Pith. sign in

REVIEW 4 major objections 5 minor 48 references

Configurations, a single-parameter attraction–repulsion clustering framework, reproduces unsupervised hierarchical organization, novelty sensitivity, and flexible adaptation seen in infant categorization.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Configurations, an existing finite-resolution clustering method, is reframed as a computational model of early cognition and reported to achieve competitive clustering, 87% novelty-detection AUC, and 35% better stability on evolving data.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A plausible cognitive framing undercut by an evaluation pipeline that aligns clusters to ground truth before measuring them; the headline ARI numbers are not evidence as reported. the 4 major comments →

arxiv 2510.19229 v2 pith:VHBOYKXZ submitted 2025-10-22 cs.LG q-bio.NC

Brain-Inspired Perspective on Configurations: Unsupervised Similarity and Early Cognition

Juntang Wang , Yihan Wang , Hao Wu , Dongmian Zou , Shixin Xu This is my paper
classification cs.LG q-bio.NC
keywords configurationsunsupervised clusteringhierarchical clusteringnovelty detectioninfant categorizationattraction-repulsion dynamicsmulti-resolution clusteringmheatmap
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper tries to establish that configurations—a clustering framework with one resolution parameter and attraction–repulsion dynamics—reproduce three hallmarks of early infant cognition without any supervision: hierarchical organization, novelty sensitivity, and flexible adaptation. If that holds, a simple unsupervised method captures a key part of how infants discover categories, and machine learning gains a label-free route to hierarchical structure, novelty detection, and graceful response to changing categories. To make these claims testable, the paper introduces mheatmap, an evaluation tool with proportional mosaic heatmaps and a reverse merge/split alignment algorithm for fair multi-resolution comparison. Experiments report strong agreement with ground truth on Salinas, infant-stimulus, and ImageNet data, 87% AUC for novelty detection, and 35% better stability during dynamic category evolution.

Core claim

On its own terms, the paper claims that configurations—the family of energy-minimizing partitions produced by the Hamiltonian H(ω) as the single resolution parameter γ sweeps from 0 to ∞—embody unsupervised organization, hierarchical flexibility, and novelty sensitivity. The same attraction–repulsion energy that drives clustering also gives a novelty score: novel items raise both attraction and repulsion costs, yielding 87% AUC in separating novel from familiar stimuli. Empirically, configurations achieve higher adjusted Rand index and normalized mutual information than standard baselines on Salinas, infant-study stimuli, and ImageNet, and they mirror infant behavior in the cue-combination e

What carries the argument

The central object is the configuration family Ω = {ω₀, …, ω∞}: all energy-minimizing partitions of n items as the single resolution parameter γ ranges over [0,∞). The carrying mechanism is the Hamiltonian energy H(ω) = −(attraction within clusters) + γ·(repulsion costs), with pairwise weights w⁺ and w⁻ derived from similarity; γ sets the granularity, low values giving coarse superordinate clusters and high values giving fine basic-level distinctions. Stability plateaus—γ intervals where one partition stays optimal—provide the hierarchical levels, and the energy value itself supplies the novelty signal. The paper's added evaluative machinery is mheatmap, whose mosaic heatmaps visualize unequ

Load-bearing premise

The reported superiority of configurations rests on the assumption that γ was chosen by a pre-specified rule rather than by searching for the value that maximizes agreement with ground truth, and that RMS alignment did not use ground-truth labels to reshape clusters before ARI/NMI were computed.

