{"id":"f5f08c3a-f46f-4463-b140-4db379c8b7e8","arxiv_id":"1908.07699","paper_version":1,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":0.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"This invited survey reviews low-rank tensor decomposition and completion methods for compact uncertainty quantification and deep learning compression, with no new experimental or theoretical result.","lead":"This paper surveys how low-rank tensor decompositions and tensor completion can shrink expensive uncertainty quantification computations and compress neural networks for edge devices. It is a high-level review that classifies tensorized deep learning approaches by whether ranks are fixed or learned, so it is best read as an entry point rather than a new research result.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Survey's central claims rely on unverified, self-cited quantitative results rather than a demonstrated low-rank property of the target tensors.","rationale":"The reader's verdict is UNVERDICTED, which is appropriate for an invited survey that presents no new proof or experiment. The reader identifies low-rankness as the weakest assumption; I agree that this is the load-bearing premise. My stress-test adds a more specific evidential concern: the paper's quantitative claims are not merely inherited from the literature—the two headline numbers (Section III.C and Section IV.D) come from the authors' own work [40] and [56], and the survey provides no experimental details for either. The paper also treats Eq. (3) as a straightforward formulation without acknowledging that tensor completion is nonconvex and that recovery guarantees require conditions such as incoherence and sufficient sampling. This reinforces the reader's correctness-risk assessment. However, because the paper is explicitly a high-level survey, the lack of independent verification does not change the verdict: UNVERDICTED remains the correct status. No internal inconsistency or fatal flaw was found, so the verdict should be unchanged rather than upgraded or rejected.","tokens_in":9881,"tokens_out":9499,"duration_ms":94274,"concrete_test":"Independently reproduce the variability-prediction result from Section III.C: run the Bayesian tensor completion method of [42] on the same or a public multi-die spatial-variation dataset using only 10% of the entries, and compare the relative Frobenius error to the claimed ~0.5%. If the reproduced error is substantially larger (e.g., more than a factor of two) or the automatically selected rank is not small relative to the tensor dimensions, the headline claim that tensor methods 'significantly reduce measurement cost' is not supported for this application.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's abstract promises to 'significantly reduce the simulation or measurement cost' and to 'significantly compress' neural networks. The mechanism is the low-rank assumption encoded in Eq. (3), where the completion solution is constrained to a low-rank manifold M, and in Section IV.C, where weight tensors are assumed compressible via CP, Tucker, or tensor-train decompositions. The paper never quantifies how low these ranks are in the reported applications, and it does not discuss the incoherence or sampling conditions under which the nonconvex completion problem in Eq. (3) can succeed. The only quantitative evidence is self-cited: a ~0.5% relative error with 10% samples for the variability-prediction method of [40], and 7.4x–137x compression ratios for [56], which is the authors' own arXiv preprint. Neither result is accompanied by dataset, model, or accuracy tradeoff details sufficient for independent verification. A reader therefore cannot tell whether the claimed cost reductions are robust or are artifacts of favorable problems and hand-chosen ranks.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This invited survey paper reviews tensor decomposition and completion methods for constructing compact models in two application areas: uncertainty/variability analysis in electronic design automation and deep neural network compression. It introduces CP, Tucker, and tensor-train decompositions, formulates the tensor completion problem as a low-rank constrained least-squares problem in Eq. (3), and then surveys applications in stochastic collocation, hierarchical uncertainty quantification, spatial variation pattern prediction, and tensorized neural network inference and training. The paper also presents a complexity comparison in Table II and highlights the authors' own Bayesian tensorized neural network with automatic rank determination, reporting 0.5% relative error at 10% samples for variation prediction and 7.4x-137x compression for neural