Formation of structural attractors in neuromorphic systems

1. Relevance and Objectives of the Study.

• ISL represents knowledge as stable invariant structures called structural attractors, formed through self-organization rather than as parameter distributions or trajectories in continuous space.The paper relates a structural attractor to the ontological sense of a concept.Source: §1 · p. 3
• The proposed theory is not presented as a direct account of brain mechanisms; its dendritic-tree implementation is a hypothesis and a biologically motivated mathematical abstraction.The authors ground the abstraction in observations concerning dendritic computation, stable-structure selection, and long-term representations.Source: §1 · pp. 3–4
• ISL replaces BackProp and gradient descent with monotonic structural reduction, forming attractor-concepts from extremely small datasets and exposing classifier decisions and errors through transparent topology.The proposal describes learning from positive examples in a single pass without preliminary or meta-learning or heuristic rules.Source: §1 · p. 3
• The study aims to formalize ISL, establish reduction convergence and class-attractor uniqueness, demonstrate pattern-recognition potential, and propose a hypothetical neurobiological interpretation.The computational work is described as a proof-of-concept on classical pattern-recognition tasks.Source: §1 · p. 4

2. Research Background.

• The work continues prior theoretical and experimental results [2] and draws on Hubel and Wiesel, Marr, Hopfield, and the tradition of structural pattern recognition.The passage describes that tradition as reaching significant theoretical and practical milestones before convolutional neural networks became widespread.Source: §2 · p. 5
• Xiao et al. (2008) treated image topology as a class invariant and built style-invariant hierarchical descriptions using region trees and graph-energy segmentation.Their pipeline extracts Difference of Gaussian regions, hierarchically merges neighboring regions, and uses graph energy to segment objects and parts.Source: §2 · p. 5
• Xiao et al. compress graph structure into an 8D vector for a GMM, losing relative positioning because distinct structures can share similar Laplacian spectra.With only 10 vectors, the authors warn that the Gaussian may be too broad or too narrow.Source: §2 · pp. 5–6
• The authors conducted an experiment on a 12-class image dataset in a real few-shot setting, using only 10 examples per class.Source: §2 · p. 6
• On Australian Sign Language (Auslan-s), the model showed an error rate of only 1.0%, surpassing the existing SOTA record of 1.5%.It also identified hidden local patterns without biological heuristics.Source: §2 · p. 7
• Geurts et al. [6] classify local pieces and aggregate predictions by voting, but their permutation-invariant bag of parts loses information about connections and scene geometry.The framework splits objects into pieces, trains a classifier on individual pieces, then votes or averages their predictions.Source: §2 · pp. 6–7
• ISL treats the structural attractor as both learning’s result and recognition criterion, shifting focus from constructing structural features to explaining their emergence.Prior approaches reduce structural descriptions to statistical estimates of feature distributions in Euclidean space.Source: §2 · pp. 7–8
• Dendritic Localized Learning relaxes weight symmetry and overlaps forward propagation with learning, but retains recursive error propagation and informational dependency between passes.Its local errors ultimately derive from output-layer error, so the authors characterize DLL as modifying error propagation rather than abandoning error-driven learning.Source: §2 · pp. 8–9
• Learning is viewed as the system’s sequential structural self-organization.Source: §2 · p. 9

1.1. Input Space 𝒳.

• The input space 𝒳 contains raw objects, including MNIST images, geometric shapes, contours, and 2D or 3D models.Source: §1 · p. 10, §1.1 · p. 10

1.2. Space of Atomic Structural Elements (Primitives) ℰ.

• Contour primitives include segments, endpoints, corners, and intersection or conjugation points, which serve as building blocks for input-pattern representations.Source: §1.2 · p. 10

1.3. Space of Hypergraphs (Representations of a Single Model Neuron).

• A model-neuron representation is a hypergraph of structural elements whose spatial edges encode connectivity and whose parametric hyperedges encode relationships among element parameters.Source: §1.3 · p. 10

2.1. Primitive Detector.

• The primitive detector maps an input x from 𝒳 to an initial hypergraph 𝒢0 in 𝒵0, while detector construction is outside this work’s scope.Source: §2.1 · p. 11

2.2. Measurements.

• Measurements associate each structural element with measured parameters, forming a connected multidimensional parametric space that structures the hypergraph representation.The coordinate systems and measurement scales are specified exogenously.Source: §2.2 · p. 11

3. Reduction

• The theory assumes each class has a hidden or noisy structural attractor that determines its stable invariant structure and manifests through reduction.Source: §3 · p. 12
• The reduction operator orders and compresses parameter values by quantizing them into stable segments, and the segmentation is idempotent: Q(Q(m)) = Q(m).Source: §3 · pp. 12–13

3.2. Structural Reduction 𝑅𝑠.

• Structural reduction selects anchors—elements present in all or almost all aligned class examples—and removes unstable elements to yield a minimal attractor structure.For incomplete training examples, the strict frequency condition can be relaxed to f(e) ≥ τ; the paper uses prepared examples with complete attractor structure.Source: §3.2 · pp. 14–15
• With the anchor-invariance criterion fixed for a class’s accumulated configuration, the anchor set A is uniquely determined.The specified criterion is the frequency measure.Source: §3.2 · pp. 15–16