What would settle it

Run the ImageNet and Salinas experiments with γ fixed by a pre-registered rule (e.g., the midpoint of the first stable plateau) and compute ARI/NMI before applying RMS alignment; if the configuration margin over k-means or GMM collapses toward zero, the claim that a single parameter reproduces infant-like categorization is not supported.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • An unlabeled dataset can be organized hierarchically without choosing the number of clusters in advance; one parameter γ sets the granularity, and the plateau structure identifies robust levels.
  • The same energy used for clustering doubles as a novelty detector, so out-of-distribution points can be flagged without training a separate model.
  • Configurations adapt to evolving categories through merge and split moves, and the paper's stability measure indicates smoother tracking than k-means or spectral clustering under category changes.
  • The mheatmap framework with mosaic heatmaps and RMS alignment provides a fairer way to compare clusterings that have different numbers of clusters and arbitrary label permutations, which standard metrics like ARI handle poorly.
  • The infant-stimulus result suggests that the method reproduces the developmental finding that combined perceptual cues support categorization whereas isolated cues do not.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: the γ parameter of configurations may correspond to a perceptual resolution or attentional scale in the brain; if so, the plateau structure predicts that categories are perceived as discrete stable levels rather than continuous gradients, a testable developmental prediction.
  • Inference: the energy-based novelty score could be directly applied to open-world recognition tasks, where unknown classes are expected to raise energy, making the method a candidate for out-of-distribution detection benchmarks.
  • Inference: because RMS alignment changes the reported ARI, the headline numbers depend on the alignment being blind to ground-truth labels; a straightforward robustness check is to compute ARI before and after RMS alignment to see how much of the margin over baselines is alignment-driven.
  • Inference: the cue-combination finding suggests a concrete cross-modal prediction: in audio or text domains, configurations should also show poor performance on isolated features and strong performance on combined features, mirroring the visual infant results.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a 'brain-inspired' interpretation of the configuration clustering framework of Liu et al. and Pitsianis et al., in which a single resolution parameter gamma controls hierarchical granularity via attraction–repulsion energy minimization. The authors claim that this framework exhibits three hallmarks of early cognition: unsupervised organization, hierarchical flexibility, and novelty sensitivity. They introduce an evaluation tool, mheatmap, with a mosaic heatmap visualization and a Reverse Merge/Split (RMS) alignment algorithm, and report strong empirical results: ARI/NMI of 0.92/0.94 on Salinas, 0.55/0.58 on Infant Stimuli, and 0.62/0.68 on ImageNet, plus 87% AUC for novelty detection and 35% better stability in dynamic evolution. The central thesis is that a single-resolution-parameter clustering method can reproduce early-cognitive categorization behavior without supervision.

Significance. If the empirical claims were validated without label leakage, the paper would be a valuable conceptual bridge between finite-resolution clustering and developmental psychology, and the mheatmap tool could be useful for multi-resolution and dynamic clustering evaluation. The paper also has the merit of shipping code and grounding its design in infant categorization studies. However, the central quantitative claims currently rest on an evaluation pipeline that appears to use ground-truth labels inside the alignment procedure before metrics are computed, and on unspecified hyperparameter selection. These issues are load-bearing: without a clean evaluation, the headline numbers and the 'brain-inspired' conclusion are not supported. The conceptual framework itself may be sound, but the evidence as presented does not yet demonstrate the claimed unsupervised abilities.