networks.","tokens_in":9997,"tokens_out":8990,"duration_ms":83826,"significance":"If the survey is reliable, it provides a useful high-level overview of how tensor methods can reduce simulation and memory costs, and it offers a clear taxonomy of tensorized neural networks (inference vs. training, fixed vs. automatic rank). The paper names a broad set of primary references, and Table II gives a quick side-by-side complexity comparison that is valuable for practitioners. However, the most striking quantitative claims are taken from the authors' own prior work and are not substantiated with experimental context, and the paper does not discuss the conditions under which the underlying low-rank assumption holds. These limitations make the survey more useful as an entry point than as a stand-alone evidence base for the abstract's strong cost-reduction claims.","major_comments":[{"comment":"The quantitative evidence underlying the abstract's central claims is thin and self-cited. In Section III.C, the paper states that the Bayesian tensor completion method achieves \"around 0.5% relative errors with only 10% testing samples\" based on reference [40]. In Section IV.D, it reports \"7.4x to 137x compression ratios\" based on reference [56], which is the authors' own arXiv preprint. Neither result is accompanied by problem dimensions, dataset characteristics, tensor ranks, baseline methods, or the accuracy of the compressed models. Because these numbers are load-bearing for the claims that tensor methods significantly reduce simulation/measurement cost and compress neural networks, please either (a) provide these experimental details, or (b) clearly state that the numbers are reported from the cited papers and direct the reader to the specific tables/figures that support them.","section":"III.C and IV.D"},{"comment":"The entire framework depends on the premise that the relevant tensors are approximately low-rank, as encoded by the set M in Eq. (3) and by the compressibility of weight tensors in Section IV.C, but the paper never states when this premise is plausible or when tensor completion can succeed. In particular, the nonconvex problem in Eq. (3) requires conditions such as incoherence and sufficient sampling, and the reported compression ratios depend on ranks that are either fixed in advance or inferred. For a survey aimed at practitioners, please add a paragraph in Section II (or in the relevant application sections) discussing these conditions and pointing to the analyses in the cited works, so that readers can judge when the claimed cost reductions apply.","section":"II, Eq. (3), and IV.C"}],"minor_comments":[{"comment":"The sentence \"Lebedev et al. [48] firstly applied CP tensor factorization to compress large-scale neural networks with fully connected layers\" misdescribes the cited paper, which is about speeding up convolutional neural networks via CP-decomposition. Please correct the description to refer to convolutional layers, or adjust the historical statement.","section":"IV.A"},{"comment":"The fully connected layer row uses the notation C and S for CP storage and FLOPS, but C and S are not defined for fully connected layers; they should be M and N (the input and output dimensions of W). Please also clarify the ambiguous notation \"cd\", \"sd\", \"dc\", \"ds\", and \"dmr\" in the high-order tensor decomposition columns, either with subscripts or with explicit sums such as r(2l + d(c+s)).","section":"Table II"},{"comment":"The sentence \"For the convolutional layer, we only counts the computational costs of a l x l x C block\" contains a grammar error and is unclear; please rephrase, e.g., \"For the convolutional layer, we count only the computational cost of processing an l x l x C input block.\"","section":"IV.C"},{"comment":"The table states that Tucker decomposition is \"not applicable to high-order tensors due to the curse of dimensionality.\" Tucker decomposition is applicable to high-order tensors, but the core tensor size grows exponentially with the number of modes. Please rephrase to say that Tucker is impractical for very high-order tensors because of the exponential growth of the core.","section":"Table II"},{"comment":"The claim that the method in [56] is the \"first Bayesian tensorized neural network\" is an overclaim that would require a careful literature search to justify. Please soften the wording to \"a Bayesian tensorized neural network\" or restrict the claim to the specific setting of automatic rank selection via Stein variational gradient descent.","section":"IV.D"},{"comment":"The caption of Fig. 1 reads \"Several popular tensor compositions\"; it should be \"Several popular tensor decompositions\".","section":"II"}],"recommendation":"major_revision","confidential_remarks":"The paper is an invited survey, and the central taxonomy is sound and reasonably broad. My main concern is that the most compelling quantitative results are drawn from the authors' own prior work, with one key reference ([56]) being an unreviewed arXiv preprint, and are presented without experimental context. This is fixable by adding details or by explicitly framing the numbers as literature reports. If the venue's survey norms permit such brief reporting, a minor revision might be acceptable, but under standard journal standards the current support for the abstract's claims is not sufficient."