3. Local Phase of Structural Reduction.

• Closure uniquely restores relations between anchors along original spatial paths and is idempotent, so repeated application leaves the reduced structure unchanged.Closure compresses existing spatial chains without introducing new spatial or parametric connections.Source: §3 · pp. 17–18
• The full reduction operator applies closure after local reduction, and its final attractor is minimal, containing no elements with f(e) < 1.The graph’s size in elements and parameters is strictly non-increasing.Source: §3 · pp. 18, 19, 21
• With finitely many initial elements, reduction terminates if each continuing step removes a non-invariant element or reaches an all-anchor state.The number of non-anchor elements strictly decreases at each continuing step.Source: §3 · p. 19
• The structural-canonical functional identifies significant element types from local topology, independently of accumulation order and the matching algorithm.Its frequency measure preserves element presence in class examples while removing dependence on the matching procedure.Source: §3 · pp. 20–21
• Under structural completeness and sufficient training-set diversity, reduction dynamics reconstruct the class’s ideal anchor structure from a finite sample.The stated condition yields A_C = A_N; f_C describes the class’s invariant structure, while f_N is used operationally.Source: §3 · p. 20

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

• On receiving a new positive example x′, the learning dynamics construct its example structure G′ = D(x′).Source: §4 · p. 22, §4.1 · p. 22

4.1. Anchor matching operator:

• The anchor-matching operator partially maps new-example anchors to existing anchors; mappings are consistent when sequential matching is order-independent.Source: §4.1 · p. 22
• The anchor-matching operator concerns contradictory identifications of the same anchors.Source: §4.1 · p. 22

4.2. Cumulative Union Operation.

• Cumulative union glues identified anchors, adds remaining hypergraph elements, and transfers edges after vertex union.The operation comprises vertex union followed by edge transfer.Source: §4.2 · pp. 22–23

4.3. Learning Iteration:

• Learning proceeds by iterating the operator 𝒢_{t+1} = R(𝒢_t⊕𝒢′) on the space of hypergraphs.This defines the learning iteration.Source: §4.3 · p. 24

4.4. Associativity and Commutativity of Learning (up to isomorphism):

• Different presentation orders for a finite set of class examples yield the same accumulated configuration up to isomorphism, although intermediate trajectories may differ.This result follows from the stated associativity and commutativity of learning up to isomorphism.Source: §4.4 · p. 24

4.5. Training Exclusively on Positive Examples.

• Class-detector neurons are trained independently and exclusively on positive examples to form an invariant hypergraph for one class.The authors hypothesize that this learning process models information processing and structural plasticity within a single pyramidal neuron’s dendritic tree.Source: §4.5 · p. 25

Axiom 3. Training of class-detector neurons is performed exclusively on positive

• Axiom 3 says the system evolves to a stable state by measuring its own structural components, rather than computing an answer using an external optimality criterion.The system’s dynamics are determined by a function of its measured state parameters.Source: Axiom 3. Training of class-detector… · p. 25

5. Principle of Invariant Reduction.

• ISL reduces structural redundancy while preserving invariants, converging to a class-determined stable core and fixed-point attractor without gradient-based optimization.The invariant sets may change as frequency estimates are refined, but converge to a stable core determined by the class.Source: §5 · pp. 26, 28, 29

6. Structural Attractor and Concept.

• A structural attractor is a canonical minimal hypergraph stable under the reduction operator that defines a class’s internal model.The paper distinguishes this structural attractor from an attractor defined as a set of states in classical dynamical systems.Source: §6 · pp. 29–30
• Every class structure reaches its structural attractor after a finite number of reduction steps, where the reduction stabilizes.The attractor is a fixed point of the reduction operator.Source: §6 · p. 31
• Learning reduces varied object representations to a single structural normal form, the convergent result of the learning process.The normal form is a canonical representation for objects of the class.Source: §6 · p. 31
• The model-based neurobiological interpretation maps attractors to stable activation patterns, anchors to stable synaptic configurations, and reduction to selection within dendritic trees.The paper explicitly presents this as an interpretation, not a direct assertion about biological implementation.Source: §6 · p. 32
• Self-organization is modeled rather than treated as a physical process, with primitive detectors, coordinate systems, and scales defined exogenously.Reduction operators model the energy dynamics of natural self-organization.Source: §6 · p. 32
• A structural attractor is a minimal invariant fixed point of reduction and defines a concept as the internal model of a class.The paper describes the concept as endogenous and as an internal model of a perceived holistic image.Source: §6 · pp. 31–32

7. Convergence Theorem for the Reduction Operator.

• Finite hypergraphs with a complexity that decreases outside fixed points stabilize in finitely many reductions at a fixed point.The stated proof condition is that each training example contains the class’s complete hidden attractor.Source: §7 · pp. 33–34

8. Theorem on the Uniqueness of the Class Structural Attractor.

• Given a finite accumulated class configuration, fixed invariance criterion and anchors, and deterministic closure, the full reduction operator has a unique fixed point.The theorem’s stated conditions also include local removal of non-anchor elements and fixed deterministic closure for restoring connectivity.Source: §8 · pp. 35–36
• The uniqueness result applies within a class under fixed formation conditions, not globally across all possible structures.The passage identifies class, anchor set, invariance criteria, and base spatial structure as contextual conditions.Source: §8 · p. 37