major comments (4)
  1. [§4.2–4.3 and §5] The RMS alignment algorithm uses ground-truth categories Y to compute overlap counts and performs cluster reassignments (merges/splits) to 'optimize the visual diagonal.' Section 4.3 states that all clustering metrics in Section 5 are calculated after RMS alignment. This means the reported ARI/NMI values are not computed on the raw clustering output; they are computed on clusters that have been modified using label information. Since ARI/NMI are not invariant to merging/splitting, this procedure can inflate scores and constitutes label leakage. This directly contradicts the claim in Section 5 that 'ground-truth labels are used solely for evaluation.' To support the headline results, the authors must either (a) report metrics computed on raw configurations without any label-informed post-processing, or (b) demonstrate that RMS alignment does not use Y in any way that could improve agreeme
  2. [§5, Table 1] No gamma selection protocol is reported. Table 1 gives no gamma values and the text does not specify how the single resolution parameter was chosen for each dataset. If gamma was selected per dataset by maximizing agreement with ground-truth labels, then the 'single parameter' is being fit to the labels and the comparison against baselines is unfair. If gamma was selected by a principled, pre-specified rule (e.g., largest plateau, a fixed default, or a stability criterion), that rule must be stated. Without this information, the reported ARI/NMI scores are not reproducible and the claim that configurations require only one parameter with no label information is not verifiable.
  3. [§5, Fig. 4b and 4c] The novelty detection result (87% AUC) and the stability result (35% better 1/ARI) lack protocol details. For novelty detection, the paper does not define how 'novel' vs 'familiar' stimuli were constructed, which energy score (raw H, ha, hr, or something else) was used, how the ROC curve was computed, or whether the energy-based score was post-processed by RMS. For dynamic adaptation, the definition of '35% better stability' is not made precise: lower 1/ARI than which baselines, over which time steps, and with what variance? Figure 4c shows curves but no error bars. These omissions prevent assessment of two of the three central cognitive claims.
  4. [§5, Table 2] The claim that configurations 'best align with infant behavior' is not quantitatively supported. All methods in Table 2 show the same qualitative pattern: low ARI on internal/external cues and higher ARI on combined cues. Configurations do have the largest gap (0.15→0.65), but no statistical test (e.g., interaction effect or difference-in-differences) is provided to show this gap is significantly larger than for the baselines (e.g., GMM 0.21→0.58, Spectral 0.19→0.56). The reader cannot conclude that configurations uniquely mirror the infant pattern.
minor comments (5)
  1. [§4.2] The RMS algorithm is central to the evaluation but the paper says 'we omit detailed algorithmic descriptions' due to space constraints. This is a significant omission; at minimum, a pseudocode or supplementary description should be included so that the label-leakage question can be resolved.
  2. [§3.2] The statement that 'dissimilar items increase both attraction costs (ha) and repulsion costs (hr)' is imprecise. In Eq. (1), attraction enters as -ha, so a dissimilar pair in the same cluster reduces the magnitude of the attractive term (less negative) and increases repulsion; the phrasing should be corrected to avoid confusion.
  3. [Fig. 1c/1d] The caption reports 'Acc=0.99' and 'Acc=0.92' but does not define what 'Acc' means (cluster accuracy? after alignment?) or how it is computed. This is needed for interpretation.
  4. [§4.3] Typo: 'futher' should be 'further.'
  5. [References] References [26] and [32] have incomplete entries (no year, no full bibliographic data). The reference formatting should be cleaned up.

Circularity Check

1 steps flagged

Reported ARI/NMI are computed after RMS alignment that takes ground-truth categories Y as input, so the headline clustering numbers are partly manufactured by label-informed merge/split reassignment.

specific steps
  1. fitted input called prediction [Sec. 4.1-4.3 and Sec. 5 (Table 1)]
    "Let Y={Y_i} be ground-truth categories and Ĉ={Ĉ_j} predicted clusters, with overlap counts N_ij=|Y_i ∩ Ĉ_j| ... we developed the Reverse Merge/Split (RMS) algorithm, which optimizes the visual “diagonal” of the mosaic heatmap by cluster reassignments considering merge-split dynamics. ... All clustering metrics reported in our experiments (Section 5) are calculated after RMS alignment to ensure fair comparison ... We provide only unlabeled inputs to clustering algorithms, using ground-truth labels solely for evaluation."

    By construction, RMS takes ground-truth categories Y as input, computes overlap counts N_ij with predicted clusters, and then reassigns/merges/splits clusters to optimize the diagonal against Y. Section 4.3 then states that every clustering metric in Section 5 is calculated after RMS alignment. Therefore the ARI/NMI values in Table 1 describe label-informed post-processed partitions, not the raw unsupervised configurations. The claim that ground-truth labels are used “solely for evaluation” is violated inside the evaluation pipeline, and the headline “configurations are competitive” reduces in part to a reassignment procedure fit to those same labels.

