acceptodds
Withdrawn from ICLR 2027

Indeterminate Cognitive Automata: Provably Fast Topological Memory Retrieval and Correlation Bounds for Autonomous Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing deep learning architectures face severe computational constraints due to the O(L^2) quadratic bottleneck of transformer self-attention and von Neumann memory-bus latency. In this work, we present Indeterminate Cognitive Automata (ICA), a non-symbolic framework for autonomous self-learning via internal topological field relaxation. ICA projects continuous sensory streams directly onto high-dimensional memory manifolds using structured two-dimensional arrays termed Perceptlets and Conceptlets. By replacing exhaustive global vector scans with localized graph traversals across an Exalted Data Stage (EDS), the architecture establishes a provably fast two-stage retrieval pipeline: first, achieving constant-time O(1) memory search and context pruning relative to repository size N; second, guaranteeing tight error bounds epsilon for dynamic correlation indexing over candidate spaces. We provide rigorous mathematical proofs for these search and correlation complexity limits, supported by an asymptotic spatial bound of O(ND + M). Furthermore, we specify a Processing-in-Memory (PIM) hardware substrate equipped with SRAM-based Topological Edge Routing Engines (TERE). Theoretical analysis confirms that ICA eliminates primary algorithmic and hardware bottlenecks, offering a scalable foundation for lifelong, non-von Neumann autonomous cognitive architectures.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Non-von Neumann Cognitive Architectures, Autonomous Reasoning & Self-Learning, Constant-Time Retrieval ($O(1)$ Complexity), Dynamic Correlation Indexing

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_jjxrOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This submission proposes Indeterminate Cognitive Automata, a graph-based memory-retrieval architecture for lifelong learning. Sensory representations are mapped by an unspecified projection to bounded-degree graph nodes; retrieval follows a greedy, potential-decreasing walk for at most HH hops and is paired with a sketched SRAM processing-in-memory substrate. The paper claims exact globally optimal retrieval in repository-size-independent O(1)O(1) time, exponentially decaying error, and O(ND+M)O(ND+M) storage. However, fixed-hop greedy search cannot generally guarantee a global optimum, the error bound assumes rather than derives contraction, and no graph-construction procedure, implementation, simulation, dataset, numerical measurements, or supplementary code supports the claims.

Critical strengths
  • Reducing the latency and energy of large, dynamically updated memory retrieval is an important systems-and-ML problem, and considering algorithm/hardware co-design could be consequential if instantiated and validated.

  • Algorithm 1 makes the intended fixed-budget greedy graph walk recognizable, and the authors transparently disclose that Gemini assisted with language, notation, and proof derivations. These positives make the claims auditable, but they do not offset the fundamental validity and evidence problems.

Critical weaknesses
  • The O(1)O(1) result assumes a constant HH, cannot guarantee the claimed global optimum even on a path graph, omits entry-point and dimensional costs, and uses undefined graph gradients/convexity. The error-bound appendix assumes rather than derives exponential contraction. Core representations, scores, learning, graph construction, and online updates are also undefined. Therefore neither the algorithm nor its "provable" guarantees are valid in their present form.

  • The "Results" section has no results; the complexity table is not a simulation, and there is no code or supplement. The PIM proposal is one paragraph without a circuit/microarchitecture, synthesis or simulation, precision/dataflow description, or latency/energy/area accounting. The comparison to Transformers and ANN systems is not task-matched, relevant dynamic-ANN literature is missing, and multiple references appear untraceable or bibliographically conflated. The disclosed use of Gemini to refine proof derivations is not itself a problem, but the combination of assertion-heavy prose, invalid proofs, and malformed citations contradicts the claim of complete human verification and substantially reduces confidence in the manuscript.

Questions for authors
  • What are the complete definitions of Φ\Phi, UU, SS, S^\hat S, graph construction/update, and query-dependent selection of v0v_0? Under exactly which checkable data-distribution and graph assumptions can a fixed HH, independent of NN, return the global optimum or a stated approximation? In particular, how does Theorem 1 avoid the bounded-degree path-graph counterexample above, and how is Eq. (2) derived rather than assumed?

  • Were any "simulation benchmarks" actually run? If so, please provide the implementation and complete results on standard static and streaming ANN benchmarks, including recall/accuracy versus latency, NN, DD, memory, index construction, insertion/deletion cost, and strong HNSW/DiskANN-style dynamic baselines. For the PIM claim, please provide a reproducible architecture/circuit model and area, power, latency, bandwidth, routing, and technology-node assumptions. This is evidence needed for the present claims, not an optional future-work suggestion.