9. Parametric Spaces of Hypergraphs and Their Separability.

• Relative parametric space uses local coordinates at each structural point for reduction and anchor formation, whereas absolute space uses one fixed coordinate system for whole-structure characteristics.Source: §9 · p. 38
• The reduction operator is deterministic, independent of example-accumulation order, and preserves global parametric invariants.The invariants include properties such as contour traversal direction and stable element positions.Source: §9 · p. 39
• Global parametric invariants can separate classes that share identical local anchors and spatial interconnections.The passage illustrates this with digits 6 and 9, which differ in contour winding and endpoint orientation.Source: §9 · p. 39
• The full class signature Ψ𝒞=(A𝒞, Φ𝒞) combines local anchors and global parametric invariants to distinguish classes and their structural attractors.The paper identifies these signatures as the basis for class and attractor separability.Source: §9 · p. 40
• Constructing the full signature as an edit distance on hypergraphs is identified as a key challenge.The paper characterizes global parametric invariants as exogenous anchors that preserve the reduced hypergraph's spatial topology.Source: §9 · p. 40

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

• The attractor is structurally unique but generates a continuum of realizations; parameter-interval widths encode intra-class variability, and augmentation expands realizations without changing the structure.The model treats the attractor as a formal class prototype that combines structural invariance with endogenously formed variability.Source: §10 · pp. 41–42
• In ISL, intra-class variability arises in parametric intervals during cumulative reduction, reflecting observed parameter scatter without an external criterion.The paper contrasts this endogenous variability with metric-learning methods, where the metric is defined or learned exogenously; full endogeneity of coordinates and primitives remains future work.Source: §10 · p. 42

11. Few-Shot Learning.

• A sufficient covering set can reconstruct the class attractor when it covers its structural elements and has connected anchor overlap; the attractor is nested within each example.Such a set ensures reconstructibility but does not by itself guarantee a unique attractor.Source: §11 · pp. 43–44
• An identifying set is a minimal subset that uniquely reconstructs the attractor, whereas a sufficient covering set need not uniquely identify it.The paper states that any identifying set is sufficient, but the converse is generally false.Source: §11 · p. 45
• Few-shot sufficiency depends on covering capacity, not sample count: 2-3 examples may suffice while 100 may not, and finite attractor elements and connections permit a finite minimal set.That set must cover the attractor’s anchors and connections for full reconstruction.Source: §11 · pp. 45–46
• In few-shot learning, data augmentation is described as playing a fundamentally different role.The supplied passage does not specify the role it differs from.Source: §11 · p. 46

12. The Role of Augmentation in Learning.

• Augmentation is presented as an operator for exploring a parametric space.Source: §12 · p. 46
• The formal setup starts with an initial example represented as a hypergraph and states that augmentation generates a family.Source: §12 · p. 46

Parametric Coverage.

• Parametric coverage uses augmentation to segment parameters from observed values, assuming variations are structure-independent and one example represents class-wide variation.Source: Parametric Coverage · pp. 46–47

Decomposition of Learning.

• Augmentation builds the parametric structure or metric, while structural examples determine topology; augmentation therefore acts as a metrization operator.This contrasts with classical augmentation that increases sample size and reduces overfitting.Source: Decomposition of Learning · p. 48

Properties of the Metric:

• The metric arises through structure-to-segmentation-to-metric learning and is induced by segmentation, but its parametric space is specified exogenously.The metric is defined on anchors and is consistent through mapping μ.Source: Properties of the Metric · p. 49

13. Attention Mechanism and Anchors.

• Attention selects locally informative structural anchors, restricting matching to local alignment and weakening—but not eliminating—the difficulty of NP-hard graph matching.The canonical definition of the attention operator remains an open problem.Source: §13 · p. 50

14. Neuron Map.

• Distinct map attractors have distance greater than δ, making them structurally distinguishable; the class detector neuron hierarchically relates to subclass or example attractors.The class is represented by a finite set of attractors corresponding to subclasses or stable within-class variations.Source: §14 · pp. 51–52
• Under the stated reduction, learning-dynamics, pseudometric, and threshold conditions, a finite attractor map forms endogenously as stable fixed points.Learning trajectories stabilize at fixed points of the reduction operator.Source: §14 · pp. 52–53
• Each fixed point reflects invariants of an example subset, and different subsets may induce different attractors.Source: §14 · p. 54
• If d(Hx, A_i) ≤ δ, accumulation and reduction leave the structure unchanged; if d(Hx, A_i) > δ, reduction yields a distinct new fixed point A_j.Source: §14 · p. 54
• The attractor map self-organizes as fixed points of learning dynamics, determined by reduction R, accumulation ⊕, and distinguishability d rather than specified externally.Source: §14 · p. 55
• The class attractor A_C is a common substructure of every attractor A_i in the map.Source: §14 · p. 55
• When attractors are separated by more than δ, they represent different stable configurations and do not merge during reduction.Source: §14 · p. 55
• The formalism gives the structural condition for a new attractor, but novelty detection and detector creation lie outside it; Parzhin et al. (2020) describe a novelty neuron.Source: §14 · p. 56

