{"id":"341897bd-4b8d-437c-853c-3ad06030dec0","arxiv_id":"2501.00345","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A linear three-layer neural network with constrained weights provably recovers the edge conductivities of a resistor network from boundary voltage-current data, with the conductivity stored in the second-layer weights.","lead":"The authors build a three-layer neural network where the trained weights, not the output, contain the answer: for a discrete electrical impedance tomography problem on a square grid, the weights in the second layer equal the unknown edge conductivities. The paper proves this for exact data and shows numerically that the approach handles noise better than the standard Curtis-Morrow algorithm, making neural network weights physically interpretable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.1 is internally inconsistent as written: the candidate minimizer (4.1) fails the paper's own boundary-flux equation (4.2) and diagonal constraint (4.3), so the central zero-loss/unique-minimizer claim is unproven until the sign conventions are fixed.","rationale":"The paper's central contribution is Theorem 4.1, so I focused on whether its proof actually establishes that zero-loss minimizers have the stated weights. It does not, as printed. The sign convention in the candidate (4.1) is incompatible with the boundary-flux equation (4.2) and the diagonal constraint (4.3). This is not a matter of consensus or external uniqueness; it is an internal algebraic inconsistency. The constructive existence of a zero-loss minimizer, which the theorem asserts, fails for the displayed weights. The uniqueness half is also affected because the proof uses (4.2) with the same incompatible sign. The likely fix is a sign change in either (2.5), (3.4), or (4.1), and the interpretability claim might survive after such a correction. Until the convention is fixed and the proof re-run, the paper's central claim is unverified as stated. I therefore keep the CONDITIONAL verdict, but for a different reason than the reader's; the Adam-convergence issue remains secondary. The reader's weakest assumption concerned practical convergence to global minimizers, which is a legitimate concern but not the most load-bearing one: the theorem that classifies those minimizers is internally inconsistent as written.","tokens_in":19216,"tokens_out":28801,"duration_ms":299590,"concrete_test":"Run the one-line algebraic check for n = 1: take γ = 1, Dirichlet datum u = e_{(0,1)}, solve (2.3) to get u_p = 1/4, and compute v_{(0,1)} = γ(u_{(0,1)} − u_p) = 3/4. Assemble fW(2) according to (4.1). Forward propagate through (3.5) and evaluate Cα; the boundary term for (0,1) is [ew(u_{(0,1)} − u_p) − v]^2 = [(−1)(3/4) − 3/4]^2 = (3/2)^2 > 0, so the alleged zero-loss minimizer has positive loss. If instead the sign in (2.5), (3.4), or (4.1) is corrected so that this residual vanishes, re-run the proof of Theorem 4.1 with the corrected convention and verify the diagonal constraint (4.3) for interior nodes adjacent to the boundary.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is not self-consistent as printed. Equation (2.5) defines the Neumann data as v_q = γ_{q,p}(u_q − u_p) for q ∈ ∂D, p ∈ N(q), and (4.2) requires ew(2)_{qp}(ex(2)_q − ex(2)_p) = v_q. For a boundary edge, (4.1) sets ew(2)_{qp} = −γ_{q,p}. Substituting gives −γ(u_q − u_p) = γ(u_q − u_p), so the boundary output in (3.4) is y_q = ew(2)_{qp}(u_q − u_p) − v_q = −2v_q ≠ 0. Hence the proposed fW(2) is not a zero-loss minimizer; the constructive existence half of Theorem 4.1 fails for the displayed weights. The same sign error breaks the architectural constraint (ii)/(4.3): for an interior node p adjacent to a boundary node, (4.1) and (4.3) require −Σ_{q∈N(p)} γ_{q,p} = Σ_{q∈N(p), q∈D} γ − Σ_{q∈N(p), q∈∂D} γ, which is false unless all boundary conductances are zero. Thus either (2.5), (3.4), or (4.1) carries the wrong sign; as written, the proof cannot establish that every zero-loss global minimizer has weights (4.1). This is a correctness gap in the theory itself, independent of whether Adam converges.