{"id":"ac12a5d6-27ae-4888-b6f4-39cae71407f7","arxiv_id":"1909.01844","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Deep kernel learning is extended to line-integral observations; a Hilbert space basis expansion turns the costly 2D integrals into 1D integrals, and neural-network pre-training stabilizes joint training.","lead":"The authors combine Gaussian processes with neural-network input warping and make the approach work when the data are line-integral measurements, as in X-ray computed tomography. Their basis-function trick converts hard double integrals into single integrals, and they demonstrate the idea on simulated and real CT scans.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central scalability claim rests on the truncated basis expansion (Eq. 9) remaining accurate for trained neural-network inputs u(x); the paper gives no error bound or post-training validation, so the equivalence to exact deep kernel learning is unproven.","rationale":"The reader's verdict is CONDITIONAL, and the weakest assumption identified there is also the most load-bearing issue I can find. The paper's central contribution is a practical way to do deep kernel learning with line-integral observations, and that practicality rests on the Hilbert-space basis approximation being accurate after the neural network is trained. The paper gives a heuristic for choosing the domain size, mentions that the chosen Lk is not optimized, and states without evidence that the resulting effect is negligible, but it never validates the approximation on the trained network's actual input distribution. This is a correctness risk rather than a stylistic or comparison issue: if the basis is not converged, the efficient single-integral computations implement a different covariance than the intended deep kernel, and the claims about scalability and reconstruction quality no longer follow. I do not see a reason to move the verdict from CONDITIONAL to REJECT, because the required validation is straightforward and the method may well survive it; the paper's derivation is otherwise clear, the pretraining idea is sensible, and the experiments, though qualitative, are consistent with the method working. The concern is addressable by an explicit numerical check, so the appropriate recommendation remains conditional acceptance pending that validation.","tokens_in":10568,"tokens_out":6141,"duration_ms":72923,"concrete_test":"After training on the Shepp-Logan or carved-cheese experiment, freeze the learned parameters and compute the exact measurement covariance entries L_ij of Eq. (4) for a random subset of about 200 measurement pairs using numerical double integration with the trained u(x); compare these with the approximate Phi*Lambda*Phi^T entries of Eq. (12). Also report the basis count m and the maximum observed |u(x)|/Lk. If the relative error exceeds roughly 10%, or if max |u(x)|/Lk is near 1, the truncated expansion is not validated for trained networks and the claims require an error bound or an adaptive domain/re-training procedure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that Eq. (9) is a faithful approximation of the squared-exponential covariance after the network has been trained. This is required for the entries in Eqs. (12)-(13) to approximate the true double integrals of Eq. (4), and hence for the predictive expressions (7) to be valid. The approximation has two uncontrolled parts. First, the Dirichlet basis is fixed on a domain whose size Lk is chosen by the alpha=5 heuristic in Supplementary 8.1 using the initial lengthscale; as the hyperparameters, especially the lengthscale lk, are optimized, the spectral mass of S(cj) moves, and the fixed frequency grid [pi/(2Lk), m-tilde*pi/(2Lk)] may no longer cover it. The paper explicitly states that Lk is not optimized and calls the effect \"negligible\" without supporting evidence. Second, u(x) is a trained neural network, and nothing prevents u(x) from approaching or leaving the Dirichlet boundary during training; the eigenfunctions vanish on the boundary, which would inject spurious variance collapse and boundary artifacts. Supplementary 8.1 requires only that Lk be clearly larger than max |u_k(x)|, but no mechanism enforces this and no post-training check is reported. Since all numerical results, including the claimed linear-in-N complexity, are for this approximate model, a failure of the approximation would mean the paper demonstrates an efficient solver for a different GP than the one intended, not for the proposed deep kernel model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for deep kernel learning when observations are line integrals of a latent function, as in computed tomography. It approximates the stationary covariance kernel on the network-warped input with a truncated Hilbert-space basis expansion (Eq. 9), turning the double integrals in the Gram matrix (Eq. 4) into single integrals (Eq. 13) and matrix operations that scale linearly in the number of measurements for fixed basis size. It also introduces a pre-training step (Eq. 14) in which the neural network is fit to the standard GP posterior mean before joint training. Experiments are presented for a one-dimensional step function and for simulated and real CT data (Shepp-Logan phantom and carved cheese), with qualitative comparisons against filtered backprojection.","tokens_in":10900,"tokens_out":4382,"duration_ms":48386,"significance":"If the basis expansion remains accurate after training, the core reduction is a useful and nontrivial contribution: it is the first deep kernel learning formulation for integral observations and offers