Pith. sign in

REVIEW 5 major objections 5 minor 70 references

The Urysohn Ladder: Recursive Metric Contraction for Scalable Continual Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read Continual learning can be solved by folding the manifold of experience, not by expanding memory.

desk verdict Four valid but circular conditional theorems, no empirical section despite the abstract's promise, and the load-bearing compressibility premise is conceded in the Limitations—a clear conceptual essay, not a research result. read the letter →

arxiv 2512.18471 v2 pith:UQPZBDI4 submitted 2025-12-20 cs.LG q-bio.NC

classification cs.LGq-bio.NC
keywords continuallearningcatastrophicinterferencequotientspacesmetriccontractionUrysohnlemmacoveringnumberstability-plasticitydilemmahierarchicalabstraction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Continual learning systems face a geometric wall: as experience accumulates, the trajectory on a flat representational manifold grows linearly, eventually exhausting any fixed capacity and causing catastrophic interference. This paper claims that the wall is an artifact of flat geometry, and that the solution is the opposite of the usual reflex to expand capacity. It proposes the Urysohn Ladder, a hierarchy of quotient maps that collapse validated submanifolds of experience into single tokens, shrinking the covering number at every level. If the stream is compressible, a fixed-capacity system can absorb an arbitrarily long stream by growing only logarithmically in hierarchy depth, and nonlinearity can be resolved by contraction rather than by the kernel trick. The paper proves bounded capacity, separability, stability via parity-partitioning, and correctness under abstraction, and it states itself that the key premise—the existence of a uniform compression factor—fails for incompressible streams.

What carries the argument

The load-bearing object is the Urysohn Ladder, a sequence of quotient maps q_k: M_k → M_{k+1} that collapse each validated submanifold into a single token; the identity carrying the argument is the covering-number inequality N(ε,M_{k+1}) ≤ ρ^{−1}N(ε,M_k), which telescopes to N(ε,M_D) ≤ ρ^{−D}N(ε,M_0). Urysohn's lemma provides the continuous separator that survives the quotient, and the parity partition M = H_odd ⊕ H_even ensures updates in the flow subspace do not disturb the scaffold subspace.

What would settle it

Construct or find a stream whose successive quotient levels have equal covering numbers—for example, a Kolmogorov-random walk in high dimension. Then N(ε,M_{k+1}) / N(ε,M_k) ≥ 1 for every quotient map, directly contradicting the ρ-compressibility premise and voiding Theorem 1's O(1) bound.

Watch

Extended reading notes

Core claim

The paper's central claim is that unbounded inference is achievable in fixed dimensions if and only if the system actively folds the manifold of experience. The mechanism is recursive metric contraction: validated temporal submanifolds are identified with points via quotient maps, producing a ladder M0 → M1 → ... → MD in which covering numbers drop by a factor ρ > 1 at each step. The consequences follow as four theorems: the Bounded Capacity Theorem (N(ε,MD) ≤ ρ^{−D}N(ε,M0), so capacity demand is O(1) per level); the Topological Collapse Separability Theorem (via Urysohn's lemma, collapsing disjoint sets to points makes them linearly separable in the same dimension); the Parity-Partitioned S

Load-bearing premise

The argument stands or falls on the assumption that the experience stream is recursively ρ-compressible: at every quotient level there exists a collapse that reduces the covering number by a constant factor ρ > 1, and the paper admits this fails for incompressible streams.