15. Competition between Attractors.

• A mixed structural–parametric distance enables Winner-Take-All competition within class maps and between maps of different classes.The distance combines structural and parametric differences; the authors note that the proposed metric remains a pseudometric in some cases.Source: §15 · pp. 56, 59
• The authors propose dendritic coincidence topology, rather than spike frequency or count, as a basic neural code relevant to binding and invariant representations.They connect this proposal to the Binding Problem [11] and invariant conceptual representations, and identify topology of coincidences in the dendritic tree.Source: §15 · p. 56
• The proposed edit-distance mechanism is justified for hypergraphs but is simplified, differs from biological competition mechanisms, and leaves defining a WTA metric an open problem.The authors note that the metric is not the only possible mechanism for WTA competition.Source: §15 · p. 59

16. Discussion of Theoretical Results.

• The authors identify formation of a natural endogenous ontology—an internal model of the external world—as an advantage of the approach.They describe the theory as conceptual and requiring further research.Source: §16 · pp. 59, 62
• Primitive detectors are specified exogenously, and the necessary and sufficient primitives for forming all attractors remain an unresolved problem.The authors suggest sensory areas may be genetically predetermined rather than learned.Source: §16 · p. 59
• The experiments also specify coordinate systems and scales exogenously; straight-line contour approximation loses curve information and introduces errors and distortions.The authors identify variations in scales and coordinate systems as a broader issue beyond this work's scope.Source: §16 · pp. 59–61
• Anchor points are learned through reduction based on candidate-point frequency, although much of the theoretical discussion assumes the initial hypergraph and candidates already exist.Local parametric reduction isolates invariant substructures before global structural reduction; its mathematical formalization is left for future work.Source: §16 · p. 61
• Formalizing global topological invariants and the attention operator remains open, and the work uses only a heuristic for the NP-hardness problem.Proposed invariants include closedness, monotonicity under parameter changes, and contour traversal direction.Source: §16 · p. 62

16. Discussion of Theoretical Results

• Independently trained detector-neurons form concept hierarchies through their organization, producing hierarchies of attractors.The discussion presents this hierarchy as addressing overfitting and as separating attractor formation for simple images, complex scenes, movements, processes, and semantic constructs across levels.Source: §16 · pp. 62–63
• Structural feature-based models show scaling potential, and ISL may match or exceed it by using richer geometric elements in 3D hypergraphs.For 3D images, anchors could be segments or arcs, with planes or curved surfaces replacing the segments used for 2D structures.Source: §16 · p. 62
• The proposed neurobiological hypothesis treats a neuron's dendritic tree as the computational system forming an attractor, with the response reflecting a generalized characteristic.This is presented as consistent with studies of independent dendritic-branch computation and excitation driven by restricted synapse subsets.Source: §16 · p. 63
• Hypergraph matching is generally NP-hard; simple MNIST structures and active-perception modeling alleviate the challenge, but complex 3D stimuli may make it critical.The authors also consider it unlikely that brain neurons perform hypergraph matching or compute edit distances.Source: §16 · p. 63
• ISL is framed as an abstraction of neuronal-level information processes, not a low-level physical brain simulation, and Section IV offers a biological and neuromorphic perspective on these problems.The authors characterize it as addressing what neuronal processes do and why.Source: §16 · p. 63
• Associative recognition is theorized as attractor relaxation followed by activity spreading through the attractor to restore missing elements of an object's internal representation.The proposed mechanism may also interpret top-down perception modulation effects.Source: §16 · p. 63
• The experiments do not test associative restoration: incomplete or broken contours are excluded, so preprocessing losses are not restored and may reduce classification accuracy.Initially incomplete contours are also excluded from the training set.Source: §16 · p. 64
• ISL is not framed as statistical learning based on risk, probability estimation, or loss minimization; its representation architecture instead self-organizes as attractors form.The discussion notes that the architecture changes during learning rather than remaining fixed.Source: §16 · p. 64

17. Conclusions of the Section.

• ISL models learning as structural transformations converging to a hypergraph fixed point, with concepts represented by stable structural invariants rather than statistical parameters.The proposed process is endogenous to the observed data and does not use an external optimality criterion.Source: §17 · pp. 64–66
• The theory establishes finite reduction, a unique class attractor up to structural equivalence under anchor connectivity, and a self-organizing attractor map without a preset subclass count.The attractor is independent of example presentation order, and attractors may form a parametric family of permissible class representations.Source: §17 · p. 65
• The proposed model separates object topology from parameter scales, while augmentation constructs an internal class metric that supports learning from ultra-small initial samples.Its stated properties include no loss function or backpropagation, single-pass learning, positive examples only, and inherent interpretability through explicit attractors.Source: §17 · p. 65

