{"id":"9d0854ab-e8fa-487c-a231-c9b27a6505e5","arxiv_id":"2501.08717","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A joint SSL and hierarchical clustering framework is proposed, but the paper provides only a visual demonstration and defers theoretical support to future work.","lead":"The paper proposes InfoHier, a framework that combines self-supervised learning with hierarchical clustering to learn multi-level data structures without labels. It reports only a qualitative visualization on CIFAR100, with no quantitative evaluation or proofs, so the central claim remains unsupported.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that joint SSL+HC training improves both clustering and representation is not supported by any controlled comparison; the only result is a 64-sample visualization, and the paper defers the needed proofs and experiments.","rationale":"The central claim requires that joint training of SSL and HC improves both clustering and representation learning. For that to be supported, the paper must show (i) the HC loss is a faithful surrogate for the hierarchical objective, and (ii) the joint objective outperforms either component alone. The paper explicitly defers the theoretical proof of (i) and the comprehensive experiments that would establish (ii) to future work (Section 5). The only presented evidence is Figure 4, a 64-sample visualization, which is not a controlled experiment: it does not compare against SSL-only or HC-only, reports no clustering or representation quality metric, and provides no hyperparameters or code for reproduction. The reader's weakest assumption about Eq. (1) is fair and important: the triplet similarity wijk is never specified, so the loss is not even fully defined as written. However, I weight the absence of a controlled comparison as the most load-bearing concern because it directly targets the causal claim of joint-training benefit. A pretrained SimCLR encoder could plausibly already separate these superclasses; without baselines, Figure 4 provides no evidence that Lhc contributed. The paper is a vision/position statement rather than a verified result, so the reader's REJECT verdict is unchanged.","tokens_in":6950,"tokens_out":6251,"duration_ms":61664,"concrete_test":"Run the same CIFAR100 pipeline under three conditions: (1) frozen pretrained SimCLR encoder with only the HC loss Lhc; (2) fine-tuned encoder with only the NT-Xent contrastive loss Lct; (3) the joint loss Eq. (3) with the same hyperparameters and training budget. Report quantitative hierarchy quality (e.g., superclass dendrogram purity or Dasgupta cost on the full test set) and representation quality (e.g., linear probe or k-NN accuracy on CIFAR100). If the joint model does not beat both single-loss baselines on both families of metrics, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (Section 1) is that jointly training SSL and HC yields representations that reflect the data hierarchy and improves both clustering and representation learning. The only supporting evidence is Figure 4 (Section 4), a visualization of 64 CIFAR100 samples. This cannot establish the joint-training benefit: a pretrained SimCLR encoder may already separate these superclasses, so the figure does not show that Lhc adds value. No quantitative metrics (e.g., clustering accuracy, dendrogram purity, Dasgupta cost, linear-probe accuracy) and no baselines (SSL-only, HC-only) are reported. The authors explicitly defer 'essential theoretical proofs' and 'comprehensive experiments' to future work (Section 5). In addition, Eq. (1) defines the HC loss only up to an unspecified triplet similarity wijk(EX;w), and its claimed equivalence to the discrete Dasgupta cost is asserted without derivation. The central claim is therefore not yet substantiated by evidence or theory.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InfoHier, a framework that jointly trains a self-supervised contrastive encoder with a hierarchical-clustering loss based on a continuous relaxation of the Dasgupta cost in hyperbolic space. The stated goal is to learn representations that reflect the underlying hierarchy of unlabeled data, improving both clustering and representation learning. The manuscript describes the joint loss, presents a schematic architecture, and reports a single qualitative visualization on 64 CIFAR100 images grouped into four superclasses. The authors explicitly frame the work as preliminary and defer comprehensive experiments and theoretical proofs to future publication.","tokens_in":7132,"tokens_out":3772,"duration_ms":39590,"significance":"The problem addressed is real: SSL methods typically produce flat representations, while classical hierarchical clustering is sensitive to fixed similarity metrics, and combining the two in an end-to-end trainable way is a plausible direction. The paper is clearly written and cites relevant prior work on hyperbolic hierarchical clustering and contrastive SSL. However, as submitted, the manuscript is a position/proposal rather than a validated research contribution: the only experimental evidence is a 64-sample visualization, the central loss is under-specified, and no quantitative evaluation or comparison is provided. If the framework were validated with controlled experiments