Editorial extensions

If this is right

  • A fixed-capacity network can, in principle, keep learning forever on a compressible stream by growing hierarchy depth logarithmically rather than adding parameters.
  • Nonlinearly entangled classes can be made linearly separable by metric contraction, so the kernel trick is mathematically unnecessary for separability.
  • Catastrophic interference can be eliminated by architectural segregation: plastic updates in one subspace, stable memories in an orthogonal subspace.
  • The cost of recalling a past event stops depending on how long ago it occurred; once collapsed, it is a short hop in quotient space.
  • The practical bottleneck shifts from memory capacity to discovering which submanifolds are safe to collapse.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the paper's core premise would be to instrument a continual-learning system to measure the covering-number ratio between successive abstraction levels on a real stream; if no level achieves a consistent ratio below 1, Theorem 1 has nothing to bind.
  • The paper implies an information-theoretic floor: on maximally incompressible streams, no fixed-capacity architecture can avoid linear capacity growth, so the approach's reach is exactly the compressible part of the task distribution.
  • The 'validated' qualifier is the real bottleneck; the paper does not specify how to certify that a submanifold is safe to collapse. One could test whether a simple validation signal—recurrence frequency, say—predicts whether contraction preserves downstream accuracy.
  • The parity-partitioning principle suggests that alternating search and consolidation phases are not just biologically inspired but geometrically necessary for interference-free learning; a test would be to compare continual learners with and without such phase separation at equal parameter count.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The manuscript proposes the 'Urysohn Ladder', a continual-learning framework in which a temporal manifold is recursively quotiented to collapse validated submanifolds into tokens. It claims four formal results: bounded capacity under recursive metric contraction (Theorem 1), separability via Urysohn quotienting (Theorem 2), stability via parity-partitioned orthogonal subspaces (Theorem 3), and correctness under compatible quotienting (Theorem 4). The abstract and introduction state that each claim is validated empirically with pretrained models and real-world datasets. My reading of the full text is that the four theorems are conditional statements whose hypotheses essentially contain the advertised conclusions, and no experiments appear anywhere in the manuscript.

Significance. If a constructive version of these results existed, the framework would offer a genuinely different geometric perspective on continual learning, potentially complementing expansion-based methods such as kernels, overparameterization, and replay. The definitions are clearly stated, and the telescoping argument in Theorem 1 is valid as a conditional derivation. However, the paper does not establish existence, computability, or learnability of the required quotient maps; the remaining theorems follow from their assumptions by construction; and the claimed empirical validation is absent. The significance of the results is therefore not demonstrated.

major comments (5)
  1. [§IV.A, Definition 4 and Theorem 1] Definition 4 defines recursive ρ-compressibility as Ceff(M_{k+1}) ≤ ρ^{-1} Ceff(M_k). Theorem 1 then assumes exactly N(ϵ,M_{k+1}) ≤ ρ^{-1} N(ϵ,M_k) and proves N(ϵ,M_D) ≤ ρ^{-D} N(ϵ,M_0) by telescoping. This is a valid conditional derivation, but the conclusion is already contained in the premise. The paper does not prove that any nontrivial stream admits such quotient maps; Section VII.B concedes that for incompressible streams no such maps exist and that discovering the deformation is an unsolved search problem. The central 'bounded capacity' claim therefore rests on an unproved environmental assumption.
  2. [§IV.B, Theorem 2] Theorem 2 defines the equivalence relation by x∼_f y iff f(x)=f(y), where f is a Urysohn separator. Consequently q(A) and q(B) are singletons by construction and the threshold rule is immediate. The theorem shows only that once an exact continuous separator is available, quotienting along its fibers trivially separates the classes. It gives no mechanism for learning f, no finite-sample analogue, and no guarantee that the required quotient can be constructed; Urysohn's lemma is non-constructive for arbitrary normal spaces. Thus the theorem does not support the paper's claim that recursive quotienting renders nonlinearly entangled structure linearly separable in a learnable system.
  3. [§IV.C, Theorem 3] The proof assumes Δθ_S=0 during Flow and Δθ_F=0 during Scaffold, together with a block-diagonal metric. Under these assumptions cross-interference vanishes by definition. The theorem does not establish that such an alternation can be realized by a learning algorithm, nor that the strict alternation is compatible with the quotient construction of Theorems 1 and 2. The Limitations section explicitly concedes that approximate orthogonality may introduce bounded interference and that quantifying leakage is future work. The stability guarantee is therefore assumed, not derived.
  4. [§V.A, Theorem 4] Theorem 4 assumes 'compatibility' x∼_k x' ⇒ f_k(x)=f_k(x') for every level. This assumption is exactly the condition that the descendant f_k is well-defined on the quotient; the induction then restates the universal property of quotient maps. In particular, the conclusion that A_{k+1} and B_{k+1} remain disjoint follows from compatibility with f_k's values 0 and 1, not from any property of metric contraction. Thus 'correctness under abstraction' is not proved from the contraction mechanism; it is an additional compatibility constraint imposed on the hierarchy.
  5. [Abstract and §I vs. full text] The abstract states 'We validate each claim empirically with pretrained models and real-world datasets,' and the Introduction repeats the promise of empirical validation. The full text, however, contains no experimental section, no datasets, no baselines, and no numerical results. This is a major discrepancy. Either the empirical claims must be supplied or the abstract and introduction must be revised to present the work as purely theoretical.
minor comments (5)
  1. [Title and §IV] The title promises the 'Urysohn Ladder', but the body primarily uses 'condensation hierarchy' and 'recursive metric contraction'; standardize terminology for readability.
  2. [§III.A and §IV.A] The quotient metrics d_k on M_k are never defined. Theorem 1 refers to N(ϵ,M_k) without specifying the metric at each level, making the covering-number comparisons not fully formal.
  3. [§IV.A, Lemma 1] Lemma 1 writes Ceff(M0)=Θ(L/ϵ). For an interval of length L the covering number is L/(2ϵ)+1 up to constants; this is acceptable but should be stated with explicit constants to avoid a false exact equality.
  4. [§VI.B] The claims that the 'magical number seven' is a geometric necessity and that cortical uniformity is a parity-alternating operator are not derived from the formal theorems. They should be labeled as speculative interpretations or conjectures.
  5. [References] There are duplicated and inconsistent entries (e.g., two Gromov references [13] and [27], and [28] is not the Gromov–Hausdorff distance paper). Please verify all citations.