full rationale

The configuration framework itself is not circular: the Hamiltonian, Parallel-DT, and the correspondence to modularity are cited from external prior work ([1,2], [27]), and the novelty-score derivation from the same energy function is a legitimate within-model construction rather than a circular prediction-equals-input case. The main circularity is in the evaluation protocol. Section 4.1 defines RMS using ground-truth categories Y; Section 4.2 says RMS “optimizes the visual diagonal ... by cluster reassignments considering merge-split dynamics”; Section 4.3 says all Section 5 clustering metrics are calculated after RMS alignment. Because ARI/NMI are sensitive to merging and splitting, this label-informed alignment can inflate scores without reflecting genuine unsupervised structure, so the Table 1 headline numbers are not evidence for purely unsupervised clustering as claimed. I did not flag gamma selection as a separate circular step because the paper does not state the selection rule, so no reduction can be exhibited from the text. No load-bearing self-citation chain is present, and the mathematical derivation of configurations is independent of the evaluation issue.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The paper contributes a perspective and an evaluation tool; the underlying Hamiltonian/Parallel-DT framework and its guarantees are assumed from [1,2]. The key additional assumptions are that energy H is a valid novelty score and that unsupervised evaluation does not use labels in gamma selection or RMS alignment (the latter is likely violated as described). No new physical or mathematical entities are introduced.

free parameters (2)
  • Resolution parameter gamma = not reported; used values include 0.3 and 1.4 in Fig. 4a
    Controls cluster granularity. If selected per dataset by inspecting ground-truth ARI, the reported clustering metrics are optimized rather than predicted.
  • kNN neighborhood size k for attraction/repulsion weights = not reported
    Attraction and repulsion weights are derived from a kNN graph [2]; k is a hyperparameter, but the paper gives no value or sensitivity analysis.
axioms (4)
  • domain assumption Minimizing the Hamiltonian energy H(omega) yields 'good' partitions; lower H corresponds to better clustering.
    This is the framework's core criterion, borrowed from [1,2], not justified in this paper beyond a cognitive analogy.
  • domain assumption Parallel-DT correctly enumerates all dominant configurations and plateaus as gamma varies.
    The paper explicitly says 'Design and proofs of the whole Omega framework, including Parallel-DT, are not our focus' and cites [1,2]; the hierarchy and stability claims rely on this unproved-in-text guarantee.
  • domain assumption The energy H is a valid novelty score such that novel items increase both attraction and repulsion costs.
    Novelty detection depends on this monotonic claim, but no derivation or independent benchmark is given.
  • domain assumption Ground-truth labels are used only for evaluation, not for learning or alignment.
    Stated in Sec. 5, but contradicted by the RMS description in Sec. 4.2, where ground-truth categories Y are used to optimize reassignments before computing metrics.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Brain-Inspired Perspective on Configurations: Unsupervised Similarity and Early Cognition." pith.science (2026). https://pith.science/paper/VHBOYKXZ

@misc{pith2026251019229,
  author       = {Pith},
  title        = {Pith review of: Brain-Inspired Perspective on Configurations: Unsupervised Similarity and Early Cognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VHBOYKXZ}},
  note         = {Machine review of arXiv:2510.19229}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Infants discover categories, detect novelty, and adapt to new contexts without supervision-a challenge for current machine learning. We present a brain-inspired perspective on configurations, a finite-resolution clustering framework that uses a single resolution parameter and attraction-repulsion dynamics to yield hierarchical organization, novelty sensitivity, and flexible adaptation. To evaluate these properties, we introduce mheatmap, which provides proportional heatmaps and reassignment algorithm to fairly assess multi-resolution and dynamic behavior. Across datasets, configurations are competitive on standard clustering metrics, achieve 87% AUC in novelty detection, and show 35% better stability during dynamic category evolution. These results position configurations as a principled computational model of early cognitive categorization and a step toward brain-inspired AI.

Figures

Figures reproduced from arXiv: 2510.19229 by Dongmian Zou, Hao Wu, Juntang Wang, Shixin Xu, Yihan Wang.

