{"id":"3dd42427-1297-460d-ae2e-4977a59e6cdf","arxiv_id":"1908.02626","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Structuring Autoencoders enforce a user-chosen class-distance geometry in the latent space via MDS and Procrustes alignment, improving sparse-label classification and confidence calibration.","lead":"This paper introduces Structuring Autoencoders, which add a user-defined class-distance structure to the latent space of a standard autoencoder using multidimensional scaling and a rotation alignment step. The method improves classification with few labels and provides more reliable confidence scores than standard classifiers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Sec. 2.3 Procrustes alignment is not a valid rotation for the paper's setups: MDS targets for 2-3 classes have rank at most 2 in latent dims 10-192, so Z*Z+ != I and tilde{Z}=RZ* need not preserve the intended class distances.","rationale":"The reader's weakest assumption identifies the same load-bearing weakness: the orthogonal alignment in Sec. 2.3 relies on a full-rank condition that the MDS targets cannot satisfy for the small number of classes used in all experiments. My analysis strengthens this by noting that the failure is not a rare edge case but holds in every reported configuration: with k classes, the MDS target after centering has rank at most k-1, so for k=2 or 3 the target is rank-deficient in all latent dimensions used. As a result, the SVD-based rotation is arbitrary on the null space, and the structural loss may be pulling encoder outputs toward a target configuration that does not preserve the class distances that are supposed to drive separation. This is the load-bearing link between the weak supervision and the claimed classification gains. I do not recommend rejection because a standard orthogonal Procrustes solution exists and is likely what the authors intended; the flaw may be a notation/derivation error rather than a failed method. But without a corrected derivation, code, or a reproducibility check, the central claim remains conditional. Therefore the reader's CONDITIONAL verdict is appropriate, so the verdict is UNCHANGED.","tokens_in":10154,"tokens_out":12191,"duration_ms":144997,"concrete_test":"Re-implement Algorithm 1 exactly as written for the MNIST 3-class setup with latent dimension d=10 and 0/1 class-distance matrix D. Stack encoder outputs as columns so Z* is d x n, compute Z+ via a Moore-Penrose routine, set P* = Z @ Z+, take its SVD, set all singular values to 1, form R = U @ S* @ V^T, and compute tilde{Z} = R @ Z*. Then check: (1) whether R^T R = I; (2) whether pairwise distances between the three class centers of tilde{Z} match D; (3) whether a linear SVM on tilde{Z} recovers the classes. Repeat with the standard Procrustes rotation R = U @ V^T from the SVD of Z @ Z*^T. If the literal algorithm fails distance preservation or class separation while the standard Procrustes succeeds, the concern lands and the paper must be corrected or verified; if both preserve D and separate classes, the rank issue is cosmetic and the outcome is unchanged.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of the SAE is the structural loss ||f_enc(x) - tilde{z}||^2, where tilde{Z}=RZ* and R is obtained by an 'orthogonal alignment' in Sec. 2.3 (Eqs. 5-8). The derivation assumes a Moore-Penrose inverse satisfying Z*Z+ = I, stated to hold whenever there are more data points than latent dimensions. That condition is wrong: for an n x d row-stacked configuration the relevant identity is Z+Z* = I and requires full column rank; for a d x n column-stacked configuration it requires full row rank. In every experiment the MDS target is constructed from only k=2 or k=3 class distances, so after centering Z* has rank at most k-1 (i.e., 1 or 2), while the latent dimension d is 10, 30, 64, or 192. Hence Z*Z+ != I in all reported setups. Consequently P* = ZZ+ is rank-deficient, its SVD has zero singular values, and forcing all singular values to 1 selects an arbitrary orthogonal completion. The resulting R is not uniquely determined by the data and is not the Procrustes rotation that best maps Z* to Z. Then tilde{Z}=RZ* can rotate the low-dimensional class structure out of the latent subspace, and the structural loss may not enforce the distances in D. Since no code or pseudo-inverse details are given, the experiments may have used a different, correct Procrustes, but the core algorithm as written is mathematically unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Structuring AutoEncoders (SAE), which augment a standard autoencoder