{"id":"eb0f0a8a-150f-4a09-add1-18fbdb6a9889","arxiv_id":"2411.13525","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GA-Planes generalizes several volume representations and offers convex, semiconvex, or nonconvex training; 2D analysis shows equivalence to low-rank plus low-resolution matrix approximation.","lead":"This paper introduces GA-Planes, a family of volume models that combine line, plane, and volume feature grids with a decoder, and shows that some variants can be trained with convex optimization. It also proves 2D equivalences to low-rank plus low-resolution matrix factorization and demonstrates competitive results on radiance field reconstruction and 3D/video segmentation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The semiconvex GA-Planes global-optimality guarantee is asserted from Sahiner et al. (2024) without verifying that eq. (6) satisfies the Burer-Monteiro conditions; Theorem 3's 'convex' objective is itself nonconvex because the mask depends on trainable U,V.","rationale":"The paper has genuine and independently checkable contributions: the 2D equivalence theorems for linear decoders (Theorems 1 and 2) are straightforward and correct in structure, the fully convex model in eq. (7) is indeed convex because it is linear in the optimizable grids, and the experiments show that the convex and semiconvex variants train stably on segmentation tasks. The conditional verdict is appropriate because the headline promise of guaranteed global optimality is broader than what is established. My stress-test identifies the semiconvex guarantee as the single most load-bearing claim: it is the part of the central contribution that would distinguish this model from a standard convex linear decoder, and it is asserted only by citing prior work. The statement in Section 6 that BM factorized problems have only global local minima is not a universal property; it requires specific conditions, and the paper neither states nor verifies them for eq. (6). The related inconsistency in Theorem 3, where the mask depends on trainable U and V, further undermines the theoretical support for the convex MLP decoder and strengthens the need for a corrected derivation. None of this is an attack on the authors; it is a request to make the scope of the guarantee precise. If the semiconvex model is instead presented as a heuristic biconvex formulation without the global-optimality guarantee, the paper remains valuable and the conditional verdict can be resolved. My recommended verdict is therefore unchanged: CONDITIONAL, pending either a proof that eq. (6) satisfies the cited Burer-Monteiro conditions or an explicit restriction of the guarantee to the fully convex model.","tokens_in":23790,"tokens_out":12691,"duration_ms":139089,"concrete_test":"Take the semiconvex GA-Planes objective in eq. (6) for a small 2D instance (e.g., 8x8 target, line resolution r1=4, feature dimension d1=3, decoder width h=4) and independently verify whether it can be expressed as a Burer-Monteiro factorization of a convex program satisfying the hypotheses of the Sahiner et al. (2024) theorem cited in Section 6; in parallel, run 10^4 random initializations to convergence and test whether all local minima share the same objective value. If the Sahiner et al. hypotheses fail, or if two distinct local minima differ by more than 1e-6, the semiconvex global-optimality guarantee is not valid for GA-Planes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central guarantee in Section 6 and Appendix A.1 claims that both the convex and semiconvex GA-Planes formulations have the property that all local optima are global, citing Sahiner et al. (2024). For the fully convex model in eq. (7) this is trivially true because the objective is linear in the optimizable grids, but this only certifies global optimality of the frozen-gate surrogate, not the original nonconvex GA-Planes problem. For the semiconvex model in eq. (6), the objective is biconvex in the grid features and decoder weights; the paper does not show that this objective can be written as a Burer-Monteiro factorization of a convex program satisfying the specific conditions under which Sahiner et al. establish the all-local-minima-are-global property. The blanket claim that BM factorized problems have only global local minima is not generally true; it depends on problem structure such as the measurement operator, feasible set, and regularizer, none of which are checked here. Theorem 3, which is the only theoretical support for the convex MLP decoder, is also internally inconsistent: the mask B_i in eq. (12) depends on the trainable factors U and V, so the stated objective is not convex and does not match the frozen-gate model in eq. (7). Thus the load-bearing assumption that the semiconvex variant inherits the global-optimality guarantee is unproven, and if it fails the headline claim must be restricted to the fully convex frozen-gate model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces GA-Planes, a family of volume representations that combine line, plane, and volume feature grids with an MLP decoder. It derives convex and semiconvex training formulations by convexifying the decoder via frozen ReLU gates, proves 2D equivalences to low-rank plus low-resolution matrix completion, states lower bounds on fitting error, and evaluates the models on radiance field reconstruction, 3D segmentation, and video segmentation. The central claim is that