Circularity Check

4 steps flagged · score 9.0 of 10

Bounded capacity, separability, and stability are each assumed in the definitions/axioms they are said to prove; the Limitations concede the required quotient maps and orthogonal phases are not constructed.

  1. self definitional [Section IV.A, Definition 4 and Theorem 1 (proof in Appendix)]
    "Definition 4 (Recursive ρ-Compressibility). A sequence of temporal manifolds {Mk}D k=0 is recursively ρ-compressible if there exists a sequence of quotient maps qk :M k → Mk+1 such that: Ceff (Mk+1)≤ρ −1Ceff (Mk) where ρ >1 is the uniform compression factor... Theorem 1: Assume that there exists ρ >1 such that for all k, N(ϵ,M k+1)≤ρ −1 N(ϵ,M k). Then N(ϵ,M D)≤ρ −DN(ϵ,M 0)."

    The bounded-capacity result is not derived from a construction; its key hypothesis is literally Definition 4 of recursive rho-compressibility. The appendix proof only telescopes that inequality. The paper's own Limitations concede the missing content: 'If the input stream is maximally entropic (incompressible), no quotient maps exist that significantly reduce the covering number' and 'Discovering the optimal topological deformation is a search problem that may be computationally expensive in the worst case.' Thus 'bounded capacity' is an assumption renamed as a theorem, not a derived guarantee.

  2. self definitional [Section IV.B, Theorem 2]
    "Define an equivalence relation x∼ f y⇐ ⇒f(x) =f(y) and let q:M → ˜M:=M/∼ f be the quotient map. Then: 1) The images q(A) and q(B) are distinct singleton points in ˜M(the equivalence classes at levels0and1)."

    The equivalence relation is defined along the fibers of the Urysohn separator f, so the separator descends to the quotient by the universal property and q(A), q(B) are singletons because A and B are level sets of f. The theorem therefore assumes the very separator it claims to deliver. It proves only that once a perfect f is known, quotienting by its fibers makes separation trivial; it supplies no mechanism for learning f. The Limitations admit this: the theorem 'does not prescribe an efficient algorithm for finding it.'

2 more flagged steps
  1. self definitional [Section IV.C, Theorem 3]
    "suppose the system alternates two update phases: (Flow phase): ∆θ S = 0,(Scaffold phase): ∆θ F = 0 . Assume the metric on parameter space is block-diagonal, g=g F ⊕g S, so that the induced inner product satisfies ⟨(uF ,0),(0, vS)⟩g = 0 . Then the cross-interference term between phases vanishes: ⟨∆θ(F) ,∆θ (S)⟩g = 0."

    The theorem's conclusion of zero cross-interference is exactly its hypothesis: updates never touch the other parameter block and the metric is block-diagonal, so the inner product of a flow-only update with a scaffold-only update is zero by definition. No argument is given that real learning can be partitioned this way. The Limitations state that 'parity separation is likely implemented only approximately... residual coupling between flow- and scaffold-associated updates may introduce bounded interference.' Stability is therefore a definitional consequence of the assumed phase separation, not a derived result.

  2. self definitional [Section V.A, Theorem 4]
    "If there exists a continuous f0 :X 0 →[0,1] separating A0, B0 such that, for each k, x∼ k x′ =⇒f k(x) = fk(x′), where f k is the descended separator on X k, then for every level k there is a continuous fk :X k →[0,1] separating A k and B k."

