmentation determines parametric coverage and acts as a metrization operator.The attractor metric is induced from local metrics over structural elements.
Decomposition of Learning: The model derives the metric endogenously through the sequence structure → segmentation → metric.This contrasts with approaches that specify a metric externally.
Attention and matching: The attention operator restricts hypergraph matching to locally informative elements, weakening the difficulty of global graph-edit comparison without necessarily eliminating NP-hardness.Its canonical definition remains an open problem and future research direction.
Self-organization of attractor maps: Under monotonic reduction, invariant selection, connectivity closure, a pseudometric, and threshold δ> 0, learning forms a finite attractor map endogenously.The map consists of stable fixed points of the learning dynamics.
Self-organization of attractor maps: The class attractor is a common invariant substructure, while distinct attractors remain separated when d(𝒜_i, 𝒜_j) > δ.The map is represented as Fix(R), induced by learning dynamics rather than specified a priori.

15. Competition between Attractors.

ISL defines competition between structural attractors through a mixed structural–parametric distance, enabling winner-take-all selection while exposing unresolved metric and representation limitations.

Neurobiological interpretation: The proposed neural interpretation treats dendritic coincidence topology, rather than spike frequency or count, as a basic element of neural coding.This connects structural attractors with invariant conceptual representations and the Binding Problem.
Distance construction: The proposed attractor distance combines structural and parametric differences to compare hypergraph-based representations.The mixed pseudometric is weighted by coefficients α and β.
Distance construction: Structural distance measures how many hypergraph elements cannot be matched between two attractors.It is defined as |E1|+|E2|−2·max match.
Competition mechanism: The formal attractor metric enables winner-take-all competition between maps, including maps from different classes.The paper presents this edit-distance mechanism as one realization of WTA competition for hypergraphs.
Limitations: The framework remains limited by simplified contour preprocessing, incomplete mathematical formalization, heuristic hypergraph matching, and unresolved biological correspondence.The paper notes information loss from straight-line approximation, excluded broken contours, an open WTA metric problem, and the implausibility that neurons literally compute hypergraph edit distances.

17. Conclusions of the Section.

The section presents ISL as a non-optimization learning framework in which concepts emerge as structural attractors through invariant extraction and structural self-organization. Its formal results establish finite convergence, attractor uniqueness, order invariance, and self-organized concept maps, while experiments demonstrate proof-of-concept feasibility under few-shot, single-pass conditions.

Theoretical results: The structural reduction process terminates in finitely many steps bounded by the initial structural complexity.This is stated as Theorem 1 and supports finite convergence to an attractor.
Theoretical results: Under anchor connectivity, each class attractor exists, is unique up to structural equivalence, and is independent of training-example order.The conclusions identify these properties with Theorem 2 and its corollary on conditional uniqueness.
Theoretical results: The attractor set self-organizes into distinguishable fixed points without a pre-specified number of subclasses.The paper describes this self-organized detector-neuron map as an alphabet of concepts.
Model properties: A concept is represented as a stable invariant structure, and augmentation constructs the class's internal parametric metric rather than merely increasing sample count.The sequence “structure → segmentation → metric” defines the geometry of permissible concept realizations.
Model properties: ISL omits loss functions, backpropagation, and negative examples while supporting single-pass learning from ultra-small positive-only training sets.The claimed properties include inherent interpretability through explicit structural attractors.
Theoretical framework: ISL formalizes learning as convergence of structural transformations to fixed-point concept representations called structural attractors.The model treats information as a stable structure emerging during reduction rather than as a statistical characteristic of observations.
Experimental verification: The MNIST proof-of-concept uses clean, unbroken contours to test few-shot classification and local explainability rather than state-of-the-art benchmark performance.The authors explicitly restrict the experiment to empirical viability of the proposed theory.

2. From Raster to Graph: Implementation of the Primitive Space.

The implementation converts raster contours into bipartite directed attributed graphs, represents nodes with typed structural and parametric features, and applies reduction to obtain concept attractors. The demonstrated pipeline uses skeletonization, polygon simplification, graph matching, and parametric aggregation, with augmentation reinforcing invariant intervals and order-independent attractors.

Graph construction: Raster images are converted into bipartite directed attributed graphs whose Point nodes represent critical skeleton points and Vector nodes represent directed segments.The pipeline includes binarization, skeletonization, and polygon simplification before graph construction.
Graph construction: Each graph vertex carries a thirteen-element feature vector spanning spatio-geometric, orientational, structural-functional, and topological-neighborhood categories.Features exclusive to one node type are skipped when matching against a different node type.
Structural reduction: The reduction process uses graph edit distance to isolate invariant cores containing critical anchor elements across recognition examples.Anchor search implements the attention operator and also addresses the graph-matching problem.
Structural reduction: Stable repetition of parameter values narrows their intervals and increases their diagnostic weights during attractor formation.Subsequent iterations may expand intervals but do not add vertices, indicating that the fixed point has been reached.
Parametric stratification: Augmentation constructs an internal metric by reinforcing permissible coordinate intervals and narrowing diagnostic-weight estimates rather than simply increasing sample size.The experiment reports uniform saturation of diagnostic weight without segmentation.
Parametric stratification: The resulting attractor is invariant to the permutation of training-example merge order.The reported reason is that the set of elements with frequency f(e) = 1 on the cumulative graph does not depend on presentation order.