and a rigorous derivation of the HC loss, it could be of interest to the hierarchical representation learning and information retrieval communities, but that validation is not present here.","major_comments":[{"comment":"The only experimental evidence for the central claim is a qualitative visualization on 64 CIFAR100 samples. No quantitative metrics (e.g., clustering accuracy, dendrogram purity, Dasgupta cost, or linear-probe accuracy), no baselines (e.g., SimCLR-only or HC-only), and no error bars or ablations are reported. Consequently, the claim in Section 1 that joint training 'enhances both clustering and representation learning' is not supported by the presented evidence.","section":"Section 4, Figure 4"},{"comment":"The HC loss is not fully specified: the triplet similarity wijk(EX; w) is described only as being determined by the tree structure or equivalently by the embedding EX, but no explicit formula, decoding procedure from the hyperbolic embedding to a tree, or derivation of the equivalence to the discrete Dasgupta cost is given. Since Section 5 explicitly defers 'essential theoretical proofs' to future publication, the core objective of the method cannot be reproduced or verified from the manuscript as it stands.","section":"Section 3, Eq. (1)"},{"comment":"The visualization does not demonstrate that the HC loss contributes beyond the pretrained SimCLR encoder. Because no comparison to a model trained with the contrastive loss alone is provided, the figure is compatible with the possibility that the pretrained encoder already separates the four CIFAR100 superclasses in the sampled 64 images, making the joint-training benefit unsubstantiated.","section":"Section 4, Figure 4"},{"comment":"The joint objective is a fixed weighted sum with no stated values, tuning procedure, or sensitivity analysis for the hyperparameters lambda_ct and lambda_hc. Given that Section 5 acknowledges different convergence rates of the encoder and embedding networks, the reported behavior cannot be attributed to the proposed joint objective rather than to an arbitrary hyperparameter choice.","section":"Section 3, Eq. (3)"}],"minor_comments":[{"comment":"The notation in Eq. (2) is imprecise: theta, z_i, z'_i, and z_j are not defined before the equation; please specify that theta denotes the trainable encoder parameters and clearly define the positive and negative pairs used in the contrastive loss.","section":"Section 3, Eq. (2)"},{"comment":"The text states that the raw input is encoded by a 'pre-trained encoder' and later says the contrastive loss 'further fine-tune[s] the encoder,' but the training protocol is not specified; clarify whether the encoder is initialized from a SimCLR checkpoint and whether both the encoder and the hyperbolic embedding map are fine-tuned together.","section":"Section 3, paragraph 2"},{"comment":"Reference [32] appears garbled as 'YM., A., C., R., A., V.'; it should be formatted as Asano, Y.M., Rupprecht, C., Vedaldi, A., with the correct title and venue.","section":"References, [32]"},{"comment":"The phrase 'demonstrated on a large-scale problem' conflicts with the actual experiment, which uses only 64 samples; please rephrase to describe the result as a small-scale illustrative study.","section":"Section 5, first paragraph"},{"comment":"The caption and the inline description of Figure 2 are hard to follow, particularly the 'GFP #Distance' table; please clarify the distance values and the intended comparison between images D02, C05, and B04.","section":"Figure 2 and surrounding text"},{"comment":"The claim that hyperbolic embeddings give InfoHier 'greater scalability over traditional methods' is not supported by any experiment or citation in the manuscript; either justify this claim or remove it.","section":"Section 5, second paragraph"}],"recommendation":"reject","confidential_remarks":"This manuscript reads as a workshop abstract or research proposal rather than a completed research article. The central claim is not backed by quantitative results, the key loss is under-specified, and the authors themselves state that essential theoretical proofs and comprehensive experiments are planned for future publication. The idea may be worth developing, but in its current form it does not meet the bar for a full-length paper; a resubmission after a controlled evaluation and a precise derivation of the HC loss would be a different and stronger submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a position/vision paper. The one genuinely new thing is the suggestion to train a contrastive encoder and a hyperbolic Dasgupta-loss embedding jointly with a weighted sum, Eq. (3). The components are both established—NT-Xent/SimCLR and the continuous hyperbolic HC loss from Chami et al. and Monath et al.—so the novelty is the combination, not the math. The paper deserves credit for being straightforward about this: it cites the sources, states that theoretical proofs are future work, and calls the experiments preliminary.\n\nWhat it does well: the motivation is clear, the background is competent, and Figure 4 does at least show that something tree-like can be extracted from 64 CIFAR100 samples. That is a nice sanity check, not a result.