    Correctness under abstraction is presented as an invariant, but the theorem assumes compatibility of every quotient map with the descended separator. The proof defines f_{k+1}(q_k(x)) := f_k(x), so the conclusion is the definition of a well-defined descended separator; there is no demonstration that quotient maps arising from learning actually satisfy the compatibility condition. The advertised guarantee is again conditional on an assumption that the paper does not establish.

full rationale

The paper's central advertised guarantees reduce to their own hypotheses. Theorem 1's bounded-capacity conclusion is the telescoped version of Definition 4's compressibility inequality; Theorem 2 constructs the quotient from an Urysohn separator it assumes and calls the resulting trivial separation a theorem; Theorem 3 hard-codes zero cross-updates and a block-diagonal metric, making interference-free learning definitional; Theorem 4 assumes compatibility and proves that compatibility implies well-defined descent. The Limitations section confirms the gap: incompressible streams admit no useful quotient maps, finding the required topological deformation is an open search problem, and orthogonality is only an idealization. The abstract also claims empirical validation, but the full text contains no experiments, datasets, or results, so there is no independent check that the assumed contractions occur. Because the load-bearing properties are placed in the definitions and axioms rather than derived, the derivation chain is circular by construction.

Assumptions & free parameters 1 free parameters · 6 assumptions · 1 invented entities

The framework leans on a small set of postulates: flat manifold growth, existence of a constant compression factor ρ, a strict parity decomposition of the state space, and compatibility conditions for quotient maps. The first three are assumed rather than derived or measured, and the paper's own limitations section acknowledges the compressibility and search-for-topology gaps.

free parameters (1)
  • ρ (uniform compression factor)
    Assumed >1 such that N(ϵ,M_{k+1}) ≤ ρ^{-1}N(ϵ,M_k) at every level; the theorem's conclusion is a telescoping consequence of this assumption. No value or estimation procedure is given.
assumptions (6)
  • domain assumption Temporal manifold is flat and its geodesic diameter grows linearly with stream length L.
    Assumed in Lemma 1 and throughout; the paper models continual learning as a trajectory on a flat Riemannian manifold, but no evidence is given that real learned representations obey this.
  • ad hoc to paper The compressibility assumption: for some ρ>1, recursive quotient maps reduce the covering number by factor ρ^{-1} at every level.
    Def. 4 and Thm. 1; this assumption essentially contains the bounded-capacity conclusion. The paper admits in Limitations that incompressible streams break it.
  • standard math Urysohn's Lemma (normal space has separating function).
    Used in Thm. 2 and Thm. 4; standard result, not an issue.
  • ad hoc to paper Parity Alternation Principle: cognitive state space decomposes into orthogonal H_odd ⊕ H_even, with learning acting only on odd and memory on even.
    Axioms 1–2 in Sec. III-B; no independent empirical or mathematical justification; the stability theorem follows from this postulate plus the block-diagonal metric assumption.
  • ad hoc to paper Block-diagonal parameter-space metric and strict phase alternation (Δθ_S=0 in Flow, Δθ_F=0 in Scaffold).
    Assumed in Thm. 3; cross-interference vanishes by construction.
  • ad hoc to paper Compatibility of all quotient maps with the descended separator (x∼_k x' ⇒ f_k(x)=f_k(x')).
    Assumed in Thm. 4; without it, separation does not descend.
invented entities (1)
  • Metric singularity / wormhole (token as a region of extreme positive curvature)
    purpose: Explains how a condensed abstraction can bridge distant experiences and reduce inference cost.
    The paper proposes that tokens are physically realizable as wormholes/regions of extreme curvature, but provides no testable prediction or evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Urysohn Ladder: Recursive Metric Contraction for Scalable Continual Learning." pith.science (2026). https://pith.science/paper/UQPZBDI4

@misc{pith2026251218471,
  author       = {Pith},
  title        = {Pith review of: The Urysohn Ladder: Recursive Metric Contraction for Scalable Continual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UQPZBDI4}},
  note         = {Machine review of arXiv:2512.18471}
}
abstract