5. Classifier Operation Principle.

Classification reduces each test-image graph toward each concept attractor, using weighted structural edits and bounded iterative comparison. Experiments report finite, order-independent learning, interpretable decisions, and strong performance from very small training sets, while exposing preprocessing and attractor-complexity limitations.

Classifier operation: Classification computes an upper bound on graph-edit distance between a test-image skeleton and each concept attractor, then selects the lower-cost concept.Vertex-substitution costs reflect attribute deviations, with greater weight for diagnostically stable features.
Classifier operation: The iterative comparison uses a polynomially initialized upper-bound generator and stops after a fixed 5-second budget per concept comparison.The procedure returns the best estimate found within that budget.
Computational behavior: More than 99% of comparisons stabilized near the optimum for graphs containing 3 to 12 reference nodes.This supports feasibility of the bounded comparison procedure across the tested graph sizes.
Interpretability: Each prediction returns a local structural explanation based on the winning attractor, its closest competitor, and the attributes determining their ranking.Verification therefore uses the classifier’s own structural quantities rather than a post-hoc model.

unique samples

The section reports computational feasibility of ISL on image-recognition tasks and identifies model-internal sources of classification error. It also outlines extensions toward richer primitives, incomplete contours, more classes, and scene attractors.

91.13% accuracy was obtained on datasets more complex than MNIST, including miniImageNet and FMNIST.
ISL achieved 61% on MLP (MNIST) and 75% on SVM (RMNIST/5-10), supporting the theory’s computational effectiveness.
Discussion of Problematic Issues: Classification errors arise from imperfect hypergraph edit-distance calculations, empirical parameter choices, primitive loss during preprocessing, and reduced parametric space.The edit distance can be non-monotonic relative to attractor specificity, and additional measured parameters or metrics may be needed.
Discussion of Problematic Issues: The authors argue that identifying these issues allows the error structure of attractor-based classification to be interpreted within the model itself.
Prospects for Future Experimental Research: Future work includes adding curve primitives, recognizing incomplete contours, scaling to national alphabets, and forming higher-level scene attractors.Scene attractors are proposed for complex handwritten-symbol datasets and future few-shot-learning experiments with many complex classes.
The experiments deliberately used simple contour images and reduced structural-element and parameter types, yielding relatively simple hypergraphs.Attention focused on informative structural elements, enabling constrained low-treewidth matching that is polynomial-time solvable in the selected input domain.
This restriction mitigates, but does not eliminate, the NP-hardness of matching more complex hypergraphs representing realistic images and scenes.

4. Adaptation Operator and Initial Redundancy of the Material Hypergraph.

The adaptation operator combines reduction with structural-parametric expansion, operating first on a redundant initial hypergraph and later adding structure when existing redundancy is exhausted. The proposed biological interpretation emphasizes local adaptation and global confirmation signals.

Initial Redundancy: The initial redundant hypergraph can reveal structural attractors through reduction without adding new elements.This initial-learning mode remains active while reduction continues and the material structure has not reached a fixed point.
Adaptive Learning: The model treats a fixed point as a computational or mathematical idealization rather than biological cessation of plasticity.
Adaptive Learning: Adaptive learning introduces the expansion operator when redundancy is exhausted and new or modified stimuli require structural reorganization.The transition requires both exhaustion of reduction at a fixed point and incompatibility between the new stimulus and the current attractor.
Adaptive Learning: A compatible stimulus produces no adaptation when projection remains within the existing attractor, yielding another reduction cycle.
Local Adaptation and Global Invariants: Assumption 2 makes adaptation local: only elements participating in the active fragment ΔH_t are modified rather than the entire neuronal structure.The paper relates this assumption to local dendritic plasticity and synaptic tagging and capture.
Local Adaptation and Global Invariants: The proposed global-invariant role may be served by backpropagating action potentials, which provide a soma-to-dendrite signal that helps establish attractor structure.The global invariant is described as emerging cumulatively across examples rather than being transmitted as pre-existing information.
Local Adaptation and Global Invariants: Because bAP efficacy decreases with distance from the soma, forming global attractor structure may require additional hierarchical mechanisms across dendritic compartments.