Figure 1
Figure 1. Figure 1: Configuration lineage and energy landscapes. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) A schematic example of merge and split happens as γ increases from i to j. (b) Comparison of normal and mosaic heatmap for a confusion matrix. Mosaic heatmap provides more intuitive visualization with interpretable diagonal structure, when cases like a split of 1 into 1 and 2 happen [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: One case of clustering versus GT before and after RMS alignment [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Brain-inspired capabilities of configurations. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 9 linked inside Pith

  1. [1]

    Digraph Clustering by the BlueRed Method

    T. Liu, D. Floros, N. Pitsianis, and X. Sun. “Digraph Clustering by the BlueRed Method”. In:2021 IEEE High Performance Extreme Computing Conference (HPEC). 2021 IEEE High Performance Extreme Computing Conference (HPEC). Sept. 2021, pp. 1–7

    2021

  2. [2]

    Parallel Clustering with Res- olution Variation

    N. Pitsianis, D. Floros, T. Liu, and X. Sun. “Parallel Clustering with Res- olution Variation”. In:2023 IEEE High Performance Extreme Computing Conference (HPEC). 2023 IEEE High Performance Extreme Computing Conference (HPEC). Boston, MA, USA: IEEE, Sept. 25, 2023, pp. 1–8

    2023

  3. [3]

    Perceptual Cues That Permit Categorical Differentiation of Animal Species by Infants

    P. C. Quinn and P. D. Eimas. “Perceptual Cues That Permit Categorical Differentiation of Animal Species by Infants”. In:Journal of Experimental Child Psychology63.1 (Oct. 1996), pp. 189–211. PMID: 8812045

    1996

  4. [4]

    Basic-Level and Superordinate-like Categorical Repre- sentations in Early Infancy

    G. Behl-Chadha. “Basic-Level and Superordinate-like Categorical Repre- sentations in Early Infancy”. In:Cognition60.2 (Aug. 1, 1996), pp. 105– 141

    1996

  5. [5]

    Categorization in Infancy

    D. Mareschal and P. C. Quinn. “Categorization in Infancy”. In:Trends in Cognitive Sciences5.10 (Oct. 1, 2001), pp. 443–450

    2001

  6. [6]

    The Development of Visual Attention in Infancy

    J. Colombo. “The Development of Visual Attention in Infancy”. In:Annual Review of Psychology52 (Volume 52, 2001 Feb. 1, 2001), pp. 337–367

    2001

  7. [7]

    Krizhevsky.Learning Multiple Layers of Features from Tiny Images

    A. Krizhevsky.Learning Multiple Layers of Features from Tiny Images. Toronto, ON, Canada, 2009, pp. 32–33

    2009

  8. [8]

    Deep Residual Learning for Image Recognition

    K. He, X. Zhang, S. Ren, and J. Sun. “Deep Residual Learning for Image Recognition”. In:arXiv:1512.03385 [cs.CV](Dec. 10, 2015). arXiv: 1512. 03385[cs.CV]

    Pith/arXiv arXiv 2015

  9. [9]

    A Simple Framework for Contrastive Learning of Visual Representations

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. “A Simple Framework for Contrastive Learning of Visual Representations”. In:arXiv:2002.05709 [cs.LG](July 1, 2020). arXiv: 2002.05709[cs.LG]

    Pith/arXiv arXiv 2002

  10. [10]

    Doersch, B

    J.-B.Grill,F.Strub,F.Altché,C.Tallec,P.H.Richemond,E.Buchatskaya, C. Doersch, B. A. Pires, Z. D. Guo, M. G. Azar, B. Piot, K. Kavukcuoglu, R. Munos, and M. Valko.Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning. Sept. 10, 2020. arXiv: 2006.07733[cs].url: http://arxiv.org/abs/2006.07733 (visited on 09/01/2025). Pre-published

    Pith/arXiv arXiv 2020

  11. [11]

    Caron, H

    M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, and A. Joulin.Emerging Properties in Self-Supervised Vision Transformers. May 24, 2021. arXiv: 2104.14294[cs].url: http://arxiv.org/abs/2104. 14294 (visited on 03/25/2025). Pre-published. Brain-Inspired Configurations 11

    Pith/arXiv arXiv 2021

  12. [12]

    B. M. Lake, T. D. Ullman, J. B. Tenenbaum, and S. J. Gershman.Build- ing Machines That Learn and Think like People. Nov. 2, 2016. arXiv: 1604.00289[cs].url: http://arxiv.org/abs/1604.00289 (visited on 09/01/2025). Pre-published

    Pith/arXiv arXiv 2016

  13. [13]