a route to non-stationary GP reconstruction with uncertainty in limited-data CT. The derivation of Eq. (13) is clean, the matrix-inversion-lemma reformulation is standard, and the authors are candid about limitations such as restricting to a single latent output and the heuristic choice of the domain size. However, the paper does not provide code, quantitative error metrics, uncertainty-calibration checks, or validation of the approximation error, so the significance is conditional on the additional evidence requested below.","major_comments":[{"comment":"The central reduction is valid only if the truncated basis expansion accurately represents the squared-exponential covariance on the range of u(x) after training, but this is asserted rather than verified. The domain size L_k is fixed before training using the alpha=5 heuristic with the initial lengthscale, while the network and hyperparameters are optimized; nothing prevents u(x) from approaching the Dirichlet boundary or the spectral mass from moving outside the frequency grid [pi/(2L_k), m-tilde*pi/(2L_k)]. The statement at the end of Section 8.1 that the effect of not optimizing L_k is \"negligible\" is unsupported. Please provide either an error bound for the approximation in Eq. (9) over the set of u values visited during training, or a post-training validation reporting max|u_k(x)|, the optimized lengthscales, and the resulting spectral coverage for each experiment. Without this, Eqs. (12)-(13) are not shown to approximate Eq. (4) for the model actually trained.","section":"Supplementary 8.1 and Eqs. (9)-(13)"},{"comment":"The empirical claims are supported only by visual inspection. For the Shepp-Logan phantom the ground truth is available, so the paper should report quantitative reconstruction errors (e.g., RMSE, SSIM, or normalized MSE) for the proposed method, FBP, and a standard GP baseline, with variability over noise realizations; for the real cheese data it should report a quantitative comparison against the dense reconstruction. In addition, on simulated data the paper should report the empirical coverage of the claimed 95% credibility regions, since the uncertainty estimates are one of the advertised outputs. Without these numbers the claims of improved performance over FBP and the standard GP cannot be assessed.","section":"Section 4"},{"comment":"Pre-training is described as having a \"crucial impact\" on convergence and final quality, but no ablation is provided. Please report results with (i) random initialization, (ii) the proposed pre-training step, and ideally (iii) a fixed u(x) equal to the standard GP mean ft(x), to establish whether the gains come from the deep kernel learning itself or from initialization near the standard GP solution. This also addresses the mild circularity that the network is trained to reproduce the standard GP mean, which the authors themselves note may contain artifacts for discontinuous functions.","section":"Algorithm 1 and Eq. (14)"},{"comment":"The headline scalability claim of linear scaling with the data size N is not demonstrated empirically. The derivation of Eq. (13) gives linear scaling for building Phi for fixed m, but m = m-tilde^{D_u} grows exponentially with the latent dimension, as acknowledged in Section 3.2, and no wall-clock times or scaling experiment with varying N are reported. Please include a short experiment measuring computation time per optimization step or per prediction as a function of N (and ideally D_u) to substantiate the claim that the method is practically feasible beyond one-dimensional problems.","section":"Section 3.1 and Section 4"}],"minor_comments":[{"comment":"In the paragraph after Figure 5, \"FPB\" should be \"FBP\" to match the abbreviation used elsewhere.","section":"Section 4.2.1"},{"comment":"Reference [1] spells the society as \"Finish Inverse Problems Society\"; the correct spelling is \"Finnish\", and the dataset citation [1, 6] should give the full technical-report information in the text.","section":"References and Section 4.2.2"},{"comment":"The equation \"alpha l_k^{-1} = max_k c_{kj}\" uses k both as a dimension index and as a summation index in Eq. (11); please clarify the notation.","section":"Supplementary 8.1"},{"comment":"The sentence \"As stated, this approach does not directly generalise to the case D_u > 1\" is unclear, since Eq. (14) could in principle be applied to each output component separately; please clarify what is meant by non-generalization.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and cites the relevant literature; the requested revisions concern validation and experimental rigor rather than a fundamental flaw in the derivation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper has a real idea and a real computational contribution, but it undersells its own assumptions and over-relies on pictures. I think it deserves referee time, not because it is complete, but because the core mechanism is clean and the gaps are fixable.\n\nWhat is actually new: the paper adapts deep kernel learning to line-integral measurements, and the trick of expanding the warped kernel in a Hilbert space basis (Eq. 9) to turn double integrals into single integrals (Eq. 13) is a genuine computational reduction. The derivation is sound, and the pre-training scheme (Eq. 14) is a sensible way to initialize the network when the cost landscape is bad. The method is described in enough detail to reimplement, and the related work section is competent.