6. Hypothesis of Dynamic Equivalence of Attractors under Structural Projection.

The paper proposes that structural correspondence can emerge from projection and reduction dynamics rather than explicit hypergraph matching. It further hypothesizes that attractors, dendritic morphology, and projection dynamics shape subjective neural representations and coding.

Repeated presentation of class-C stimuli is proposed to materialize an attractor in the material hypergraph through a two-mode reduction cycle.
The paper concludes that ISL provides a mathematical model of structural self-organization in hypergraph space and hypothesizes dynamic equivalence under structural projection.
Structural projection reframes correspondence as an emergent result of reduction dynamics in an activated material fragment rather than explicit hypergraph matching.
Neural Coding: The proposed neural code is an attractor dynamic whose observable response is a projection, rather than a code defined solely by firing rate, timing, or population activity.The dendritic structure is treated as the substrate of the dynamics, not itself the code.
Neural Coding: MNIST experiments indicate that linear and more complex readouts did not adequately map attractor dynamics into neuronal output signals.
Neural Coding: The paper hypothesizes that a neuron encodes an reached decision or internal justification rather than the stimulus itself.
Projection Subjectivity: Projection attractors define representational spaces in which conceptual attractors subsequently emerge, allowing identical stimulus streams to yield different stable representations.
Projection Subjectivity: Dendritic morphology may implement different selection operators, so systems exposed to identical input statistics can form different stable connection structures.

9. Hypothesis of Counter (Cross-System) Learning.

Hypothesis 4 proposes cross-system coactivation as a mechanism for selecting neurons, forming inter-neuronal connections, and creating structural attractors without classical supervision. The mechanism uses top-down recruitment and joint top-down/bottom-up activity, followed by local structural plasticity and reduction.

Core proposal: Cross-system coactivation can trigger inter-neuronal connection formation and attractor formation without classical supervision.The proposed mechanism is presented as addressing biological-plausibility concerns associated with supervised learning.
Motivation: Hypothesis 4 addresses coactivation before connection formation, interconnection formation, and structural-attractor formation across functional systems.It is intended to explain the origin of projection attractors and replace predefined class labels with internal teaching signals.
Core proposal: Counter learning forms a new structural attractor in one functional system under joint activity from local sensory representations and an attractor in another system.The Base system may be sensory, motor, attentional, interoceptive, proprioceptive, emotional, or another system, while the New system receives the induced structural organization.
Learning phases: Top-down activity first recruits a free concept-neuron, then jointly activates it with bottom-up inputs to initiate structural plasticity.The recruited neuron receives independent top-down and bottom-up excitation; prolonged coactivation crosses a threshold for structural plasticity.
Learning phases: Local spatiotemporal coactivation stabilizes eligible contacts, and structural reduction produces the new structural attractor.Cross-system coactivation selects the concept-neuron and initiates plasticity, whereas intra-system coactivation determines the specific connection architecture.

1. Connection Formation Phase.

The connection-formation phase uses an activated attractor in the Base system to recruit a free concept-neuron in the New system. Top-down signaling determines the future attractor carrier and initiates structural plasticity.

Connection Formation Phase: An activated Base-system attractor sends a top-down signal that recruits a free concept-neuron in the New system.The signal determines which neuron will carry the future structural attractor.

2. Learning Phase.

The learning phase interprets neural responses as projections of internal structural-attractor dynamics and extends ISL toward cross-system concept formation. The paper presents these neurobiological mechanisms as hypotheses and identifies substantial questions for future testing.

Learning Phase: Top-down signals can act as learning triggers during attractor formation and later support associative activation when bottom-up input is absent or incomplete.The model assigns top-down activity multiple roles across learning and post-learning activation.
Learning Phase: Innate attractors are proposed as fixed points that provide initial conditions for the ontogenetic formation of later conceptual attractors.Their proposed origins lie on the phylogenetic timescale and may involve selection and genetic assimilation of adaptive predispositions.
Learning Phase: The model treats observable neuronal responses as multidimensional projections of internal attractor dynamics rather than direct carriers of information.Information is proposed to arise through interactions between incoming signals and the receiving neuron's internal organization.
Learning Phase: Hypothesis 4 proposes that cross-system top-down projections can form new attractors and concepts without an external teacher.The hypothesis links self-organization across interacting functional systems to a possible mechanism for concept emergence.
Learning Phase: The formal ISL theory describes reduction as converging to stable structural attractors interpreted as class concepts, while experiments show stable representations from limited positive examples without error backpropagation.The experiments demonstrate an alternative non-statistical learning mechanism, not superiority over modern machine-learning architectures.
Learning Phase: The neurobiological proposals remain hypothetical and exclude detailed models of local reduction, anchor-point identification, dendritic dynamics, and innate-attractor evolution.The authors state that validating the framework requires further theoretical and experimental investigation.