    Least Squares Quantization in PCM

    S. Lloyd. “Least Squares Quantization in PCM”. In:IEEE Transactions on Information Theory28.2 (Mar. 1982), pp. 129–137

    1982

  14. [14]

    Normalized Cuts and Image Segmentation

    Jianbo Shi and J. Malik. “Normalized Cuts and Image Segmentation”. In: IEEE Transactions on Pattern Analysis and Machine Intelligence22.8 (Aug. 2000), pp. 888–905

    2000

  15. [15]

    A Survey of Clustering Data Mining Techniques

    P. Berkhin. “A Survey of Clustering Data Mining Techniques”. In:Group- ing Multidimensional Data: Recent Advances in Clustering. Ed. by J. Ko- gan, C. Nicholas, and M. Teboulle. Berlin, Heidelberg: Springer, 2006, pp. 25–71

    2006

  16. [16]

    A Density-Based Algorithm for Dis- covering Clusters in Large Spatial Databases with Noise

    M. Ester, H.-P. Kriegel, and X. Xu. “A Density-Based Algorithm for Dis- covering Clusters in Large Spatial Databases with Noise”. In:Second Inter- national Conference on Knowledge Discovery and Data Mining (KDD’96). Proceedings of a conference held August 2-4(1996), pp. 226–331

    1996

  17. [17]

    VisualExperienceinInfants:DecreasedAttentiontoFamiliar Patterns Relative to Novel Ones

    R.L.Fantz.“VisualExperienceinInfants:DecreasedAttentiontoFamiliar Patterns Relative to Novel Ones”. In:Science146.3644 (1964), pp. 668–

    1964

  18. [18]

    Fränken, C

    J.-P. Fränken, C. G. Lucas, N. R. Bramley, and S. T. Piantadosi.Model- ing Infant Object Perception as Program Induction. Aug. 28, 2023. arXiv: 2309.07099[q-bio].url: http://arxiv.org/abs/2309.07099 (visited on 09/10/2025). Pre-published

    Pith/arXiv arXiv 2023

  19. [19]

    Eight- Month-Old Infants Meta-Learn by Downweighting Irrelevant Evidence

    F. Poli, T. Ghilardi, R. B. Mars, M. Hinne, and S. Hunnius. “Eight- Month-Old Infants Meta-Learn by Downweighting Irrelevant Evidence”. In:Open Mind: Discoveries in Cognitive Science7 (2023), pp. 141–155. PMID: 37416070

    2023

  20. [20]

    X. Ke, S. Tsutsui, Y. Zhang, and B. Wen.Discovering Hidden Visual Con- cepts beyond Linguistic Input in Infant Learning. June 13, 2025. arXiv: 2501.05205[cs].url: http://arxiv.org/abs/2501.05205 (visited on 09/10/2025). Pre-published

    Pith/arXiv arXiv 2025

  21. [21]

    Comparing Partitions

    L. J. Hubert and P. Arabie. “Comparing Partitions”. In:Journal of Clas- sification2.2–3 (1985), pp. 193–218

    1985

  22. [22]

    Objective Criteria for the Evaluation of Clustering Meth- ods

    W. M. Rand. “Objective Criteria for the Evaluation of Clustering Meth- ods”. In:Journal of the American Statistical Association66.336 (1971), pp. 846–850. JSTOR: 2284239

    1971

  23. [23]

    Cluster Ensembles — A Knowledge Reuse Frame- work for Combining Multiple Partitions

    A. Strehl and J. Ghosh. “Cluster Ensembles — A Knowledge Reuse Frame- work for Combining Multiple Partitions”. In:Journal of Machine Learning Research3 (Dec 2002), pp. 583–617

    2002

  24. [24]

    V-Measure: A Conditional Entropy- Based External Cluster Evaluation Measure

    A. Rosenberg and J. Hirschberg. “V-Measure: A Conditional Entropy- Based External Cluster Evaluation Measure”. In:Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL).EMNLP- 12 J. Wang et al. CoNLL 2007. Ed. by J. Eisner. Prague, Czech Republic: Association f...