1. Goals and Limitations.

• The MNIST experiment aims to test the theory at proof-of-concept level, assessing ultra-small-sample accuracy, local explainability, and the listed attractor properties without targeting state-of-the-art benchmarks.The proposed classifier is evaluated on a subset of MNIST with clean, unbroken contours.Source: §1 · p. 67
• The study does not examine raw-image primitive extraction in detail and restricts attractor formation to simple two-dimensional digit contours without background or complex scene substructures.Established preprocessing procedures are used; the authors state that this restriction makes more complex benchmarks difficult for primitive extraction.Source: §1 · p. 68
• The study excludes images with broken or fragmented contours because their classification requires a distinct associative recognition mechanism and mathematical approach.The paper also notes that this alternative has a different neurobiological motivation.Source: §1 · p. 68
• Training maps same-class objects into hypergraphs, combines them into a cumulative hypergraph, and reduces it to an attractor.Alignment maps elements of different hypergraphs into a shared structural space, where occurrence frequencies and stabilization under reduction can be analyzed.Source: §1 · p. 68
• Hypergraph alignment for matching is among the most computationally expensive training and inference procedures and is generally NP-hard.The procedure aligns elements across hypergraphs in a common structural space.Source: §1 · p. 68
• The theory uses parametric hyperedges, but the experiment simplifies these to thirteen continuous and categorical attributes per vertex in a directed attributed graph.The hyperedge layer is represented implicitly through a diagnostic-weighted node-substitution cost linking attributes of matching vertex pairs; this is described as a model simplification, not a theoretical departure.Source: §1 · p. 69

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

• Raster conversion binarizes the image, uses Growing Neural Gas [32] to skeletonize it, and applies Ramer-Douglas-Peucker simplification [33] to retain endpoints, corners, and junctions.Simplification removes intermediate vertices whose deviation from the chord between neighboring vertices is below a permissible value.Source: §2 · p. 69
• The result is a bipartite directed attributed graph with a 13-element node feature vector spanning spatial, orientational, structural-functional, and topological-neighborhood properties.Source: §2 · pp. 69–70

3. Reduction Operator and Attractor Formation.

• The digit-7 demonstration uses nine manually selected MNIST examples with intact, structurally complete contours; graph edit distance identifies shared anchors, while vertices occurring at frequency f(e) < 1 are removed.Source: §3 · pp. 72, 74
• After several iterations, attractor 7_1 stabilizes as a chain of three Point and two Vector nodes; later iterations may widen intervals but add no vertices, marking the reduction fixed point.The chain alternates between the starting point, horizontal vector, upper corner, diagonal vector, and endpoint.Source: §3 · pp. 74–75, Figure 3 · p. 74
• For preserved vertices, coordinate intervals across absorbed examples reflect expected digit topology, and interval width determines diagnostic weight, with stably repeated parameter values weighted more heavily.Source: §3 · p. 75
• The digit “7” has a compact topological-invariant attractor with complexity C = 9, but its broad segmentation causes excessive activity that competes with more complex concepts.The attractor has vector-type vertices connected by four edges.Source: §3 · p. 75
• Reduction updates parameter ranges, removes vertices absent from any absorbed example, then reconnects the remaining graph while preserving overall spatial topology.Permissible continuous-value intervals expand to observed extremes; removed vertices have f(e) < 1.Source: §3 · pp. 75–76
• Reduction stops when absorbing the next example leaves the attractor unchanged, indicating that the fixed point of operator R has been reached.The passage identifies this fixed point with the attractor described by Theorem 1.Source: §3 · p. 76

4. The Role of Augmentation: Parametric Stratification.

• Augmentation forms an internal metric from the input data through segmentation, rather than increasing the training sample size in the classical sense.Each original is expanded into structurally equivalent variants with independently varying parameters.Source: §4 · pp. 76–77
• Augmentation randomly rotates originals by [−10°, +10°] and shifts them by [−10%, +10%] of a 100 × 100 pixel window, producing 5–10 variants per original.The range boundaries and number of variants are chosen empirically; variant count depends on parametric variability.Source: §4 · p. 76
• Augmentation reinforces permissible coordinate bounds and narrows diagnostic-weight estimates; without it, 3–9 source instances yield nearly degenerate intervals and uniformly saturated weights.Parameter segmentation Q is defined by a stable interval rather than a single point.Source: §4 · pp. 76–77
• The augmented family empirically makes the resulting attractor invariant to the permutation of the training-example merge order.The set of cumulative-graph elements with f(e) = 1 does not depend on presentation order.Source: §4 · p. 77

5. Classifier Operation Principle.

• Inference computes an upper-bound estimate of graph-edit distance between a test image’s initial skeleton graph and each concept attractor, assigning higher substitution weights to diagnostically stable features.At a five-second budget per concept comparison, the procedure stops and returns the best result.Source: §5 · p. 78
• The classifier returns a local explanation with the winning concept, likelihood, and complexity values, enabling verification by comparing the winner with its closest competitor.The paper says this decision explanation requires no post-hoc model.Source: §5 · p. 79