GA-Planes is the first class of implicit neural volume representations that can be trained by convex optimization, with guarantees that all local optima are global.","tokens_in":24093,"tokens_out":5508,"duration_ms":56582,"significance":"If the global-optimality claims are correct, the paper would be significant: it would offer a principled way to avoid nonconvex optimization instability in a broad family of volume models, backed by a clean 2D theoretical characterization. Theorems 1 and 2 are correct and provide a useful interpretation of GA-Planes as low-rank plus low-resolution approximation. The empirical study is extensive, spans three tasks, and includes code release. However, the key global-optimality guarantee for the semiconvex formulation is inherited from prior work without verification for this model, and Theorem 3 appears inconsistent with the convex model it claims to describe. The lower bounds in Section 4.2 are also stated without proof. These issues are load-bearing for the paper's headline claims and need to be addressed before the paper can be accepted.","major_comments":[{"comment":"The mask matrix B_i in Eq. (12) is defined as B_i = 1[Σ_j W̄_{i,j} U_j V_j^T ≥ 0], where U and V are optimization variables. The objective is therefore not convex, and it does not match the frozen-gate convex model of Eq. (7), where the gates are evaluated at the initialization of the grids and remain fixed. The proof in Appendix A.2.3 (Eq. (36)) confirms that the masks depend on the trainable grids g1, g2. As stated, Theorem 3 describes a nonconvex masked low-rank problem, and the claimed equivalence to the convex MLP decoder is not established. Please reformulate the theorem to match the frozen-gate model or explicitly state that it applies to a different, nonconvex problem.","section":"Section 4.1, Theorem 3 / Eq. (12)"},{"comment":"The assertion that 'both the convex and semiconvex GA-Planes formulations come with guarantees that all local optima are also global (Sahiner et al., 2024)' is not supported for the semiconvex model in Eq. (6). The objective is biconvex in the grid features f(q) and decoder weights W, and it is not shown that this problem is a Burer-Monteiro factorization of a convex program satisfying the conditions of Sahiner et al. (2024). In particular, the input features f(q) are themselves trainable parameters, unlike the standard BM setting for convex neural networks where the data matrix is fixed. Without verifying these conditions, the global-optimality guarantee is unproven. The paper should either provide a derivation for this specific model or restrict the claim to the fully convex frozen-gate surrogate.","section":"Section 6 / Appendix A.1"},{"comment":"The paper states that subsampling ReLU patterns 'results in a convex program with tractable size, whose solution is one of the stationary points of the original non-convex problem' (citing Pilanci & Ergen 2020), yet the abstract, Section 6, and Appendix A.1 imply global optimality for the convex and semiconvex GA-Planes formulations. This is a mismatch: the convex program is a surrogate, and its global optimum need not be the global optimum of the original nonconvex GA-Planes problem. Please clarify that the global-optimality guarantee applies to the frozen-gate surrogate, not to the original nonconvex problem, and discuss what this implies for the practical value of the convex formulation.","section":"Section 2, last paragraph"},{"comment":"The lower bounds on the Frobenius norm error are asserted without proof. They are load-bearing for the paper's interpretation that MLP decoders increase attainable rank, yet no derivation or reference is supplied in the main text or appendix. Please add proofs or provide a detailed reference for each bound, including the definitions of L* and the conditions under which the bounds are tight or loose.","section":"Section 4.2, Eqs. (14)-(19)"}],"minor_comments":[{"comment":"The captions for Figures 10-17 are essentially identical and do not identify the scene, making them hard to navigate; please add the scene name to each caption.","section":"Appendix A.6, Figures 10-17"},{"comment":"The notation W for the trainable weights and W̄ for the frozen gate weights is easy to confuse; consider using a more distinct symbol for the frozen weights, such as W_gate or W_0.","section":"Section 3.2, Eq. (6)"},{"comment":"The equivalence in Theorem 4 is essentially a reparameterization and the statement 'U* = g1*, V* = g2*, W* = W*, α* = α*' is tautological. Consider phrasing it explicitly as a reparameterization of the optimization variables rather than an equivalence of two distinct optimization problems.","section":"Section 4.1, Theorem 4"},{"comment":"The theorem statements would be clearer if the assumptions on interpolation linearity, feature-grid resolution matching the target, and the Frobenius objective were repeated in each theorem statement or stated as a single set of formal assumptions immediately before Theorem 1.","section":"Section 4.1, Assumptions paragraph"},{"comment":"The phrase 'temporal superresolution of object masks' is used, but the task is essentially interpolation of masks between labeled frames; consider clarifying the relationship to classical video segmentation to avoid overclaiming.","section":"Section 5.3, video segmentation"}],"recommendation":"major_revision","confidential_remarks":"The paper contains a correct and elegant 2D theory for linear