\n\nWhere the soft spots are: the truncated basis approximation is load-bearing but under-validated. The paper chooses the domain size Lk using the alpha=5 heuristic from the initial lengthscale, then fixes it while the network and hyperparameters are trained. The paper says the effect is \"negligible\" but offers no evidence. Since the neural network can push u(x) into regions where the fixed frequency grid no longer covers the spectral mass, the finite-basis approximation could silently become poor. No post-training check or error bound is reported. This is a real hole, not a fatal one: the approximation is standard practice, but with a warped input it deserves more care.\n\nThe empirical support is also weaker than the claims. The CT results are visual only: no quantitative reconstruction error, no comparison to a standard GP on the same CT data, no reporting of the number of basis functions, and no runtime or convergence details. The blurriness is acknowledged and blamed on the network architecture, but without ablation it is hard to know whether the method is doing what the paper says. The toy example does compare against a standard GP and looks better, but still visually.\n\nOne mild self-reference: pre-training the network against the standard GP reconstruction of the same data is a heuristic initialization, not a circular prediction. It is fine, but it should be described as such.\n\nOverall, the math is honest and the contribution is real. This paper is for people working on limited-data CT, strain tomography, or any integral-observation GP problem who need non-stationary behavior with uncertainty. A serious referee could push the authors to validate the basis approximation after training, add quantitative metrics, and include a standard GP baseline in the CT experiments. I would recommend accepting it conditionally rather than desk-rejecting.","headline":"A genuinely useful extension of deep kernel learning to line-integral data, with a real computational trick at its core, but the evidence is visual and the key approximation is not stress-tested.","tokens_in":11393,"tokens_out":1827,"would_cite":false,"duration_ms":19794,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A finite sine-basis expansion makes deep kernel learning tractable for line-integral measurements such as computed tomography.","keywords":["deep kernel learning","Gaussian process","line integral measurements","computed tomography","non-stationary covariance","Hilbert space approximation","neural network pre-training","limited-data reconstruction"],"falsifier":"On a trained model, compute the exact double-integral covariance entry (4) by numerical integration and compare it with the basis approximation (12) for many pairs of measurement lines, including lines crossing sharp edges; if the relative error becomes comparable to the noise level after training, the central feasibility claim fails.","tokens_in":10361,"feed_emoji":"🩻","tokens_out":9681,"duration_ms":89460,"temperature":0.7,"pith_summary":"This paper sets out to make deep kernel learning work when the data are line integrals of the target function, as in X-ray computed tomography. The authors argue that a finite Hilbert space basis expansion of the covariance function removes the main obstacle: naive implementations need numerical double integrals whose number grows quadratically with data, while the expansion reduces each entry to a single integral and makes total cost linear in data size. They add a practical pre-training step in which the neural-network input warp is fitted to a standard Gaussian-process reconstruction before joint training, which they find to be crucial for convergence. If the method holds, limited-data CT-style problems gain non-stationary GP reconstructions with explicit uncertainty, and related integral-measurement problems such as strain-field tomography become natural next targets.","feed_headline":"Basis functions cut CT deep kernel costs to single integrals","feed_subtitle":"A Hilbert-space expansion plus pre-training lets Gaussian processes handle X-ray line data with uncertainty.","key_machinery":"The load-bearing object is the finite Hilbert space basis expansion (9), in which a stationary covariance is written as a weighted sum of products of sine eigenfunctions on a rectangular domain, with weights given by the spectral density. Because this factorizes the kernel in its two inputs, each line-integral covariance entry becomes a product of single line integrals of sines (13), which is what reduces the numerical work from double to single integrals. The second piece of machinery is the pre-training objective (14), which fits the network warp to the mean of a standard GP reconstruction before joint training, providing an informed starting point.","core_discovery":"The paper claims that deep kernel learning—a Gaussian process with a neural-network-warped input space—is feasible for line-integral measurements if the covariance is replaced by a truncated Hilbert space basis expansion. This turns the covariance matrix between measurements into $L\\approx \\Phi\\Lambda\\Phi^T$, so the challenging double integrals reduce to single integrals and the computation scales linearly with the number of measurements. The paper further claims that pre-training the network warp against a standard GP reconstruction, before joint training of all parameters, substantially improves convergence and the final reconstruction. Demonstrated on one-dimensional step-function inference and on limited-data computed tomography, the method yields non-stationary GP reconstructions with tighter credibility regions than the standard stationary GP and cleaner region separation than filtered back projection.","pith_inferences":["Because the