6. Results of Experimental Studies.

• Attractor complexity cannot be determined solely from graph vertex and edge counts, because counts coincide for several concept pairs, including 1_1 and 5_1, so those counts alone cannot separate all attractor classes.Source: §6 · p. 79
• Accuracy was 91.13% on the valid MNIST set: 8,685 of 8,707 test images yielded valid skeleton graphs, while 22 were excluded as preprocessing errors.Figure 4 presents the confusion matrix for the 8,685 valid images.Source: §6 · pp. 79–80, Figure 4 · p. 81
• The dominant mismatch was “2” → “7” (111 cases, 51.6% of class “2” errors), alongside “4” → “7” (62), “1” → “7” (26), and “3” → “7” (10).The compact 7_1 attractor represents two segments and a corner inflection, matching several stylistic variants of these digits.Source: §6 · pp. 80–81, Table 4 · p. 82
• Preprocessing can prevent simpler examples from activating more complex attractors when skeletonization and local reduction remove their anchor points, revealing a precision limitation.The passages attribute some victories by attractor 7_1 to this mechanism.Source: §6 · p. 83
• The “9” → “4” error occurred in 46 cases (75.4% of class “9” errors); attractors 4_1 and 9_2 share structural complexity, while contour approximation loses distinguishing curved-line information.The two attractors have eight vertices, eight edges, and one loop.Source: §6 · p. 83

7. Discussion of Experimental Results and Prospects for Future Research.

• All 13 concepts stabilized in finitely many steps, with the 5-node attractor 7_1 stabilizing faster than the 11-node 8_1, consistent with the complexity-based estimate.The reduction operator reached a fixed point for each class.Source: §7 · pp. 83–84
• Changing the order of original examples left each resulting attractor’s vertices and attribute intervals unchanged, supporting that f(e) depends on the absorbed set, not its order.Presentation order varied for each concept.Source: §7 · p. 84
• Classes 1, 2, and 4 required multiple attractors for full coverage, with subclass divisions arising iteratively from structural incompatibility and corresponding to observed handwriting variants.The number of attractors per class was not predetermined.Source: §7 · p. 84
• Attractors expose inspectable structural features, and the model explains classification errors through its own structures rather than treating them as inexplicable anomalies.The paper contrasts this local interpretability with latent vectors and post-hoc explanations such as LIME and SHAP.Source: §7 · p. 84
• A covering set of 3 to 11 originals per class is sufficient for learning, defined as convergence to an attractor, without error-functional optimization; the method combines single-pass learning without backpropagation, ultra-small samples, and local model-level interpretability.Source: §7 · p. 84
• The MNIST comparison is contextual rather than controlled; the study frames its ultra-small-sample results as proof of concept rather than a state-of-the-art benchmark.The comparison is presented in Table 5.Source: §7 · pp. 84–85

ISL (MNIST)

• 91.13% accuracy on conditionally complete MNIST was reported for the model, versus 84.38% for CNN (RMNIST/5), 61% for MLP (MNIST), and 75% for SVM (RMNIST/5-10).The paper presents this result as confirmation of effectiveness within the considered theory, while noting that some compared results came from more complex datasets.Source: ISL (MNIST) · pp. 85–86
• Classification errors are attributed to imperfect hypergraph edit distance, empirical parameter choices, lost primitives, and impoverished parameter space; the error structure is interpreted within the model.The paper says the reduced parameter space motivates adding measured parameters or metrics.Source: ISL (MNIST) · p. 86
• Future work includes adding curve primitives, recognizing images with incomplete contours, scaling to more classes, and forming higher-level scene attractors.The proposed scene work includes complex handwritten symbols and aims to extend few-shot learning to many complex classes.Source: ISL (MNIST) · p. 86

IV. Neurobiological Hypotheses of Structural Projection and Neural Coding

• Experiments used simple contour images and constrained hypergraphs with candidate anchors, making matching polynomial-time solvable in that domain but not eliminating difficulty for realistic complex hypergraphs.The paper proposes hierarchical matching and parametric filtering as computational approaches; it also notes that learnable GNN attention would conflict with ISL’s ideology, though it could help investigate operator ℱ.Source: §IV · p. 87

1. Material Attractor.

• The material-attractor hypothesis treats a neuron’s dendritic tree as an evolving internal hypergraph shaped by prior experience, onto which stimuli are projected.The tree’s synapses and synaptic structures form functional relationships; the resulting structure is described as an internal model of the world.Source: §1 · pp. 89–90

2. Structural Projection.

• The section introduces a structural projection operator Π.Source: §2 · p. 90
• Stimuli activate structurally relevant internal elements through corresponding reduction dynamics, even when stimulus and internal hypergraphs have different topologies.The projection is not a vertex mapping or topological isomorphism; it is determined by the structure of neuronal interactions.Source: §2 · pp. 90–91

3. Dynamic Equivalence.

• Dynamic equivalence preserves the structure of convergence and relevant invariants, not topology; equivalent attractors need not be topologically identical.The relation applies when both attractors induce the same class of stable structural invariants.Source: §3 · pp. 91–92
• Stimulus hyperedges and dendritic connections may differ as structures yet implement the same constraints on reduction and admissible outcomes.Their anchor-element relationships are functionally equivalent.Source: §3 · p. 92
• Equivalent structures converge to a fixed point in a finite number of reduction steps, shifting the issue from structural isomorphism to dynamical-system equivalence.The passage states the finite-step relation as n∼m.Source: §3 · p. 92