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-layer linear feed-forward network for the discrete inverse conductivity problem on an n×n square lattice. The first layer is trained to map Dirichlet boundary data to the interior solution, while the second-layer weights are constrained to be symmetric and to satisfy the diagonal condition (ii), so that the trained weights can be interpreted as edge conductivities. Theorem 4.1 claims that, with m≥3n noiseless Cauchy pairs whose Dirichlet data contain a basis of the last 3n coordinate directions, every zero-loss global minimizer has second-layer weights equal to the true conductivity profile, with a stated sign convention on boundary edges; Corollary 4.2 claims uniqueness when m≥4n. Section 6 develops a first-order sensitivity analysis of the minimizer under additive noise, and Section 5 reports numerical comparisons with the Curtis-Morrow algorithm for full, incomplete, and partial data, including noisy data. The central theoretical identification and the numerical superiority claim both depend on the zero-loss minimizer analysis and on the assumption that Adam converges to such a minimizer.","tokens_in":19525,"tokens_out":16303,"duration_ms":148912,"significance":"The paper's aim—an interpretable network whose trained weights literally represent the discrete conductivities—is attractive, and the appeal to Curtis-Morrow uniqueness to identify zero-loss minimizers is a sound strategy if the sign convention is repaired. The sensitivity formula in Theorem 6.2 is a useful complement, and the numerical study covers full, incomplete, and partial data with noisy comparisons. These strengths, however, cannot offset the sign inconsistency in the main theorem; the contribution is presently not self-consistent as printed, and the numerical claims rest on an unverified training assumption. If the sign issue is corrected and the optimization caveat is made explicit, the paper would make a worthwhile contribution to interpretable network-based inversion.","major_comments":[{"comment":"Equations (2.5), (3.4), (4.1), and (4.2) are mutually inconsistent. With v_q defined by (2.5) as γ_{q,p}(u_q-u_p) and the boundary output in (3.4) as y_r = ew(2)_{pr}(ex(2)_r - ex(2)_p) - bx(2)_r, formula (4.1) sets ew(2)_{pr} = -γ_{p,r} for a boundary edge, so at the proposed minimizer one obtains y_r = -γ(u_r-u_p) - v_r = -2v_r, not zero. The same contradiction appears in the boundary equation of (4.2), where the claimed solution gives ew(2)_{qp}(u_q-u_p) = -v_q instead of v_q. Moreover, formula (4.1) violates the diagonal constraint (4.3): for an interior node p adjacent to a boundary node, (4.1) gives ew(2)_{pp} = -Σ_{q∈N(p)}γ_{q,p}, whereas (4.3) together with (4.1) would require ew(2)_{pp} = -Σ_{q∈N(p)∩D}γ_{q,p} + Σ_{q∈N(p)∩∂D}γ_{q,p}; equality holds only if all boundary edge conductivities vanish. Thus the constructive existence half of Theorem 4.1 and the subsequent uniqueness argument are not established as printed. The sign convention in at least one of (2.5), (3.4), or (4.1) must be changed, and the boundary diagonal entries in (4.1) must be re-derived consistently with constraint (ii).","section":"Theorem 4.1; Eqs. (2.5), (3.4), (4.1)-(4.3)"},{"comment":"The numerical comparison in Section 5, including the abstract's claim that the method outperforms Curtis-Morrow for noisy data, is conditional on the Adam optimizer reaching a zero-loss global minimizer of the nonconvex loss Cα. The paper provides no convergence guarantee, and Section 5.2.1 itself reports that the loss stagnates at positive values for noisy data and that the method fails to give accurate reconstructions for n=20. Theorem 4.1 only characterizes minimizers after they are attained. To make the numerical claims load-bearing, the authors should report the distribution of final loss values and success rates over random initializations, and either prove or clearly state as an assumption the global convergence of the training procedure; in the present form, the comparison tests the optimizer together with the recovery architecture, not the recovery map characterized by Theorem 4.1.","section":"Section 5.2.1; Figs. 3-5"}],"minor_comments":[{"comment":"The paper uses D both for the interior grid in (2.1) and for the full set including ∂D in the architecture discussion; for example, (3.4) indexes outputs by r∈D even though the boundary nodes ∂D are included. Please introduce a separate symbol, e.g., \\bar D, and use it consistently throughout.","section":"Notation in Sections 2-3"},{"comment":"Weights are indexed by ordered pairs (p,q) although B is a set of unordered edges; specify an orientation convention so that expressions like ew(2)_{pq} and the condition (p,q)∈B are unambiguous.","section":"Eq. (4.1)"},{"comment":"The statement ∇εγ_{T′ε}|_{ε=0}=S†_{T′} should be spelled out as the Jacobian of γ_{T′ε} with respect to the noise vector restricted to T′; as written, the left-hand side has size 2n(n+1)×|T′|, matching the matrix S†_{T′}, but this is not immediately clear from the notation.","section":"Theorem 6.2"},{"comment":"The abstract contains