\n\nThe soft spots are exactly where the reader put them. The central claim—joint training makes representations hierarchy-aware and improves both clustering and representation learning—is not supported by any controlled comparison. The only evidence is a visualization on 64 samples, and a SimCLR pretrained encoder might already separate those superclasses, so the figure does not isolate the contribution of Lhc. There are no clustering metrics, no dendrogram purity, no Dasgupta cost measurements, no SSL-only or HC-only baselines, and no ablations on the loss-balancing hyperparameters. Eq. (1) is asserted: the triplet similarity wijk(EX;w) is never defined, and the claimed equivalence to the discrete Dasgupta cost is not derived. The paper says so itself, which is honest but means the central objective is currently a placeholder.\n\nI do not think the skeptic overshot here. If this were a workshop position paper, it would be acceptable as a proposal. As a regular paper, the load-bearing claim is unsubstantiated. That said, the idea is not silly and the citation pattern is fine. My recommendation: desk reject for a full venue; if the venue accepts vision statements, it can pass as one. I would not spend referee time on it until there is a comparison showing joint training beats SSL-only and HC-only. When that experiment exists, it is worth a re-look.","headline":"InfoHier is an honest, clearly written vision for combining contrastive SSL with hyperbolic hierarchical clustering, but it contains no evidence that the combination works, so it reads as a project proposal rather than a paper with a supported claim.","tokens_in":7669,"tokens_out":2114,"would_cite":false,"duration_ms":23027,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"InfoHier claims that jointly training a contrastive self-supervised loss with a continuous hierarchical-clustering loss makes learned representations reflect the data's underlying hierarchy, recovering CIFAR100 superclasses from unlabeled…","keywords":["hierarchical representation","hierarchical clustering","self-supervised learning","joint learning","information retrieval","hyperbolic embedding","Dasgupta loss","contrastive learning"],"falsifier":"Train InfoHier on a dataset with a known ground-truth hierarchy and compare the decoded tree against that hierarchy with a tree-agreement measure while monitoring the HC loss; if the loss keeps decreasing but agreement with the ground truth does not improve, the surrogate loss is unfaithful. A simpler check is to disable the hierarchical term entirely and show that the decoded tree is no better than chance at recovering the known superclasses.","tokens_in":6752,"feed_emoji":"🌳","tokens_out":10350,"duration_ms":89256,"temperature":0.7,"pith_summary":"InfoHier proposes to train an encoder and a hyperbolic embedding together so that unlabeled data yield both a flat representation and a hierarchical tree structure. The paper's claim is that jointly minimizing a self-supervised contrastive loss and a continuous version of the Dasgupta hierarchical-clustering loss makes the representations reflect the data's underlying hierarchy, improving both clustering and representation learning. If this holds, multi-level structure could be extracted from large unlabeled datasets without human labels, supporting retrieval, storage management, and data organization. The supporting evidence is a preliminary CIFAR100 experiment in which most samples sharing a superclass fall into the same unsupervised cluster. The paper treats this as an initial demonstration, with theoretical analysis deferred.","feed_headline":"No labels needed: joint training recovers hidden data hierarchies","feed_subtitle":"InfoHier ties contrastive learning to a hyperbolic tree loss, so the learned structure can guide retrieval and storage.","key_machinery":"The load-bearing object is the joint loss $\\mathcal{L} = \\lambda_{\\text{ct}} \\mathcal{L}_{\\text{ct}} + \\lambda_{\\text{hc}} \\mathcal{L}_{\\text{hc}}$, where $\\mathcal{L}_{\\text{hc}}$ is a continuous Dasgupta hierarchical-clustering loss and $\\mathcal{L}_{\\text{ct}}$ is a unified contrastive loss. The HC term is written as a sum over triples $(i,j,k)$ of pairwise similarities minus a triplet similarity $w_{ijk}(\\mathcal{E}_X; w)$ that is determined by the decoded tree, equivalently by the embedding in two-dimensional hyperbolic space. This continuous relaxation is what makes the tree learnable by gradient descent, and the contrastive term is what steers the encoder toward representations that respect the hierarchy. The two gradient flows refine each other: hierarchical clustering gives self-supervised learning a structural target, and self-supervised learning gives clustering adaptive similarity measures.","core_discovery":"The central discovery claim is that self-supervised learning and hierarchical clustering can be co-trained through one differentiable objective, producing latent representations that are hierarchy-aware. The objective pairs a contrastive loss, expressed in a unified form that covers multiple SSL methods, with an HC loss obtained by reformulating the Dasgupta cost as a triplet sum over a two-dimensional hyperbolic embedding. The embedding decodes into a rooted binary tree whose leaves are data points, and the two loss terms are balanced by