Continual learning systems face a fundamental geometric obstacle: as experience accumulates on a fixed-capacity manifold, covering numbers grow linearly with time, eventually forcing representational overlap and catastrophic interference. Prevailing approaches attack this problem by \emph{expansion} - projecting into higher-dimensional spaces via kernels, overparameterization, or replay. We argue the solution is the opposite: \emph{contraction}. We formalize abstraction as the \textbf{Urysohn Ladder}, a hierarchy of quotient maps that recursively collapse validated metric neighborhoods into compact tokens, converting unbounded ambient-space search into bounded navigation on a low-dimensional intrinsic scaffold. Geometrically, each collapsed token acts as a shortcut - a region of extreme metric contraction that bridges distant experiences, much like a wormhole in the representational manifold. We establish four results that collectively guarantee \emph{separability} (metric contraction renders nonlinearly entangled structure linearly separable at each quotient level, and this separability propagates faithfully through the entire hierarchy), \emph{bounded capacity} (covering numbers remain $O(1)$ per quotient level, independent of stream length), \emph{stability} (parity-partitioned flow/scaffold subspaces enable unbounded plasticity without catastrophic interference), and \emph{scalability} (inference cost scales with quotient distance, not ambient distance). We validate each claim empirically with pretrained models and real-world datasets. Moreover, we demonstrate the potential of Urysohn Ladder for scalable continual learning via scaffold amortization.

Figures

Figures reproduced from arXiv: 2512.18471 by the authors.

Figure 1
Figure 1. The Topological Trinity Transformation. The diagram demonstrates the core mechanism of MAI. Left: The raw input stream (Hodd) contains complex, intertwined temporal trajectories that violate Cover’s Theorem for linear separability. Right: After applying the Condensation Operator Ψ (Metric Collapse), the trajectories are topologically quotiented into single points in the Scaffold space (Heven). In this collapsed metr… view at source ↗
Figure 2
Figure 2. Geometric Expansion vs. Topological Contraction/Folding. (A) The Flat Manifold Problem: In standard continual learning, the temporal manifold [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Logarithmic Scaling via Recursive Condensation. The diagram illustrates how a linear input stream of length L on the base manifold M0 is progressively compressed through a hierarchy of quotient maps qk. Each contraction reduces the metric volume by a factor ρ. To achieve a bounded effective diameter at the top manifold MD (necessary for constant-width inference), the depth of the hierarchy D must scale logarithmical… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 4 linked inside Pith

  1. [1]

    Geometrical and statistical properties of systems of linear inequalities with applications in pattern recognition,

    T. M. Cover, “Geometrical and statistical properties of systems of linear inequalities with applications in pattern recognition,”IEEE Transactions on Electronic Computers, vol. EC-14, no. 3, pp. 326–334, 1965

  2. [2]

    V . N. Vapnik,Statistical Learning Theory. Wiley, 1998

  3. [3]

    Understanding machine learning: From theory to algorithms,

    S. Shalev-Shwartz and S. Ben-David, “Understanding machine learning: From theory to algorithms,”Foundations and Trends in Machine Learning, vol. 4, no. 2, pp. 1–325, 2014

  4. [4]

    Catastrophic interference in connection- ist networks: The sequential learning problem,

    M. McCloskey and N. J. Cohen, “Catastrophic interference in connection- ist networks: The sequential learning problem,”Psychology of Learning and Motivation, vol. 24, pp. 109–165, 1989

  5. [5]

    Catastrophic forgetting in connectionist networks,

    R. M. French, “Catastrophic forgetting in connectionist networks,”Trends in Cognitive Sciences, vol. 3, no. 4, pp. 128–135, 1999

  6. [6]

    Continual lifelong learning with neural networks: A review,

    G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, and S. Wermter, “Continual lifelong learning with neural networks: A review,”Neural Networks, vol. 113, pp. 54–71, 2019

  7. [7]

    Buzs ´aki,Rhythms of the Brain

    G. Buzs ´aki,Rhythms of the Brain. Oxford University Press, 2006

  8. [8]

    The columnar organization of the neocortex,

    V . B. Mountcastle, “The columnar organization of the neocortex,”Brain, vol. 120, no. 4, pp. 701–722, 1997

Show all 70 references
  1. [9]

    A hierarchy of temporal receptive windows in human cortex,

    U. Hasson, E. Yang, I. Vallines, D. J. Heeger, and N. Rubin, “A hierarchy of temporal receptive windows in human cortex,”Journal of Neuroscience, vol. 28, no. 10, pp. 2539–2550, 2008

  2. [10]

    The architecture of complexity,

    H. A. Simon, “The architecture of complexity,”Proceedings of the American Philosophical Society, vol. 106, no. 6, pp. 467–482, 1973

  3. [11]

    A global geometric framework for nonlinear dimensionality reduction,

    J. B. Tenenbaum, V . de Silva, and J. C. Langford, “A global geometric framework for nonlinear dimensionality reduction,”Science, vol. 290, no. 5500, pp. 2319–2323, 2000. 11