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a survey, and it says so in the first line; judging it as a research paper would be a category error. As a survey it is actually decent. The Fig. 2 taxonomy — tensorized inference vs training, crossed with fixed vs automatic rank — is a clean way to organize the deep-learning half, and Table II's complexity comparison is handy for someone choosing a decomposition. Joining UQ and deep learning in one paper is useful; they share low-rank tensor tools but are usually treated separately.\n\nWhat it does not do is present a new algorithm, theorem, or experiment. That is fine for an invited survey, but it means the quantitative claims carry the burden. Here the stress-test is right: the 0.5% relative error with 10% samples (Section III.C) and the 7.4x–137x compression ratios (Section IV.D) are asserted in one line each, with no dataset, model, or accuracy tradeoff. Both point to the authors' own prior work — [40] is their paper and [56] is their own arXiv preprint. A referee cannot check these numbers from the text alone. The 'first Bayesian tensorized neural network' claim in IV.D is a red flag too; [42] already did Bayesian tensor completion with automatic rank, and the paper doesn't explain what makes the NN version first.\n\nThe other soft spot is the load-bearing low-rank assumption. The completion problem (3) restricts to M, and the compression claims assume weight tensors are well approximated by low CP/TT/Tucker ranks. The paper never says how low these ranks are in the reported applications, or under what sampling/incoherence conditions the nonconvex completion succeeds. A few sentences on this would make the survey much more useful to a newcomer.\n\nNone of this is fatal. The central claim — that tensor methods can build compact models for UQ and neural nets — is well supported by the literature the paper cites. The survey is a fair map. Minor issue: Table II marks Tucker as inapplicable to high-order tensors due to the curse of dimensionality; Tucker can be written in high order but the core grows exponentially, so the 'inapplicable' label is too strong.\n\nWho is this for: a graduate student or practitioner wanting a quick map of tensor methods for EDA or model compression. It deserves serious review as a survey; I would accept it with revisions — substantiate or qualify the headline numbers, tone down 'first', and add a paragraph on when the low-rank assumption is safe.\n\nRecommendation: send to referees, but reviewers should treat it as a survey and focus on accuracy and completeness, not novelty.","headline":"A serviceable invited survey of tensor methods for UQ and deep-learning compression; no new results, and some self-cited numbers need scrutiny, but it does what a survey should.","tokens_in":10571,"tokens_out":2842,"would_cite":true,"duration_ms":92914,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A survey positions low-rank tensor factorization as the shared mechanism for cutting simulation cost and compressing neural networks.","keywords":["tensor decomposition","tensor completion","uncertainty quantification","compact neural networks","tensor-train decomposition","CP decomposition","Tucker decomposition","automatic rank determination"],"falsifier":"Take one of the paper's target problems, such as a circuit-simulation output tensor over random parameters, and compute the tensor-train reconstruction error as the ranks grow; if the relative error does not drop below a few percent until the ranks approach the full tensor dimensions, the low-rank assumption fails and the claimed sample and memory savings would not hold.","tokens_in":9615,"feed_emoji":"🧮","tokens_out":6476,"duration_ms":231121,"temperature":0.7,"pith_summary":"The paper's thesis is that representing simulation outputs and neural network weights as low-rank tensors yields compact models that are much cheaper to build and run. In uncertainty quantification, tensor completion estimates the full set of stochastic-collocation simulation samples from a small subset, reducing the number of expensive runs. In deep learning, CP, Tucker, and tensor-train factorizations shrink pretrained networks