hyperparameters. On a 64-sample subset of CIFAR100 covering 16 classes and four superclasses, InfoHier reports that most samples of the same superclass form distinct clusters with no label information, and argues that the recovered structure can expose intrinsic hierarchies beyond the human-defined categories.","pith_inferences":["If the claim holds, the Dasgupta cost of the tree induced by a learned representation becomes a label-free proxy for representation quality, giving SSL an evaluation signal that does not need downstream classifiers.","A testable extension: run InfoHier on ImageNet's full class hierarchy and measure whether the decoded tree agrees with the known superclass structure at several depths, not just one level.","The same loop could be inverted: hierarchy-aware clustering could choose hard negatives for contrastive learning, making the structural signal shape the representation instead of only checking it afterward."],"forward_implications":["Unlabeled datasets can be organized into trees at customizable levels of granularity, enabling category-specific retrieval.","The recovered hierarchy supports balanced sampling, giving a structured way to handle imbalanced datasets.","Vector databases can use the hierarchy for hierarchical indexing, improving retrieval efficiency and relevance.","Multi-tiered storage systems could use the structure to identify high-priority data and reduce latency.","Nearest-neighbor search can be narrowed by first descending the hierarchy, benefiting recommendation and search systems."],"supporting_citations":[{"why":"Defines the Dasgupta cost function for similarity-based hierarchical clustering that InfoHier reformulates into a differentiable loss.","marker":"[10]"},{"why":"Supplies the hyperbolic-embedding approach that decodes continuous embeddings back into trees, the basis for InfoHier's HC component.","marker":"[6]"},{"why":"Provides gradient-based hierarchical clustering with continuous tree representations in hyperbolic space, another direct foundation for the HC loss.","marker":"[26]"},{"why":"SimCLR is the contrastive learner used in the CIFAR100 implementation, providing the encoder and the NT-Xent loss.","marker":"[7]"},{"why":"Unifies contrastive SSL losses into the form InfoHier uses for the self-supervised term in the joint objective.","marker":"[29]"},{"why":"Supplies the CIFAR100 dataset and its superclass structure used in the preliminary experiment.","marker":"[21]"},{"why":"ImageNet's category hierarchy motivates the claim that real datasets have multi-level structure SSL currently ignores.","marker":"[12]"},{"why":"Demonstrates ultrametric fitting by gradient descent, supporting the viability of continuous HC optimization.","marker":"[9]"}],"fun_headline_variants":["Self-supervised learning meets hierarchical clustering for hierarchy-aware embeddings","Joint SSL and clustering reveal data hierarchies without labels","InfoHier: one loss to learn both representations and hierarchies","Co-training SSL and HC yields latent structures that reflect hierarchies"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the smooth hierarchical-clustering loss really captures the discrete Dasgupta cost, so that minimizing it yields a tree matching the true data hierarchy; the paper does not prove this equivalence and says the proof is planned for future work.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised learning meets hierarchical clustering for hierarchy-aware embeddings","Joint SSL and clustering reveal data hierarchies without labels","InfoHier: one loss to learn both representations and hierarchies","Co-training SSL and HC yields latent structures that reflect hierarchies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1217,"prompt_tokens":877,"completion_tokens":340,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":275}},"tokens_in":493,"tokens_out":340,"duration_ms":3915,"temperature":1.0,"reasoning_tokens":275,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:18:27.079736+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train InfoHier on a dataset with a known ground-truth hierarchy and compare the decoded tree against that hierarchy with a tree-agreement measure while monitoring the HC loss; if the loss keeps decreasing but agreement with the ground truth does not improve, the surrogate loss is unfaithful. A simpler check is to disable the hierarchical term entirely and show that the decoded tree is no better than chance at recovering the known superclasses.","supporting_citations":[{"cited_title":"In: Proceedings of the 34th Interna- tional Conference on Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"Supplies the hyperbolic-embedding approach that decodes continuous embeddings back into trees, the basis for InfoHier's HC component."},{"cited_title":"Advances in Neural Information Processing Systems35, 19511–19522 (2022)","cited_arxiv_id":null,"evidence_quote":"Unifies contrastive SSL losses into the form InfoHier uses for the self-supervised term in the joint objective."},{"cited_title":"Advances in neural information processing systems32 (2019)","cited_arxiv_id":null,"evidence_quote":"Demonstrates ultrametric fitting by gradient descent, supporting the viability of continuous HC optimization."}],"review_version":1}