  4. [12]

    Testing the manifold hypothesis,

    C. Fefferman, S. Mitter, and H. Narayanan, “Testing the manifold hypothesis,”Journal of the American Mathematical Society, vol. 29, no. 4, pp. 983–1049, 2016

  5. [13]

    Gromov,Metric Structures for Riemannian and Non-Riemannian Spaces

    M. Gromov,Metric Structures for Riemannian and Non-Riemannian Spaces. Birkh ¨auser, 2007

  6. [14]

    F. R. K. Chung,Spectral Graph Theory. American Mathematical Society, 1997

  7. [15]

    Catastrophic forgetting, rehearsal and pseudorehearsal,

    A. Robins, “Catastrophic forgetting, rehearsal and pseudorehearsal,” Connection Science, vol. 7, no. 2, pp. 123–146, 1995

  8. [16]

    Prioritized experience replay,

    T. Schaul, J. Quan, I. Antonoglou, and D. Silver, “Prioritized experience replay,”arXiv preprint arXiv:1511.05952, 2015

  9. [17]

    Overcoming catas- trophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanu, N. Rabinowitzet al., “Overcoming catas- trophic forgetting in neural networks,”Proceedings of the National Academy of Sciences, vol. 114, no. 13, pp. 3521–3526, 2017

  10. [18]

    Feudal reinforcement learning,

    P. Dayan and G. E. Hinton, “Feudal reinforcement learning,”Advances in Neural Information Processing Systems, vol. 5, 1993

  11. [19]

    Hierarchical reinforcement learning with the maxq value function decomposition,

    T. G. Dietterich, “Hierarchical reinforcement learning with the maxq value function decomposition,”Journal of Artificial Intelligence Research, vol. 13, pp. 227–303, 2000

  12. [20]

    Benefits of depth in neural networks,

    M. Telgarsky, “Benefits of depth in neural networks,”Proceedings of the 29th Annual Conference on Learning Theory, 2016

  13. [21]

    Proto-value functions: A laplacian framework for learning representation and control in markov decision processes,

    S. Mahadevan, “Proto-value functions: A laplacian framework for learning representation and control in markov decision processes,” inInternational Conference on Machine Learning (ICML), 2005

  14. [22]

    A comprehensive survey of continual learning: Theory, method and application,

    L. Wang, X. Zhang, H. Su, and J. Zhu, “A comprehensive survey of continual learning: Theory, method and application,”IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 8, pp. 5362– 5383, 2024

  15. [23]

    J. M. Lee,Introduction to Riemannian Manifolds. Springer, 2018

  16. [24]

    Vershynin,High-Dimensional Probability: An Introduction with Applications in Data Science

    R. Vershynin,High-Dimensional Probability: An Introduction with Applications in Data Science. Cambridge University Press, 2018

  17. [25]

    A. N. Kolmogorov and V . M. Tikhomirov, ϵ-Entropy and ϵ-Capacity of Sets in Function Spaces, 1959, vol. 14, no. 2

  18. [26]

    J. R. Munkres,Topology, 2nd ed. Prentice Hall, 2000

  19. [27]

    Gromov,Metric Structures for Riemannian and Non-Riemannian Spaces

    M. Gromov,Metric Structures for Riemannian and Non-Riemannian Spaces. Birkh ¨auser, 1999

  20. [28]

    On the gromov–hausdorff distance,

    F. M´emoli, “On the gromov–hausdorff distance,”Discrete & Computa- tional Geometry, vol. 38, no. 1, pp. 43–70, 2007

  21. [29]

    Diffusion maps,

    R. R. Coifman and S. Lafon, “Diffusion maps,”Applied and Computa- tional Harmonic Analysis, vol. 21, no. 1, pp. 5–30, 2006

  22. [30]

    Burago, Y

    D. Burago, Y . Burago, and S. Ivanov,A Course in Metric Geometry. American Mathematical Society, 2001

  23. [31]

    The hippocampus as a cognitive graph

    R. U. Muller, M. Stead, and J. Pach, “The hippocampus as a cognitive graph.”The Journal of general physiology, vol. 107, no. 6, pp. 663–694, 1996

  24. [32]

    Memory, navigation and theta rhythm in the hippocampal-entorhinal system,

    G. Buzs ´aki and E. I. Moser, “Memory, navigation and theta rhythm in the hippocampal-entorhinal system,”Nature Neuroscience, vol. 16, no. 2, pp. 130–138, 2013