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

• The model assumes initial redundancy in H0 means R can reveal structural attractors without adding elements, with excess degrees of freedom rather than hidden pre-existing attractors.This assumption makes the initial material hypergraph capable of forming the attractors in {𝒜𝒞}.Source: §4 · pp. 93–94
• During initial learning, reduction alone continues to reveal latent attractors until Ht reaches a fixed point, a mathematical idealization rather than biological cessation of plasticity.The proposed expansion operator is not required while Assumption 1 holds.Source: §4 · p. 94
• Adaptive learning uses stimulus-driven expansion followed by reduction when redundancy is exhausted and a new or modified stimulus is incompatible with the current attractor.The transition requires both an exhausted reduction process and a stimulus that reduction alone cannot accommodate.Source: §4 · pp. 94–95

1. Local Adaptation.

• The adaptation operator Λ is assumed to act locally: an element changes only through its participation in the active fragment ΔHt, not the neuron’s global state.The paper relates this assumption to local dendritic activity and synaptic tagging and capture as possible biological analogues.Source: §1 · p. 96

2. Global Invariants.

• The paper hypothesizes that backpropagating action potentials may instruct global attractor structure, which emerges cumulatively across class examples rather than being transmitted as pre-existing information.Because bAP efficacy decreases with distance from the soma, global structure may require additional hierarchical mechanisms to propagate plasticity among dendrites.Source: §2 · p. 97
• Spine elimination is not random, which supports Assumption 2.Source: §2 · p. 98

5. Subjectivity of Attractors.

• Different systems may have different attractors, yet their classifications based on subjective “models of the world” remain the same.Source: §5 · p. 98
• Structural-parametric adaptation modifies the existing dendritic-tree hypergraph through a stimulus projection, acting on the internal model rather than on the stimulus itself.Source: §5 · p. 98
• Attractors are stable configurations of internal models, and subjectivity follows from differences in those models and in attractor-formation histories, not noise.Source: §5 · pp. 98–99

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

• Repeated presentation of a stimulus class is hypothesized to materialize an attractor through a reduction cycle using two modes of the adaptation operator.The account treats perception as structural self-organization; differences in material histories, rather than uncertainty or noise, are proposed to underlie subjectivity.Source: §6 · p. 99
• The model proposes that hypergraph matching is not performed explicitly: its computational complexity shifts from the algorithmic level to the system's physical reduction dynamics.The stimulus-representation correspondence is described as emerging from reduction of the activated fragment of the material hypergraph.Source: §6 · p. 100

7. Hypothesis of Neural Coding.

• The neural code is hypothesized to reside in converging internal structural dynamics, with a neuron's observable response acting as their projection; ISL experiments found linear and more complex readouts inadequate.The material hypergraph is the substrate, not the code, and the proposed correspondence between stimulus and representation emerges through reduction dynamics rather than explicit hypergraph matching.Source: §7 · pp. 100–101
• A neuron is hypothesized to encode the decision reached, not the stimulus itself, and the same activity pattern may have different meanings across neurons and networks.The account locates a concept in a single neuron's structural attractor, with its response as a local projection interpreted by selected neighboring neurons.Source: §7 · p. 102
• The proposed observable neuronal response is multidimensional: action-potential generation may mark recognition, while response parameters may reflect attractor properties and convergence.The response may include a relatively stable, invariant component associated with the attractor and the fact of recognition.Source: §7 · p. 103
• Contextual EPSPs may contribute to the observable response without changing recognition, while excitation type may be reflected in response dynamics.Source: §7 · p. 103
• Neuronal responses are observable projections of internal dynamics formed by interaction between an attractor and current context, rather than direct concept encodings.Source: §7 · p. 103
• Informational meaning arises during interpretation by the receiving neuron, where dendritic structure and inter-neuronal connections may constrain future conceptual attractors.Source: §7 · p. 104

8. Hypothesis of the Projection Attractor.

• The proposed biologically plausible selection of inter-neuronal connections is related to structural projection operator Π, whose formalization remains incomplete.Source: §8 · p. 105

8.1. Formal Problem Formulation.

• The formal problem is to determine how outputs from primary primitive detectors connect to concept-detector dendrites without using a predefined or random map.Source: §8.1 · p. 105

8.2. Space of Potential Projections.

• Π0 is an overcomplete space of anatomically possible synaptic contacts treated as an initial condition, while its specific structure is beyond this work’s formalization.Source: §8.2 · p. 106

8.3. Hypothesis of the Existence of a Projection Attractor.

• The hypothesis proposes that Π emerges as an attractor of projection reduction through statistical coactivation and dendritic selection; a stable fixed point defines the representational space for conceptual attractors.Source: §8.3 · pp. 106–107
• During the critical period, neurons detecting frequently co-occurring natural-scene primitives may coactivate, and STDP may strengthen connections between them.The account concerns anatomically available candidates from M1 and treats natural-stimulus structural regularities as non-random.Source: §8.3 · p. 108
• Natural-stimulus structural regularities, together with the genetic program, may determine the topology of connections being formed.Source: §8.3 · p. 108
• The coactivation operator assigns higher values to neuron pairs activated together more often, formalizing a Hebbian principle primarily as an illustration.It is defined in terms of neuron activities.Source: §8.3 · p. 108