with a structural loss that encourages a user-defined class structure in the latent space. The desired structure is specified by a distance matrix between classes, and at each training iteration the encoder's current latent positions are used to initialize Multidimensional Scaling (MDS), producing target positions Z*. An \"orthogonal alignment\" step computes a rotation R to align Z* with the current latent matrix Z, and the structural loss penalizes the squared distance between encoded points and the aligned targets R Z*. The authors report experiments on MNIST, Fashion-MNIST, DeepFashion2, and a 3D body-shape dataset, showing improved classification with few labels, more reliable decision confidence, effective guided labeling, and class morphing.","tokens_in":10467,"tokens_out":6248,"duration_ms":66640,"significance":"If the proposed method works as claimed, it offers a simple and general mechanism for injecting weak supervision into autoencoders, with practical applications in semi-supervised classification, active labeling, and latent-space interpretability. The paper covers a broad range of datasets and architectures, and it explicitly analyzes the trade-off between reconstruction error and structure via the balancing parameter gamma. The guided-labeling idea is appealing and the morphing demonstration is visually compelling. However, the central Procrustes alignment derivation contains a mathematical error that, as written, invalidates the claimed guarantee that the structural loss enforces the desired class distances. The experimental evaluation also lacks error bars and does not fully control for the use of unlabeled data in the semi-supervised setting. These issues prevent the paper from being accepted in its current form, but they are addressable in a revision.","major_comments":[{"comment":"The derivation of the rotation R is mathematically incorrect. The paper assumes that the Moore-Penrose inverse Z+ of Z* satisfies Z*Z+ = I whenever there are more data points than latent dimensions. For an n x d row-stacked matrix with n>d, the meaningful identity would be Z+Z* = I, and it requires Z* to have full column rank d. In all experiments, Z* is obtained by MDS from class-level distances over k=2 or k=3 classes, so rank(Z*) <= k-1 (1 or 2), while the latent dimension d is 10, 30, 64, or 192. Hence Z*Z+ != I and P* = ZZ+ is singular. Setting nonzero singular values to 1 then produces a matrix R that is a partial isometry rather than an orthogonal rotation, and it is not the standard Procrustes solution. Consequently, tilde{Z} = R Z* need not preserve the distances in D, and the structural loss in Eq. (2) does not provably enforce the intended class structure. This is a load-bearing issue because the structural loss is the core mechanism of the SAE. Please correct the derivation, state the exact pseudo-inverse/Procrustes computation used in the implementation, and verify that the resulting tilde{Z} preserves the target distances in all reported settings.","section":"Sec. 2.3"},{"comment":"The classification curves in Fig. 6 are reported as single runs without error bars or repeated-run statistics. Given that the claimed advantage over baselines is especially pronounced at low labeled-sample counts, the results may be sensitive to random initialization and data subsampling. Please report mean and standard deviation over at least five independent runs with different seeds and subsamples. In addition, clarify whether the \"comparable neural network\" baseline is trained only on the labeled subset while the SAE additionally uses all unlabeled data through its reconstruction loss; if so, the comparison is semi-supervised versus supervised and should be described as such, and the AAE baseline should be clearly distinguished.","section":"Sec. 3.3"}],"minor_comments":[{"comment":"Algorithm 1 says \"set all singular values >= 0 to 1\", which would include zero singular values; the text in Sec. 2.3 correctly says \"setting all nonzero singular values to 1\". Please align the algorithm listing with the text.","section":"Algorithm 1"},{"comment":"The paper describes 6000 labeled samples as \"a very sparse set of data\", but 6000 out of 60000 is 10%. Later experiments use 600 samples. Please use consistent and precise terminology for the amount of supervision.","section":"Sec. 3.1"},{"comment":"The 2D projections in Fig. 4 would benefit from axis labels and a legend; the current figure is hard to read in print.","section":"Fig. 4"},{"comment":"The optimal gamma for the 3D HumanPose dataset is reported as 0.0041, but Fig. 11 does not clearly show the scale of the curves. A log-scale inset or a table would make the behavior more transparent.","section":"Sec. 3.6"},{"comment":"No code or implementation details for the MDS and Procrustes steps are provided; making the code available would substantially improve reproducibility, especially given the mathematical issue raised above.