  25. [33]

    Simplified neuron model as a principal component analyzer,

    E. Oja, “Simplified neuron model as a principal component analyzer,” Journal of mathematical biology, vol. 15, no. 3, pp. 267–273, 1982

  26. [34]

    Optimal unsupervised learning in a single-layer linear feedforward neural network,

    T. D. Sanger, “Optimal unsupervised learning in a single-layer linear feedforward neural network,”Neural Networks, vol. 2, no. 6, pp. 459–473, 1989

  27. [35]

    Reactivation of hippocampal ensemble memories during sleep,

    M. A. Wilson and B. L. McNaughton, “Reactivation of hippocampal ensemble memories during sleep,”Science, vol. 265, no. 5172, pp. 676– 679, 1994

  28. [36]

    Hierarchical process memory: memory as an integral component of information processing,

    U. Hasson, J. Chen, and C. J. Honey, “Hierarchical process memory: memory as an integral component of information processing,”Trends in cognitive sciences, vol. 19, no. 6, pp. 304–313, 2015

  29. [37]

    A continual learning survey: Defying forgetting in classification tasks,

    M. De Lange, R. Aljundi, M. Masana, S. Parisot, X. Jia, A. Leonardis, G. Slabaugh, and T. Tuytelaars, “A continual learning survey: Defying forgetting in classification tasks,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 7, pp. 3366–3385, 2021

  30. [38]

    The homological brain: Parity principle and amortized inference,

    X. Li, “The homological brain: Parity principle and amortized inference,” arXiv preprint arXiv:2512.10976, 2025

  31. [39]

    Active inference: a process theory,

    K. Friston, T. FitzGerald, F. Rigoli, P. Schwartenbeck, and G. Pezzulo, “Active inference: a process theory,”Neural computation, vol. 29, no. 1, pp. 1–49, 2017

  32. [40]

    Memory consolidation,

    L. R. Squire, L. Genzel, J. T. Wixted, and R. G. Morris, “Memory consolidation,”Cold Spring Harbor perspectives in biology, vol. 7, no. 8, p. a021766, 2015

  33. [41]

    Relationships between nondeterministic and deterministic tape complexities,

    W. J. Savitch, “Relationships between nondeterministic and deterministic tape complexities,”Journal of Computer and System Sciences, vol. 4, no. 2, pp. 177–192, 1970

  34. [42]

    Shawe-Taylor and N

    J. Shawe-Taylor and N. Cristianini,Kernel methods for pattern analysis. Cambridge university press, 2004

  35. [43]

    Willard,General topology

    S. Willard,General topology. Courier Corporation, 2012

  36. [44]

    Bellman,Dynamic Programming

    R. Bellman,Dynamic Programming. Princeton University Press, 1957

  37. [45]

    The theta-gamma neural code,

    J. E. Lisman and O. Jensen, “The theta-gamma neural code,”Neuron, vol. 77, no. 6, pp. 1002–1016, 2013

  38. [46]

    The hippocampo-neocortical dialogue,

    G. Buzs ´aki, “The hippocampo-neocortical dialogue,”Cerebral cortex, vol. 6, no. 2, pp. 81–92, 1996

  39. [47]

    Theory of deep learning iii: Explaining the non-overfitting puzzle,

    T. Poggio, K. Kawaguchi, Q. Liaoet al., “Theory of deep learning iii: Explaining the non-overfitting puzzle,”arXiv preprint arXiv:1801.00173, 2017

  40. [48]

    Hatcher,Algebraic topology, 2005

    A. Hatcher,Algebraic topology, 2005

  41. [49]

    Kahneman,Thinking, fast and slow

    D. Kahneman,Thinking, fast and slow. macmillan, 2011

  42. [50]

    M. L. Minsky,Computation: Finite and Infinite Machines. Prentice-Hall, 1967

  43. [51]

    Nakahara,Geometry, topology and physics

    M. Nakahara,Geometry, topology and physics. CRC press, 2018

  44. [52]

    Amortized inference in probabilistic reasoning,

    S. Gershman and N. Goodman, “Amortized inference in probabilistic reasoning,” inProceedings of the annual meeting of the cognitive science society, vol. 36, no. 36, 2014

  45. [53]

    The magical number seven, plus or minus two: Some limits on our capacity for processing information