or allow small tensorized networks to be trained from scratch, and a Bayesian tensorized neural network selects tensor ranks automatically during training. If these methods work as described, the practical payoff is high-dimensional uncertainty analysis and neural network deployment on memory-limited hardware at a fraction of the original cost.","feed_headline":"Low-rank tensors shrink costly simulations and neural networks","feed_subtitle":"The same factorization trick lowers simulation cost and squeezes deep networks onto small devices.","key_machinery":"The load-bearing objects are three tensor factorizations—CP (a sum of rank-one outer products), Tucker (a small core tensor multiplied by orthogonal factor matrices), and tensor-train (a chain of three-dimensional cores)—plus tensor completion, which recovers missing entries by minimizing the fit on observed entries subject to a low-rank constraint. These formats replace a d-dimensional array with a product of small factors, so storage drops from exponential in d to a sum of small factor terms; the same factorizations, applied to neural-network weight tensors, yield the parameter and flop reductions collected in the paper's complexity tables. Tensor completion is what makes the uncertainty-quantification gains possible: only a few simulation samples are needed because the recovery problem is constrained to a low-rank set.","core_discovery":"The paper's central claim is that low-rank tensor factorization is a single practical tool for building compact models in two settings that look very different: expensive simulation-driven uncertainty analysis and memory-limited deep learning. In the first setting, stochastic spectral methods require many simulation samples that grow exponentially with the number of random parameters; tensor completion reconstructs the full sample tensor from a small subset, and tensor-train methods compress high-dimensional integration. In the second setting, the weight arrays of convolution and fully connected layers are folded into high-order tensors and represented in CP, Tucker, or tensor-train formats, either after training or during training; a Bayesian tensorized neural network treats the tensor-train cores as random variables and automatically selects ranks. The paper organizes these results into concrete storage and flop-complexity tables and argues that the same low-rank mechanism underlies both families of gains.","pith_inferences":["Not in the paper: the same low-rank completion logic could be applied hierarchically, treating outputs of subsystems as random inputs, to push the 184-parameter MEMS/IC example toward even larger systems.","Not in the paper: automatic rank selection could double as an interpretability tool, since the effective tensor-train ranks per layer reveal which layers are inherently low-dimensional and therefore safe to compress aggressively.","Not in the paper: if low-rank structure is absent, the framework's own recovery problem would fail to fit, so measuring the reconstruction-error decay of tensor completion across a range of ranks on new data would provide a direct checklist before committing to tensorized training."],"forward_implications":["In uncertainty quantification, tensor completion can replace the full grid of stochastic-collocation simulations with a small sampled subset, making problems with dozens of random parameters (the paper cites up to 57) tractable without dense Monte Carlo.","In variability modeling, Bayesian tensor completion can predict full-die spatial variation maps from about 10% of the measurement sites, reducing chip-testing cost.","In deep learning, tensorized convolution and fully connected layers cut storage and arithmetic by replacing a dense weight array with CP, Tucker, or tensor-train factors, with the paper quantifying parameter and FLOP counts.","For deployment, the train-then-compress path and the train-compact-from-scratch path both produce models that fit memory-limited hardware; the Bayesian tensorized network selects ranks during training and reports compression ratios between 7.4 times and 137 times."],"supporting_citations":[{"why":"Supplies the tensor-completion stochastic collocation scheme that recovers missing simulation samples from a small subset, the core uncertainty-quantification cost reduction.","marker":"[36]"},{"why":"Provides the tensor-train integration method that turns exponential system-level basis-function evaluation into linear cost in the hierarchical uncertainty-quantification pipeline.","marker":"[37]"},{"why":"Reports the multi-die spatial variation prediction result, about 0.5% relative error from 10% testing samples, that anchors the variability section.","marker":"[40]"},{"why":"Serves as the virtual-probe baseline