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central idea is attractive and the experimental scope is broad, but the Procrustes error is a serious flaw in the paper's formalism. The authors may well have implemented a correct orthogonal Procrustes step in code, so the issue could be fixable by rewriting Sec. 2.3 and re-evaluating whether the experiments change. I would encourage the editor to request a careful revision that also adds repeated-run statistics and a fair treatment of the supervised baselines. The novelty relative to existing semi-supervised autoencoder methods could also be sharpened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: Structuring Autoencoders has a genuinely useful idea — inject user-specified class distances into an autoencoder latent space via MDS plus a structural loss — and the paper shows it can work impressively. It deserves peer review, not a desk reject, but the math in the alignment step is wrong and needs a rewrite.\n\nWhat's new and good: Combining MDS with Procrustes alignment inside an AE training loop is a new integration. The method is simple to state: define a distance matrix between classes, run MDS to get target positions, then train the AE to pull latent vectors toward those targets. The experiments are encouraging: clean separation on MNIST, Fashion-MNIST, and 3D body shapes, better few-label classification than a plain CNN classifier, and the confidence calibration plots in Fig. 8 are a nice touch. Reconstruction quality stays about the same. The guided labeling application is sensible.\n\nWhere the soft spots are: The Sec. 2.3 Procrustes derivation is mathematically wrong. The paper claims Z*Z+ = I with the justification \"more data points than latent dimensions.\" That's backwards: for a row-stacked MDS target, you need full row rank (which would require n ≤ d), and even then the rank of Z* is at most k–1 for k classes, far below d in all experiments. The \"ideal rotation\" is not the optimal Procrustes rotation, and as written the structural loss may not be enforcing the intended distances. The stress-test note is on point. This is a load-bearing flaw in the formalism, though the algorithm might still work in practice — any rotation of the class centers preserves their pairwise distances, and training can adapt to the chosen rotation. But the paper must fix this.\n\nSecond, there is no code or data release and no error bars. The classification curves in Fig. 6 are single runs, I assume. That makes the quantitative claims hard to verify. Third, the baselines are thin. Comparing to a CNN classifier and an AAE is fine, but not including a simple supervised AE (add a classification loss on the labeled data) or any modern semi-supervised method leaves the superiority claim under-supported.\n\nIn proportion: the central idea is sound; the problems are in the presentation, the theoretical justification, and the experimental rigor. They are fixable.\n\nThis paper will be useful to people working on latent space structuring and semi-supervised image classification. I'd like to see it in the literature after a major revision. Send it to peer review; a solid referee will catch the Procrustes issue and ask for code and stronger baselines, but the core contribution is worth engaging with.","headline":"A useful idea from an integration of MDS and autoencoder training, with a real mathematical error in the alignment step that needs fixing, but the core approach deserves peer review.","tokens_in":10978,"tokens_out":5770,"would_cite":true,"duration_ms":62519,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Structuring Autoencoder can learn a latent space in which user-chosen class distances are preserved, so a linear classifier on the latent variables outperforms comparable models trained directly for classification when labels are scarce.","keywords":["structuring autoencoder","latent space structure","weak supervision","multidimensional scaling","few-label classification","guided labeling","decision confidence","class morphing"],"falsifier":"Take a three-class problem with latent dimension two, train the SAE on a standard dataset, and at an epoch compute the SVD of the MDS target $Z^*$: if $Z^*$ has a zero singular value, or the