    G. A. Miller, “The magical number seven, plus or minus two: Some limits on our capacity for processing information.”Psychological review, vol. 63, no. 2, p. 81, 1956

  46. [54]

    Canonical microcircuits for predictive coding,

    A. M. Bastos, W. M. Usrey, R. A. Adams, G. R. Mangun, P. Fries, and K. J. Friston, “Canonical microcircuits for predictive coding,”Neuron, vol. 76, no. 4, pp. 695–711, 2012

  47. [55]

    The “wake-sleep

    G. E. Hinton, P. Dayan, B. J. Frey, and R. M. Neal, “The “wake-sleep” algorithm for unsupervised neural networks,”Science, vol. 268, no. 5214, pp. 1158–1161, 1995

  48. [56]

    R. S. Sutton, A. G. Bartoet al.,Introduction to reinforcement learning. MIT press Cambridge, 1998, vol. 135

  49. [57]

    Pearl,Causality

    J. Pearl,Causality. Cambridge university press, 2009

  50. [58]

    Hippocampal place-cell sequences depict future paths to remembered goals,

    B. E. Pfeiffer and D. J. Foster, “Hippocampal place-cell sequences depict future paths to remembered goals,”Nature, vol. 497, no. 7447, pp. 74–79, 2013

  51. [59]

    The mechanisms for pattern completion and pattern separation in the hippocampus,

    E. T. Rolls, “The mechanisms for pattern completion and pattern separation in the hippocampus,”Frontiers in Systems Neuroscience, vol. 7, p. 74, 2013

  52. [60]

    The importance of mixed selectivity in complex cognitive tasks,

    M. Rigottiet al., “The importance of mixed selectivity in complex cognitive tasks,”Nature, vol. 497, no. 7451, pp. 585–590, 2013

  53. [61]

    Linking connectivity, dynamics, and computations in low-rank recurrent neural networks,

    F. Mastrogiuseppe and S. Ostojic, “Linking connectivity, dynamics, and computations in low-rank recurrent neural networks,”Neuron, vol. 99, no. 3, pp. 609–623, 2018

  54. [62]

    G. F. Marcus,The Algebraic Mind. MIT Press, 2001

  55. [63]

    Building machines that learn and think like people,

    B. M. Lake, T. D. Ullman, J. B. Tenenbaum, and S. J. Gershman, “Building machines that learn and think like people,”Behavioral and Brain Sciences, vol. 40, p. e253, 2017

  56. [64]

    How does the brain solve visual object recognition?

    J. J. DiCarlo, D. Zoccolan, and N. C. Rust, “How does the brain solve visual object recognition?”Neuron, vol. 73, no. 3, pp. 415–434, 2012

  57. [65]

    Predictive reward signal of dopamine neurons,

    W. Schultz, “Predictive reward signal of dopamine neurons,”Journal of neurophysiology, 1998

  58. [66]

    Prefrontal phase locking to hippocampal theta oscillations,

    A. G. Siapas, E. V . Lubenov, and M. A. Wilson, “Prefrontal phase locking to hippocampal theta oscillations,”Neuron, vol. 46, no. 1, pp. 141–151, 2005

  59. [67]

    Sparks of artificial general intelligence: Early experiments with gpt-4,

    S. Bubeck, V . Chandrasekaran, R. Eldan, J. Gehrke, E. Horvitz, E. Kamar, P. Lee, Y . T. Lee, Y . Li, S. Lundberget al., “Sparks of artificial general intelligence: Early experiments with gpt-4,”arXiv preprint arXiv:2303.12712, 2023. APPENDIX Proof of Lemma 1. A 1-dimensional ...

  60. [68]

    Bounded Search Space Assumption:The lemma posits that the metric entropy of the action space is bounded by a constant C, independent of the global state dimension or 12 history length: |Sϵ|=N(ϵ,A k(t))≤C This assumption reflects theMetric Collapseproperty, where the topologica...

  61. [69]

    Discretization of the Inference Operator:The inference operator Φk approximates the optimal action a∗ by minimizing a local objective function L over the cover Sϵ, rather than the continuous spaceA k(t).ˆa= arg mins∈Sϵ L(s)

  62. [70]

    Complexity Analysis:The total computational cost per step, Tstep, is the sum of the evaluation costs for all candidates in the representative set Sϵ. Tstep = P s∈Sϵ Cost(evaluates) Given that the evaluation cost per candidate is Teval(ϵ), we have:T step =|S ϵ| ·Teval(ϵ) Substi...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.