the typo 'post-trainig', and Section 6 appears to contain a dimension typo 'R2n(2n+1)' for the conductivity vector, which should be R^{2n(n+1)}.","section":"Abstract and Section 6"},{"comment":"The proof of local smoothness of the extension map is only a sketch; please state the base case of the induction and explicitly note that invertibility of each B_Qk is an open condition that persists in a sufficiently small neighborhood.","section":"Lemma 6.1"}],"recommendation":"major_revision","confidential_remarks":"The sign inconsistency in Theorem 4.1 is the main obstacle; it appears repairable by changing a sign convention consistently, but until that is done the central theorem cannot be considered proved. I would also ask the authors to temper the numerical superiority claim unless they provide convergence diagnostics and success rates for the Adam training, and to state that the Curtis-Morrow baseline is an unregularized pseudoinverse implementation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the preprint. The core idea is genuinely good: a three-layer linear network where the second-layer weights are forced to be symmetric with zero row sums, so that zero training loss on exact DtN data should force the weights to equal the true edge conductivities. That is a clean way to make post-training weights interpretable, and the Curtis-Morrow uniqueness theorem is the right external tool. The sensitivity analysis (Theorem 6.2 and the Jacobian formula in Proposition 6.1) is also correct as far as I can tell; the Hadamard-product expression for grad_gamma Lambda_gamma and the first-order perturbation argument using S_{T'} full column rank are solid. So there is real substance here.\n\nBut there is a load-bearing sign error in Theorem 4.1. Equation (2.5) defines Neumann data v_q = gamma(u_q - u_p) for boundary node q and interior neighbor p. The output layer (3.4) computes y_q = ew^(2)(u_q - u_p) - v_q, so to get zero loss you need ew^(2) = +gamma on boundary edges. Theorem 4.1 says ew^(2) = -gamma for boundary edges. Substitute that into (4.2) and you get -gamma(u_q - u_p) = gamma(u_q - u_p), which only holds if gamma = 0. The listed candidate minimizer is not a minimizer, and the same sign error makes the diagonal constraint (4.3) fail for nodes adjacent to the boundary. So the proof of Theorem 4.1 collapses as written. It looks fixable - flip the sign convention consistently - but the claim that every zero-loss global minimizer recovers the conductivity is currently unproven.\n\nThe other weaknesses are secondary: the numerical comparison lacks error bars, uses a single conductivity profile, and depends on Adam actually reaching a zero-loss minimum, which the paper does not prove (and admits fails for n=20). The code is promised but not shipped. None of those by themselves would sink the paper, but they make the numbers hard to trust.\n\nVerdict: I'd send it to a serious referee. The idea and the sensitivity analysis deserve refereeing, and the sign issue is the kind of thing a good referee would catch and require fixed. As it stands, I would not rely on Theorem 4.1 or the interpretability framing until the sign conventions are corrected and the proof rechecked. This is a conditional accept requiring major revision, not a desk reject.","headline":"Genuinely good idea and solid sensitivity analysis, but Theorem 4.1 has a sign error that invalidates the central claim as written.","tokens_in":20080,"tokens_out":7096,"would_cite":false,"duration_ms":57450,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N21","35R30","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A three-layer linear network encodes the discrete conductivity profile in its second-layer weights, and the paper proves these weights recover the true edge conductivities exactly when the data are exact and sufficient.","keywords":["discrete inverse conductivity problem","interpretable neural network","Dirichlet-to-Neumann map","resistor networks","global minimizer uniqueness","sensitivity analysis","electrical impedance tomography","conductivity reconstruction"],"falsifier":"Enumerate all zero-loss minimizers of the loss (3.6) for a small grid, say $2 \\times 2$ interior nodes, with the $4n$ standard basis Dirichlet data; if any minimizer has second-layer weights that disagree with the signed edge conductivities of (4.1), then Theorem 4.1 falls.","tokens_in":18968,"feed_emoji":"⚡","tokens_out":13392,"duration_ms":116039,"temperature":0.7,"pith_summary":"This paper tries to establish that a deliberately structured three-layer linear network can solve the discrete inverse conductivity