orthogonal alignment $R=US^*V^T$ fails the orthogonality check $R^TR=I$, then the targets are not reachable by rotation and any observed class separation cannot be attributed to the stated structural-loss mechanism. Reimplementing the algorithm and looking for such checkpoints settles whether the mechanism is doing the work.","tokens_in":9939,"feed_emoji":"🧩","tokens_out":9904,"duration_ms":96946,"temperature":0.7,"pith_summary":"This paper proposes Structuring Autoencoders (SAEs), autoencoders whose training includes a weak-supervision structural loss that pushes the latent representation to match positions computed by multidimensional scaling from a user-defined distance matrix between classes. The claim is that this makes hidden semantic groupings, ones a traditional autoencoder ignores because they are not dominant in reconstruction error, visible and separable in latent space. A consequence the paper demonstrates is that a linear SVM on the structured latent variables classifies sparsely labeled data more accurately than neural networks trained solely for classification, while reconstruction quality stays close to that of an ordinary autoencoder. The same structure also yields meaningful decision confidences and a guided-labeling procedure that identifies which unlabeled examples are most useful to annotate.","feed_headline":"Structured autoencoders beat pure classifiers on few labeled samples","feed_subtitle":"A weak-supervision loss arranges latent classes at preset distances, enabling reliable confidence scores and guided labeling.","key_machinery":"The central mechanism is the per-epoch calculation of target latent positions using multidimensional scaling (MDS), combined with an orthogonal alignment step and a structural loss $L_S = \\|f_{enc}(x) - \\tilde{z}\\|_2^2$. MDS derives target class centers from a fixed distance matrix $D$, and the alignment rotates these targets to match the current latent positions $Z$ before the encoder is pulled toward $\\tilde{Z}=RZ^*$; the reconstruction loss $L_{AE}$ anchors the representation to the data. This treats class geometry as coordinates rather than a classifier boundary, so a linear classifier on the latent variables reads off the intended structure.","core_discovery":"The paper's central claim is that an autoencoder can be trained to honor a user-specified class distance structure in its latent space, even when those classes are invisible in the raw data. At each epoch the encoder projects the whole training set to latent variables $Z$; multidimensional scaling, seeded with the current $Z$, produces target positions $Z^*$ from the fixed class-distance matrix $D$, and an orthogonal alignment rotates $Z^*$ to fit $Z$ as closely as possible. The structural loss $\\mathcal{L}_S = \\|f_{enc}(x) - \\tilde{z}\\|_2^2$ then drives the encoder toward those positions while the reconstruction loss keeps the representation decodable. With this mechanism, class separation appears in latent space for MNIST digits grouped into arbitrary classes, for season-based groupings of Fashion-MNIST, for skirts versus shorts in DeepFashion2, and for male versus female 3D body shapes; a linear SVM on the latent variables outperforms standard classifiers, and the margin information doubles as a calibrated confidence and a guide for which samples to label next.","pith_inferences":["Editorial inference: the same construction should carry over to ordinal or hierarchical label distances, where the MDS target is genuinely high-dimensional and the rank condition is easier to satisfy; the paper only tests equal-distance, few-class settings.","Editorial inference: the monotone score-to-precision relation suggests the latent margin could serve as a reject option for automated decisions in safety-critical settings, an application the paper does not develop.","Editorial inference: comparing SAE-guided labeling with standard uncertainty sampling, such as softmax entropy, would isolate whether the benefit comes from the structured latent geometry or from the boundary-distance heuristic itself."],"forward_implications":["A linear SVM on the SAE latent space gives lower test error than a same-architecture classifier and an adversarial autoencoder baseline, and the gap grows as the number of labeled samples shrinks (Fig. 6).","The reconstruction loss stays close to that of a traditional autoencoder for a wide range of the balancing parameter $\\gamma$; the structural constraint can be added without visibly degrading image or body-shape reconstruction (Figs. 11-13).","The SVM margin in the SAE latent space is monotonically