that the tensor-completion approach compares against in variability prediction.","marker":"[41]"},{"why":"Gives the Bayesian CP completion machinery with automatic rank determination used for variation-pattern prediction.","marker":"[42]"},{"why":"Presents the Bayesian tensorized neural network with automatic rank selection, the paper's key method for tensorized training.","marker":"[56]"},{"why":"Introduces CP-based compression of convolutional networks, the starting point for tensorized inference with fixed rank.","marker":"[48]"},{"why":"Shows how to train fully connected layers directly in tensor-train format, the basis of tensorized training with fixed rank.","marker":"[52]"},{"why":"Supplies the Tucker-2 convolution layer formulation whose storage and FLOP counts appear in the paper's complexity tables.","marker":"[57]"}],"fun_headline_variants":["Tensor factorization slices high-dimensional costs","Low-rank tensors slim down simulations and deep nets","One tensor trick compresses both simulations and neural nets","Tensors compact costly models for UQ and deep learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's case stands or falls on arrays of simulation outputs, measured variation maps, and trained network weights being approximately low-rank; without that hidden structure, tensor decomposition and completion cannot recover the missing data or compress the weights.","fun_headline_variants_meta":{"raw":{"variants":["Tensor factorization slices high-dimensional costs","Low-rank tensors slim down simulations and deep nets","One tensor trick compresses both simulations and neural nets","Tensors compact costly models for UQ and deep learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000305,"raw_usage":{"total_tokens":1695,"prompt_tokens":832,"completion_tokens":863,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":448,"completion_tokens_details":{"reasoning_tokens":803}},"tokens_in":448,"tokens_out":863,"duration_ms":7525,"temperature":1.0,"reasoning_tokens":803,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:59:36.930596+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one of the paper's target problems, such as a circuit-simulation output tensor over random parameters, and compute the tensor-train reconstruction error as the ranks grow; if the relative error does not drop below a few percent until the ranks approach the full tensor dimensions, the low-rank assumption fails and the claimed sample and memory savings would not hold.","supporting_citations":[{"cited_title":"Big-data tensor recovery for high-dimensional uncertainty quan- tification of process variations,","cited_arxiv_id":null,"evidence_quote":"Supplies the tensor-completion stochastic collocation scheme that recovers missing simulation samples from a small subset, the core uncertainty-quantification cost reduction."},{"cited_title":"Enabling high-dimensional hierarchical uncertainty quantification by ANOV A and tensor-train decomposition,","cited_arxiv_id":null,"evidence_quote":"Provides the tensor-train integration method that turns exponential system-level basis-function evaluation into linear cost in the hierarchical uncertainty-quantification pipeline."},{"cited_title":"Prediction of multi-dimensional spatial variation data via bayesian tensor completion,","cited_arxiv_id":null,"evidence_quote":"Reports the multi-die spatial variation prediction result, about 0.5% relative error from 10% testing samples, that anchors the variability section."},{"cited_title":"Virtual probe: a statistical framework for low-cost silicon characteriza- tion of nanoscale integrated circuits,","cited_arxiv_id":null,"evidence_quote":"Serves as the virtual-probe baseline that the tensor-completion approach compares against in variability prediction."},{"cited_title":"Bayesian cp factorization of incomplete tensors with automatic rank determination,","cited_arxiv_id":null,"evidence_quote":"Gives the Bayesian CP completion machinery with automatic rank determination used for variation-pattern prediction."},{"cited_title":"Speeding-up convolutional neural networks using fine-tuned cp- decomposition,","cited_arxiv_id":null,"evidence_quote":"Introduces CP-based compression of convolutional networks, the starting point for tensorized inference with fixed rank."},{"cited_title":"Tensorizing neural networks,","cited_arxiv_id":null,"evidence_quote":"Shows how to train fully connected layers directly in tensor-train format, the basis of tensorized training with fixed rank."},{"cited_title":"Tensor contraction layers for parsimonious deep nets,","cited_arxiv_id":null,"evidence_quote":"Supplies the Tucker-2 convolution layer formulation whose storage and FLOP counts appear in the paper's complexity tables."}],"review_version":1}