factorization of the basis product is what does the work, the same linear-scaling reduction should extend to other linear functionals—weighted integrals, derivatives, boundary fluxes—provided they act separably on the sine basis; only the single-integral integrand changes.","The domain-size choice used for the basis is a heuristic without an error bound; a post-training check comparing the exact double-integral covariance with the basis approximation at the input locations the warp visits would test whether the approximation survives training.","Pre-training by regressing the warp onto a GP reconstruction is a generic warm-start for deep kernel learning and may help point-measurement problems, where poor local minima are also a known difficulty."],"forward_implications":["Deep kernel learning becomes computationally feasible for line-integral observations: the double-integral bottleneck is replaced by single integrals, and the per-iteration cost scales linearly instead of quadratically in the number of measurements.","Limited-data CT reconstructions can be produced with credibility regions, so uncertainty information is available where filtered back projection gives only a single image.","Pre-training the warp on a standard GP reconstruction makes joint training robust enough for two-dimensional problems; without it, convergence depends strongly on initialization.","The same line-integral machinery transfers to other settings where measurements are integrals of a field, such as strain-field estimation, which the paper identifies as a natural extension."],"supporting_citations":[{"why":"It supplies the Hilbert-space reduced-rank approximation that rewrites the kernel as a sum of spectral-density-weighted basis products, reducing the double integrals to single integrals.","marker":"[47]"},{"why":"It defines the deep kernel learning framework of neural-network-warped kernels that this paper extends to integral measurements.","marker":"[54]"},{"why":"It presents the manifold Gaussian process formulation of input warping, the other root of the proposed model.","marker":"[7]"},{"why":"It provides the Gaussian-process and linear-functional background, including the predictive equations used throughout.","marker":"[34]"},{"why":"It supplies the Shepp-Logan phantom used as ground truth in the simulated computed tomography experiment.","marker":"[42]"},{"why":"It provides the filtered-back-projection routine used as the comparison baseline for the CT reconstructions.","marker":"[49]"},{"why":"It supplies the real carved-cheese X-ray CT dataset used in the demonstration.","marker":"[1]"},{"why":"It documents the carved-cheese dataset and its reference reconstruction, grounding the real-data comparison.","marker":"[6]"}],"fun_headline_variants":["Deep kernel learning cuts CT integral costs to single","Basis expansion makes deep kernel GP practical for CT","Neural GP handles line integrals via single integrals","Pre-trained deep kernel GP improves CT reconstruction","Gaussian process with network warp tackles X-ray line data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The finite sine-basis approximation, with a domain size chosen by a rule of thumb rather than tuned, stays accurate for the input locations the trained neural network actually produces.","fun_headline_variants_meta":{"raw":{"variants":["Deep kernel learning cuts CT integral costs to single","Basis expansion makes deep kernel GP practical for CT","Neural GP handles line integrals via single integrals","Pre-trained deep kernel GP improves CT reconstruction","Gaussian process with network warp tackles X-ray line data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000495,"raw_usage":{"total_tokens":2311,"prompt_tokens":709,"completion_tokens":1602,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":325,"completion_tokens_details":{"reasoning_tokens":1529}},"tokens_in":325,"tokens_out":1602,"duration_ms":11023,"temperature":1.0,"reasoning_tokens":1529,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:05:47.998457+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a trained model, compute the exact double-integral covariance entry (4) by numerical integration and compare it with the basis approximation (12) for many pairs of measurement lines, including lines crossing sharp edges; if the relative error becomes comparable to the noise level after training, the central feasibility claim fails.","supporting_citations":[{"cited_title":"Wilson, Zhiting Hu, Ruslan Salakhutdinov, and Eric P","cited_arxiv_id":null,"evidence_quote":"It defines the deep kernel learning framework of neural-network-warped kernels that this paper extends to integral measurements."},{"cited_title":"Rasmussen, and Marc P","cited_arxiv_id":null,"evidence_quote":"It presents the manifold Gaussian process formulation of input warping, the other root of the proposed model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the Gaussian-process and linear-functional background, including the predictive equations used throughout."},{"cited_title":"Shepp and Benjamin F","cited_arxiv_id":null,"evidence_quote":"It supplies the Shepp-Logan phantom used as ground truth in the simulated computed tomography experiment."},{"cited_title":"Schönberger, Juan Nunez-Iglesias, François Boulogne, Joshua D","cited_arxiv_id":null,"evidence_quote":"It provides the filtered-back-projection routine used as the comparison baseline for the CT reconstructions."},{"cited_title":"Accessed: 2019-04-20","cited_arxiv_id":null,"evidence_quote":"It supplies the real carved-cheese X-ray CT dataset used in the demonstration."}],"review_version":1}