    2007

  25. [25]

    From Louvain to Lei- den: Guaranteeing Well-Connected Communities

    V. A. Traag, L. Waltman, and N. J. Van Eck. “From Louvain to Lei- den: Guaranteeing Well-Connected Communities”. In:Scientific Reports 9.1 (Mar. 26, 2019), p. 5233

    2019

  26. [26]

    Learning Multiple Layers of Features from Tiny Images

    A. Krizhevsky. “Learning Multiple Layers of Features from Tiny Images”. In: ()

  27. [27]

    Fast Un- folding of Communities in Large Networks

    V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre. “Fast Un- folding of Communities in Large Networks”. In:Journal of Statistical Me- chanics: Theory and Experiment2008.10 (Oct. 2008), P10008

    2008

  28. [28]

    The MNIST Database of Handwritten Digit Images for Ma- chine Learning Research [Best of the Web]

    L. Deng. “The MNIST Database of Handwritten Digit Images for Ma- chine Learning Research [Best of the Web]”. In:IEEE Signal Processing Magazine29.6 (Nov. 2012), pp. 141–142

    2012

  29. [29]

    Zhong, X

    Y. Zhong, X. Hu, C. Luo, X. Wang, J. Zhao, and L. Zhang. “WHU-Hi: UAV-borne Hyperspectral with High Spatial Resolution (H2) Benchmark Datasets and Classifier for Precise Crop Identification Based on Deep Con- volutional Neural Network with CRF”. In:Remote Sensing of Environment 250 (Dec. 2020), p. 112012

    2020

  30. [30]

    The Development of Categorization in Early Childhood:AReview

    K. Owen and C. Barnes. “The Development of Categorization in Early Childhood:AReview”.In:Early Child Development and Care191.1(2021), pp. 13–20

    2021

  31. [31]

    Grana, M

    M. Grana, M. A. Veganzons, and B. Ayerdi.Hyperspectral Remote Sensing Scenes. Grupo de Inteligencia Computacional (GIC). 2020.url: https: //www.ehu.eus/ccwintco/index.php/Hyperspectral_Remote_Sensing_ Scenes

    2020

  32. [32]

    ImageNet: A Large-Scale Hierarchical Image Database

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. “ImageNet: A Large-Scale Hierarchical Image Database”. In: ()

  33. [33]

    An Image Is Worth 16x16 Words: Transformers for Im- age Recognition at Scale

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Un- terthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. “An Image Is Worth 16x16 Words: Transformers for Im- age Recognition at Scale”. In:arXiv:2010.11929 [cs.CV](June 3, 2021). arXiv: 2010.11929[cs.CV]

    Pith/arXiv arXiv 2010

  34. [34]

    Maximum Likelihood from Incomplete Data via theEMAlgorithm

    A. P. Dempster, N. M. Laird, and D. B. Rubin. “Maximum Likelihood from Incomplete Data via theEMAlgorithm”. In:Journal of the Royal Statistical Society Series B: Statistical Methodology39.1 (Sept. 1, 1977), pp. 1–22

    1977

  35. [35]

    Machine Learning Accurately Classifies Neural Responses to Rhythmic Speech vs. Non-Speech from 8-Week-Old Infant EEG

    S. Gibbon, A. Attaheri, Á. Ní Choisdealbha, S. Rocha, P. Brusini, N. Mead, P. Boutris, H. Olawole-Scott, H. Ahmed, S. Flanagan, K. Mandke, M. Keshavarzi, and U. Goswami. “Machine Learning Accurately Classifies Neural Responses to Rhythmic Speech vs. Non-Speech from 8-Week-Old Infant EEG”. In:Brain and Language220 (Sept. 1, 2021), p. 104968

    2021

  36. [36]

    Tsutsui, A

    S. Tsutsui, A. Chandrasekaran, M. A. Reza, D. Crandall, and C. Yu.A Computational Model of Early Word Learning from the Infant’s Point of View. June 4, 2020. arXiv: 2006.02802[cs].url: http://arxiv.org/abs/ 2006.02802 (visited on 09/10/2025). Pre-published. Brain-Inspired Configurations 13

    Pith/arXiv arXiv 2020

  37. [37]