• The proposed dendritic selection operator DS(H0) treats initial morphology as a structural template: resonant coactivation patterns are stabilized, while others are pruned.Source: Dendritic Selection Operator · pp. 108–109
• DS(i,k,v) scores compatibility between input timing and segment characteristics: values near one indicate high compatibility, while values near zero imply low stabilization probability.Source: Dendritic Selection Operator · pp. 109–110
• The specific DS(H0) function is an ISL hypothesis, not a direct consequence of existing models, although data support dendritic selectivity, morphology-dependent integration, and local strengthening of coordinated inputs.Source: Dendritic Selection Operator · p. 110
• Different dendritic morphologies implement different selection operators, so identical input-stream statistics can yield different stable connection structures.Source: Dendritic Selection Operator · p. 110
• The hypothesized projection attractor reflects joint action of input statistics and morphological constraints, formalized as DS(H0)∘Co.Source: Dendritic Selection Operator · p. 111
• Under the projection-attractor hypothesis, identical stimuli may produce different representational spaces in different systems.Source: Dendritic Selection Operator · p. 112

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

• Cross-system coactivation of attractors may provide an internal teaching signal for forming connections and structural attractors without an external teacher or predefined class labels.Global biological reinforcement may modulate this process without determining the attractor’s content.Source: §9 · pp. 112–113

9.1. The Term "Counter Learning".

• Counter (cross-system) learning forms a new structural attractor in one functional system under the joint activity of local sensory representations and an already formed structural attractor in another functional system.Source: §9.1 · p. 113

9.2. Hypothesis of Counter Learning.

• An active attractor in a Base system recruits a free neuron in a New system; prolonged top-down and bottom-up coactivation induces structural plasticity and connection formation.Activity-independent contact formation alone does not select the correct neuron, and the specific molecular mechanism is not specified.Source: §9.2 · pp. 114, 116, 117

Phase 3. Formation of directed connections.

• After structural plasticity begins, potential contacts are stabilized when local coactivation exceeds its threshold, and structural reduction then yields a new structural attractor.The proposed contact set also requires an active presynaptic detector; the hypothesis treats this as a staged mechanism rather than specifying its molecular implementation.Source: Phase 3. Formation of directed… · pp. 117–118
• Cross-system coactivation selects the concept-neuron and initiates structural plasticity, while intra-system coactivation and local contact-growth mechanisms form and stabilize connection architecture.The account distinguishes selection and initiation from the local formation of specific connections.Source: Phase 3. Formation of directed… · p. 118

9.3. The Role of Top-Down Signals.

• The same top-down signal is proposed to perform three different functions at different stages of the attractor life cycle.The supplied passage does not specify the other two functions.Source: §9.3 · p. 118

1. Connection Formation Phase.

• During connection formation, the top-down signal recruits a free concept-neuron as the carrier of the future structural attractor and brings it into a structural state.The passage describes the signal's role in recruiting and preparing the concept-neuron.Source: §1 · p. 119

2. Learning Phase.

• Top-down signals trigger coactivation of recruited neurons and active pre-detectors to form a structural attractor after primary connections are formed.Source: §2 · p. 119

3. Functional Phase.

• After learning, top-down signals can activate a formed attractor when bottom-up input is absent or incomplete, initiating completion and association.Source: §3 · p. 119

9.4. Consequences of Hypothesis 4.

• The account assumes innate fixed-point attractors provide initial conditions for later concepts and proposes that attractor potential also depends on cross-system coactivation events.Source: §9.4 · p. 119

10. Conclusions of the Section.

• The proposed account locates concepts in stable internal dendritic structure and treats observable neuronal responses as projections of attractor states, not direct information carriers.These are neurobiological hypotheses, not established biological findings.Source: §10 · p. 120
• The hypotheses distinguish development of interneuronal architecture from attractor formation within it, and propose that an attractor in one system can initiate another through top-down projections.Source: §10 · p. 121
• The neurobiological contribution is a system of experimentally testable consequences of ISL, not proof that the proposed hypotheses are valid.Source: §10 · p. 122

V. Conclusion

• The work integrates a mathematical description of concept formation, a feasibility demonstration, and a possible neurobiological interpretation, with its formal framework representing objects as hypergraphs.Source: §V · p. 123
• The reduction process is a dynamical system converging to a stable structural attractor interpreted as a class concept, with its fundamental properties formalized by definitions and theorems.This formalization frames learning as the formation of stable structural representations rather than statistical optimization of model parameters.Source: §V · p. 123
• Despite substantially less training data and no error backpropagation, experiments demonstrate formation of stable structural representations from few positive examples, not superiority over modern architectures.The results support the feasibility of an alternative, non-statistical learning mechanism based on self-organization of structural and parametric invariants.Source: §V · p. 123
• The neurobiological interpretation is hypothetical: information resides in an internal attractor’s structure, observed responses are local projections, and interacting systems may form concepts without an external teacher.Projection attractors and cross-system formation are proposed as mechanisms linking learning to self-organization among functional systems.Source: §V · p. 123
• The work leaves local parametric reduction, anchor-point identification, quantitative dendritic dynamics, and innate-attractor formation and evolution for independent investigation.These are identified as directions for further development of the theory.Source: §V · p. 124
• The theory is not a definitive account of intelligence; validating its unified mathematical and biological framework for concept formation and information organization requires further theoretical and experimental work.If its principles prove valid, learning may be reconsidered as self-organization of structural attractors that constitute an internal model.Source: §V · p. 124