  • Please provide stable bibliographic identifiers or anonymized-access artifacts for [2], [3], [5], and [9]. In particular, [5] and [9] appear to combine titles, authors, venues, and page ranges from different real PIM publications. Please also reconcile the claim of simulation benchmarks with the absence of any reported experiment.

Reviewer_VFX8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Indeterminate Cognitive Automata (ICA), a conceptual framework for associative memory retrieval using topological graph traversal and processing-in-memory hardware. Sensory inputs are represented as Perceptlets and mapped to Conceptlets organized within an Exalted Data Stage (EDS) graph. Retrieval proceeds through bounded-degree local potential minimization, followed by correlation evaluation. The authors claim constant-time O(1) retrieval independent of memory size, exponentially decreasing correlation error, and O(ND + M) storage complexity. An SRAM-based processing-in-memory architecture is also proposed to reduce data movement. The evaluation primarily consists of theoretical complexity arguments.

Critical strengths

Relevant problem motivation. The paper addresses scalable associative memory retrieval and memory-access bottlenecks, which are important challenges for efficient AI systems.

Interesting hardware–algorithm perspective. Combining graph-based local retrieval with processing-in-memory hardware is a potentially useful direction. The bounded-degree graph formulation also provides a clear starting point for complexity analysis.

Critical weaknesses

The central theoretical guarantees are not adequately established. Appendix A assumes that the maximum number of traversal hops H is independent of repository size N, which essentially presupposes the claimed O(1) complexity. Bounded graph degree and local convexity alone do not guarantee constant-time retrieval of the global optimum. Furthermore, Algorithm 1 terminates at a local minimum, without establishing that this corresponds to the globally optimal match. Appendix B similarly assumes a contraction factor without deriving it from the proposed graph construction or retrieval dynamics. Consequently, the claimed exponential correlation-error bound is insufficiently justified.

Lack of empirical validation and concrete implementation. Despite claims of superior scalability and elimination of hardware bottlenecks, the paper provides no quantitative retrieval experiments, accuracy–latency comparisons, or measured hardware results. Table 1 presents asymptotic claims rather than experimental evidence. The proposed PIM architecture is described conceptually without implementation details or cost analysis. This makes it difficult to assess either the practical advantages or the feasibility of the framework.

Questions for authors

Can the authors provide a non-circular proof that the traversal-hop bound H remains independent of N while guaranteeing retrieval of the global optimum? How are local minima handled?

What assumptions on graph construction and the potential function U ensure the contraction property used in Appendix B? Can the authors rigorously derive the claimed correlation-error bound?

Do the authors have quantitative retrieval results comparing ICA against HNSW or DiskANN in recall, latency, memory consumption, and scaling behavior?

How are the Perceptlet-to-Conceptlet mapping, graph construction, and incremental memory updates implemented? What are their computational costs?

Reviewer_FaRKOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This manuscript is a theory paper claiming that if memory vectors are placed on a graph of degree at most K and a greedy search of at most H hops is run from a given entry point, the optimal memory is found in O(1) time, independent of the number of stored items N. There are no experiments, and the O(1) proof starts from the assumption that the number of hops is independent of N.

Critical strengths

None. I don't see any novelty.

Critical weaknesses

1. The main result, O(1) retrieval independent of N (Theorem 1), is assumed in its proof, and no experiment tests it.

  • The proof in Appendix A states that the number of traversal hops H "is upper-bounded by a constant independent of total node count N" and concludes that search costs K × H = O(1). With K fixed, a bound on H that does not depend on N is the statement of Theorem 1 itself, so the proof assumes its conclusion.
  • Section 6 states that "simulation benchmarks demonstrate the scalability of ICA", but the manuscript reports no simulation or experiment. Table 1 lists only asymptotic complexities.
  • O(1) search is the headline claim of the abstract, the contributions (p. 2), Table 1 and the Conclusion. With neither a valid proof nor an empirical test, the main contribution is unsupported.

2. Three of the ten references do not match the publications they cite.

  • [5] (Mutlu et al., IEEE Micro 39(5), 2019) does not appear to exist. That issue contains no processing-in-memory article by these authors, and its pages 8 to 16 belong to a different article (Moreau et al.).
  • [9] gives the volume, issue and first author of a real article, "In-Memory Computing: Advances and Prospects" (doi:10.1109/MSSC.2019.2922889), but with a different title and three listed co-authors (Lu Jia, Hundal Jung, Vivek De) who are not authors of that article.
  • [6] (DiskANN, NeurIPS 2019) lists two authors, "Fadhil Devaguptapu" and "Rohan Kulkarni", who are not authors of that paper.
  • [5] and [9] are the only works cited for the processing-in-memory substrate (p. 2 and Sec. 4), so that contribution has no checkable support. Citations to publications that do not exist are also an integrity problem, separate from the technical content.
Questions for authors

None

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.