problem by storing the answer in its second-layer weights. The authors construct a network whose output measures the residual of the discrete current-conservation equations, train it on boundary voltage and current pairs, and prove (Theorem 4.1) that when the noiseless data include enough Dirichlet vectors, every global minimizer with zero loss has second-layer weights equal to the true edge conductivities, up to a sign on boundary edges. With a full basis of $4n$ measurements the zero-loss minimizer is unique (Corollary 4.2), so the trained weights are not just a black-box fit but a readable physical quantity. The practical payoff would be a neural-network reconstruction method whose post-training weights are interpretable and that handles noisy full and partial boundary data better than the classical algebraic baseline.","feed_headline":"Trained weights equal true edge conductivities","feed_subtitle":"A provable route to read the conductivity map out of a trained network's weights.","key_machinery":"The load-bearing object is the constrained three-layer linear network itself: an input layer of $8n$ neurons carrying $4n$ boundary voltages and $4n$ measured currents, a hidden layer of $n^2+8n$ neurons, and an output layer of $n^2+4n$ neurons. The learned block of the first layer maps boundary voltages to interior potentials; the second-layer weights are constrained to be symmetric with diagonal entries equal to the negative sum of the row, which makes these weights the parameters in which the conductivity is encoded. The output neurons form the residual between predicted boundary flux and measured Neumann data, together with interior current-conservation residuals, and the loss averages the squared residuals over the training samples. A zero-loss minimizer forces the second-layer weights to define a conductivity whose Dirichlet-to-Neumann matrix agrees with the data on enough columns, and the cited structural result then forces that conductivity to be the true one.","core_discovery":"The central claim is that the unknown conductivity profile $\\gamma$ is exactly recoverable from the weights of the trained network, not from its output. For a square grid with $n \\times n$ interior nodes and $4n$ boundary nodes, the learned first-layer weights $\\widehat{W}^{(1)}$ solve the discrete Dirichlet problem, approximating the discrete Green kernel, while the second-layer weights $\\widehat{W}^{(2)}$ are constrained to be symmetric with diagonal entries equal to the negative row sums. Theorem 4.1 states that for $m \\ge 3n$ noiseless Cauchy pairs whose Dirichlet vectors span the subspace generated by the last $3n$ coordinates, any zero-loss global minimizer of the loss (3.6) has $\\widehat{W}^{(2)}$ given by (4.1): the weight on an interior edge equals the edge conductivity, the weight on a boundary edge equals its negative, non-edge weights are zero, and diagonal entries are signed sums of incident conductivities. Corollary 4.2 strengthens this to uniqueness of the minimizer when $m \\ge 4n$ and the Dirichlet vectors form a basis of $\\mathbb{R}^{4n}$, with the first-layer weights equal to the true Green-kernel matrix. The proof reduces to a structural theorem of the resistor-network literature, Theorem 5.1 of [23], which says that the last $3n$ columns of the Dirichlet-to-Neumann matrix determine the full matrix and therefore the conductivity.","pith_inferences":["One implication the authors leave implicit is that the same architecture could be adapted to other network inverse problems where the forward map is a linear system with symmetric positive parameters, because the proof only needs a finite measurement set that determines the parameter vector.","Theorem 4.1 and Corollary 4.2 classify zero-loss minimizers but say nothing about how optimization reaches them; a natural testable extension is to map the loss landscape for small grids and check whether random initializations consistently land in the identified basin, since the paper's own $n=20$ experiments suggest the practical bottleneck is optimization, not identifiability.","The sensitivity formula could be used beyond the paper as an experimental-design tool: choosing boundary excitations that maximize the smallest singular value of the relevant Jacobian submatrix would select informative measurements before training.","The reported robustness to noise suggests a broader conjecture, not proven in the paper: that an interpretable network with physical constraints can inherit the regularization of overparameterized training without explicit penalty terms."],"forward_implications":["With $4n$ exact measurements forming a basis of the boundary data space, the loss has a unique zero-loss global minimizer, so a successful training run delivers the signed edge conductivities directly from the second-layer weights.","The first-layer weights, trained without any imposed relation to the conductivities, coincide at the unique minimizer with the discrete Green kernel, so the network's internal computation is itself a readable solution map.","Only the last $3n$ columns of the Dirichlet-to-Neumann matrix are needed for all zero-loss minima to share the same conductivity weights, extending the identifiability claim to partial boundary data.","The sensitivity analysis gives an explicit first-order prediction of reconstruction error under small noise through the pseudoinverse $S_{T'}^\\dagger$ of the Jacobian, and the paper reports that this prediction matches the observed errors except in one partial-data configuration."],"supporting_citations":[{"why":"Supplies the existence and uniqueness of the discrete Dirichlet solution, the theorem that the full Dirichlet-to-Neumann map determines the conductivity, and the full column-rank Jacobian result used in the sensitivity analysis.","marker":"[22]"},{"why":"Its Theorem 5.1 shows the last 3n columns of the Dirichlet-to-Neumann matrix determine the full matrix and the conductivities; this is the load-bearing uniqueness input for Theorem 4.1 and the baseline inversion algorithm.","marker":"[23]"},{"why":"Adam is the optimizer used in all numerical experiments, so every reported reconstruction depends on it reaching the minimizers classified by Theorem 4.1.","marker":"[32]"},{"why":"Provides the exponential instability result that motivates the discrete finite-dimensional setting and the comparison of noise robustness.","marker":"[36]"},{"why":"Supplies the depth-dependent sensitivity bounds used to explain why reconstructions are more accurate near the boundary than in the center.","marker":"[26]"}],"fun_headline_variants":["Weights, not outputs, reveal edge conductivities","Conductivity hidden in neural net weights","Read conductivity straight from trained weights","Neural net weights solve inverse conductivity","Interpretable NN: weights equal conductivities"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's success rests on the unproven hope that the training algorithm, started from a random guess, always lands in a perfectly fitting minimum; the paper's own experiments show this can fail for larger grids.","fun_headline_variants_meta":{"raw":{"variants":["Weights, not outputs, reveal edge conductivities","Conductivity hidden in neural net weights","Read conductivity straight from trained weights","Neural net weights solve inverse conductivity","Interpretable NN: weights equal conductivities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1300,"prompt_tokens":973,"completion_tokens":327,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":262}},"tokens_in":589,"tokens_out":327,"duration_ms":3897,"temperature":1.0,"reasoning_tokens":262,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:53:40.436785+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Enumerate all zero-loss minimizers of the loss (3.6) for a small grid, say $2 \\times 2$ interior nodes, with the $4n$ standard basis Dirichlet data; if any minimizer has second-layer weights that disagree with the signed edge conductivities of (4.1), then Theorem 4.1 falls.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the existence and uniqueness of the discrete Dirichlet solution, the theorem that the full Dirichlet-to-Neumann map determines the conductivity, and the full column-rank Jacobian result used in the sensitivity analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Its Theorem 5.1 shows the last 3n columns of the Dirichlet-to-Neumann matrix determine the full matrix and the conductivities; this is the load-bearing uniqueness input for Theorem 4.1 and the baseline inversion algorithm."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Adam is the optimizer used in all numerical experiments, so every reported reconstruction depends on it reaching the minimizers classified by Theorem 4.1."},{"cited_title":"Mandache","cited_arxiv_id":null,"evidence_quote":"Provides the exponential instability result that motivates the discrete finite-dimensional setting and the comparison of noise robustness."},{"cited_title":"Garde and K","cited_arxiv_id":null,"evidence_quote":"Supplies the depth-dependent sensitivity bounds used to explain why reconstructions are more accurate near the boundary than in the center."}],"review_version":1}