related to actual precision, so the produced confidence values are informative where a softmax classifier's scores are noisy (Fig. 8).","In guided labeling, adding the 100 most uncertain points found by the latent margin to a 600-sample MNIST training set lowers test error from about 4% to 3%, outperforming random sample selection (Fig. 10).","Moving a latent code along the vector between male and female class centers produces a smooth morph between body shapes while the pose stays fixed (Fig. 14)."],"supporting_citations":[{"why":"Supplies the multidimensional scaling procedure that computes target latent positions from the distance matrix.","marker":"[13]"},{"why":"Supplies the orthogonal alignment method used to rotate target positions onto current latent positions.","marker":"[23]"},{"why":"Establishes the autoencoder reconstruction objective that the structural loss extends.","marker":"[11]"},{"why":"Provides the adversarial autoencoder baseline used in semi-supervised classification comparisons.","marker":"[20]"},{"why":"Supplies the MNIST handwritten-digit benchmark used for the artificial three-class experiments.","marker":"[16]"},{"why":"Supplies the Fashion-MNIST dataset with the season-based class decomposition.","marker":"[31]"},{"why":"Supplies the SMPL model used to generate the 3D human body shape dataset.","marker":"[19]"},{"why":"Supplies the DeepFashion2 skirts-versus-shorts subset used for borderline classification and confidence analysis.","marker":"[8]"},{"why":"Supplies the linear SVM classifier applied to the latent variables.","marker":"[28]"}],"fun_headline_variants":["Autoencoders with structured latent spaces beat few-label baselines","Weakly supervised autoencoders arrange classes in latent space","Preset distances help autoencoders separate unseen classes","Structured latent space boosts autoencoder confidence and labeling","SAE: autoencoders get structured latent space via weak labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the MDS target matrix $Z^*$ has full column rank whenever there are more data points than latent dimensions, so that the orthogonal alignment can rotate the target positions to fit the current latent space; with only a few classes the target configuration lies in a lower-dimensional subspace, the pseudoinverse condition $Z^*Z^+=I$ can fail, and the structural loss may then pull the encoder toward positions that do not encode the intended class distances.","fun_headline_variants_meta":{"raw":{"variants":["Autoencoders with structured latent spaces beat few-label baselines","Weakly supervised autoencoders arrange classes in latent space","Preset distances help autoencoders separate unseen classes","Structured latent space boosts autoencoder confidence and labeling","SAE: autoencoders get structured latent space via weak labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1406,"prompt_tokens":905,"completion_tokens":501,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":419}},"tokens_in":521,"tokens_out":501,"duration_ms":5010,"temperature":1.0,"reasoning_tokens":419,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:38:52.364639+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a three-class problem with latent dimension two, train the SAE on a standard dataset, and at an epoch compute the SVD of the MDS target $Z^*$: if $Z^*$ has a zero singular value, or the orthogonal alignment $R=US^*V^T$ fails the orthogonality check $R^TR=I$, then the targets are not reachable by rotation and any observed class separation cannot be attributed to the stated structural-loss mechanism. Reimplementing the algorithm and looking for such checkpoints settles whether the mechanism is doing the work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the multidimensional scaling procedure that computes target latent positions from the distance matrix."},{"cited_title":"Makhzani, J","cited_arxiv_id":null,"evidence_quote":"Provides the adversarial autoencoder baseline used in semi-supervised classification comparisons."},{"cited_title":"Loper, N","cited_arxiv_id":null,"evidence_quote":"Supplies the SMPL model used to generate the 3D human body shape dataset."},{"cited_title":"DeepFashion2: A Versatile Benchmark for Detection, Pose Estimation, Segmentation and Re-Identification of Clothing Images","cited_arxiv_id":"1901.07973","evidence_quote":"Supplies the DeepFashion2 skirts-versus-shorts subset used for borderline classification and confidence analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the linear SVM classifier applied to the latent variables."}],"review_version":1}