    Curiosity and Exploration: Facilitating Positive Subjective Experiences and Personal Growth Oppor- tunities

    T. B. Kashdan, P. Rose, and F. D. Fincham. “Curiosity and Exploration: Facilitating Positive Subjective Experiences and Personal Growth Oppor- tunities”. In:Journal of Personality Assessment82.3 (June 2004), pp. 291–

    2004

  38. [38]

    Simplicity and Validity in Infant Research

    J. F. Kominsky, K. Lucca, A. J. Thomas, M. C. Frank, and J. K. Hamlin. “Simplicity and Validity in Infant Research”. In:Cognitive Development 63 (July 1, 2022), p. 101213

    2022

  39. [39]

    Eighteen-Month-Old Infants Represent Nonlo- cal Syntactic Dependencies

    L. Perkins and J. Lidz. “Eighteen-Month-Old Infants Represent Nonlo- cal Syntactic Dependencies”. In:Proceedings of the National Academy of Sciences118.41 (Oct. 12, 2021), e2026469118

    2021

  40. [40]

    Infants Tailor Their Attention to Maximize Learning

    F. Poli, G. Serino, R. B. Mars, and S. Hunnius. “Infants Tailor Their Attention to Maximize Learning”. In:Science Advances6.39 (Sept. 23, 2020), eabb5053

    2020

  41. [41]

    Comparing Clusterings—an Information Based Distance

    M. Meilă. “Comparing Clusterings—an Information Based Distance”. In: Journal of Multivariate Analysis98.5 (May 1, 2007), pp. 873–895

    2007

  42. [42]

    D. O. Hebb.The Organization of Behavior; a Neuropsychological Theory. The Organization of Behavior; a Neuropsychological Theory. Oxford, Eng- land: Wiley, 1949, pp. xix, 335. xix, 335

    1949

  43. [43]

    Attention Is All You Need

    A.Vaswani,N.Shazeer,N.Parmar,J.Uszkoreit,L.Jones,A.N.Gomez,Ł. ukasz Kaiser, and I. Polosukhin. “Attention Is All You Need”. In:Advances in Neural Information Processing Systems. Vol. 30. Curran Associates, Inc., 2017

    2017

  44. [44]

    Neural Networks and Physical Systems with Emergent Col- lective Computational Abilities

    J. J. Hopfield. “Neural Networks and Physical Systems with Emergent Col- lective Computational Abilities.” In:Proceedings of the National Academy of Sciences79.8 (Apr. 1982), pp. 2554–2558

    1982

  45. [45]

    Infants Use Contextual Memory to Attend and Learn in Naturalistic Scenes

    K. Tummeltshammer and D. Amso. “Infants Use Contextual Memory to Attend and Learn in Naturalistic Scenes”. In:Infancy: The Official Journal of the International Society on Infant Studies28.3 (May 2023), pp. 634–

    2023

  46. [46]

    Baby Intuitions Benchmark (BIB): Discerning the Goals, Preferences, and Actions of Oth- ers

    K. Gandhi, G. Stojnic, B. M. Lake, and M. R. Dillon. “Baby Intuitions Benchmark (BIB): Discerning the Goals, Preferences, and Actions of Oth- ers”. In:Advances in Neural Information Processing Systems. Vol. 34. Cur- ran Associates, Inc., 2021, pp. 9963–9976

    2021

  47. [47]

    How Infant-Directed Actions Enhance Infants’ Attention, Learning, and Exploration: Evidence from EEG and Computational Modeling

    M. Meyer, J. E. van Schaik, F. Poli, and S. Hunnius. “How Infant-Directed Actions Enhance Infants’ Attention, Learning, and Exploration: Evidence from EEG and Computational Modeling”. In:Developmental Science26.1 (Jan. 2023), e13259. PMID: 35343042

    2023

  48. [48]

    Rapid Infant Learning of Syntactic–Semantic Links

    M. Barbir, M. J. Babineau, A.-C. Fiévet, and A. Christophe. “Rapid Infant Learning of Syntactic–Semantic Links”. In:Proceedings of the National Academy of Sciences120.1 (Jan. 3, 2023), e2209153119

    2023

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.