decoder models (Theorems 1 and 2) and a broad empirical study, but the headline claims about global optimality are currently overstated. The semiconvex guarantee is asserted without checking the Burer-Monteiro conditions for this specific model, and Theorem 3 is internally inconsistent because its masks depend on the optimization variables. I believe these issues are fixable by careful reframing and additional proofs, so major revision rather than rejection is appropriate. If the authors cannot verify the semiconvex guarantee, they should remove or clearly qualify that claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is showing that a wide family of grid-based volume representations can be trained with a convex program by freezing ReLU gates, and most of the 2D equivalence results are correct. That is a useful application of the Pilanci–Ergen framework to volume modeling, and the paper deserves credit for it.\n\nThe 2D theorems (1, 2, 4) are the strongest part. The equivalences to low-rank plus low-resolution matrix completion are clean and the lower bounds follow from rank arguments. The experiments on segmentation with convex and semiconvex variants show stable training, and the small-model seed study in the appendix gives a concrete demonstration of the convexity benefit. The radiance field experiments are solid but secondary.\n\nThe soft spots are real and load-bearing. The semiconvex model in eq. (6) is biconvex, and the claim that all local minima are global is cited from Sahiner et al. (2024) without checking whether this objective satisfies the Burer–Monteiro conditions. That is not a small omission; the property is not true for biconvex problems in general. Theorem 3 does not match the frozen-gate convex model: the mask B_i in eq. (12) depends on trainable U and V, so the stated problem is nonconvex, and the proof in A.2.3 repeats the same issue. The fully convex model in eq. (7) is a linear decoder over fixed gates, which is honestly not an \"implicit neural representation\" in the usual sense—there is no trainable nonlinear decoder. And the abstract's \"first class... trained by convex optimization\" is true for the convex surrogate, but it does not imply global optimality for the original nonconvex volume fitting problem. The experiments also report no error bars for the main tables, and the low-rank+low-res versus low-rank+sparse comparison rests on a single image.\n\nWho is this for? Researchers working on convex optimization for inverse problems and on volume representations will get real value from the 2D equivalences and the architectural recipe. The flaws are addressable: the authors can fix Theorem 3, either prove the semiconvex guarantee or drop it, and add error bars. I would send this to a serious referee rather than desk-reject, but the referee should insist on those revisions.","headline":"A genuinely useful convexification of volume grids with a sound 2D analysis, but the global-optimality claims are overstated and one theorem describes a different model than the one being trained.","tokens_in":24717,"tokens_out":2833,"would_cite":true,"duration_ms":30740,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C25","15A23","68T45"],"pacs":[],"model":"deepseek-v4-flash","headline":"GA-Planes introduces a family of implicit neural volume models—features on line, plane, and volume grids decoded by an MLP—that can be trained by convex optimization, with convex and semiconvex formulations whose local optima are global.","keywords":["implicit neural representations","convex optimization","volume rendering","tensor factorization","low-rank matrix completion","radiance fields","3D segmentation","Burer-Monteiro factorization"],"falsifier":"One concrete check is to fit a small 2D image with the convex GA-Planes model while exhaustively enumerating all possible ReLU activation patterns for that problem, then compare that true global optimum with the optimum found by the paper's randomly subsampled convex program; if the two objectives diverge substantially, the surrogate optimum is not the original problem's optimum. For the semiconvex model, one can run many random initializations of the frozen gating grids on a fixed segmentation task and test whether every run converges to the same objective value; a spread of final values would contradict the claim that all local optima are global.","tokens_in":23505,"feed_emoji":"🧊","tokens_out":8304,"duration_ms":79344,"temperature":0.7,"pith_summary":"This paper introduces GA-Planes, a family of volume models in which features interpolated from line, plane, and volume grids are combined and decoded by an MLP. The central claim is that any member of this family that avoids multiplying features together can be trained by convex optimization whenever the fitting objective is convex, making GA-Planes the first implicit neural volume representation with a convex training route. For the convex and semiconvex formulations, the paper claims all local optima are global, removing sensitivity to initialization. In 2D the paper proves GA-Planes is equivalent to low-rank plus low-resolution matrix approximation and shows this beats the classic low-rank plus sparse decomposition on a natural image; in 3D it demonstrates the model on radiance fields, 3D segmentation, and video segmentation. If the claim holds, volume fitting for many inverse problems becomes initialization-insensitive and globally solvable.","feed_headline":"GA-Planes: implicit volumes trained by convex optimization","feed_subtitle":"The GA-Planes family adds global-optimality guarantees to volume reconstruction and segmentation.","key_machinery":"The load-bearing structure is the GA-Planes parameterization itself—interpolated line, plane, and volume feature grids combined by concatenation, addition, or elementwise multiplication and decoded by an MLP—together with the convexification recipe borrowed from convex neural network theory: freeze a random sample of ReLU activation patterns as fixed gates, then optimize the remaining weights. The semiconvex variant is a Burer-Monteiro factorization of a convex objective, in which every local minimum is claimed to be global. The 2D analysis converts these models into low-rank plus low-resolution matrix completion, which is the mechanism that yields the rank bounds, the equivalence theorems, and the interpretation of GA-Planes as first fitting a low-resolution volume and then completing the high-frequency residual.","core_discovery":"On its own terms, the paper's discovery is that implicit neural volume fitting does not have to be nonconvex. GA-Planes parameterizes a volume by interpolating features from 1D line grids, 2D plane grids, and a 3D volume grid, combining them by concatenation, addition, or geometric-algebra multiplication, and decoding with an MLP. When features are concatenated rather than multiplied, the decoder's ReLU gates can be frozen at random initialization so that the entire training objective becomes a convex program; the semiconvex variant keeps the decoder weights separate but factorizes the convex objective in the Burer-Monteiro style, and the fully convex variant fuses decoder weights into the feature grids. The paper proves the 2D versions are exactly low-rank, or low-rank plus low-resolution, matrix completion, with MLP decoders raising the attainable rank from the feature dimension to the grid resolution. Experiments then show nonconvex GA-Planes matches or exceeds established radiance-field baselines across model sizes, and that the convex and semiconvex versions match nonconvex performance on segmentation tasks where a standard tri-plane baseline degrades sharply under convex training.","pith_inferences":["The author does not pursue it, but the same frozen-gate convexification could apply to 2D and other implicit neural representations beyond volumes, since the 2D analysis already casts GA-Planes as matrix approximation.","The low-rank plus low-resolution view suggests an allocation rule the paper only demonstrates by hand: spend parameter budget on high-resolution line grids for high-frequency content and low-resolution volume grids for smooth background; this rule could be tested as an automatic architecture search.","Because the semiconvex guarantee rests on an unproved condition, a numerical check—random restarts on an ill-conditioned segmentation task—would reveal whether non-global local minima actually occur in practice.","The image-fitting comparison hints that low-rank plus low-resolution may generally beat low-rank plus sparse for natural imagery, a claim broader than this paper's single-image experiment and worth testing on other image classes and medical volumes."],"forward_implications":["Volume segmentation and other convex fitting tasks can be trained with convex or semiconvex GA-Planes and are guaranteed to reach a global optimum regardless of initialization.","The 2D analysis identifies GA-Planes as low-rank plus low-resolution matrix approximation, so design choices of grid resolution and feature dimension directly control attainable rank and error floors.","GA-Planes generalizes existing volume representations such as voxels, tensor factorizations, tri-plane, K-Planes, and mixture-of-primitives models, so the convexification applies to those architectures whenever they avoid feature multiplication.","Nonconvex GA-Planes with feature multiplication retains competitive radiance-field quality at small model sizes, extending the size, expressiveness, and optimizability tradeoff frontier.","The convex and semiconvex formulations are compatible with off-the-shelf convex solvers, not only first-order gradient training."],"supporting_citations":[{"why":"Supplies the convexification recipe: enumerate or sample ReLU activation patterns and train a convex program whose solution is a stationary point of the original network.","marker":"Pilanci & Ergen (2020)"},{"why":"Supplies the Burer-Monteiro factorization framework and the result that every local minimum of the factored convex problem is global, on which the semiconvex GA-Planes guarantee rests.","marker":"Sahiner et al. (2024)"},{"why":"Extends convexification to deep networks via path regularization, used to justify convexity of deeper decoders.","marker":"Ergen & Pilanci (2024)"},{"why":"Provides the volume rendering formulation and Blender dataset used for the radiance-field experiments.","marker":"Mildenhall et al. (2020)"},{"why":"TensoRF is a baseline the nonconvex GA-Planes radiance model is compared against and is a special case of the GA-Planes family.","marker":"Chen et al. (2022)"},{"why":"K-Planes is a baseline and a special case whose plane-feature multiplication motivates the nonconvex GA-Planes variant.","marker":"Fridovich-Keil et al. (2023)"},{"why":"Tri-Plane is the baseline used for convex, semiconvex, and nonconvex segmentation comparisons.","marker":"Chan et al. (2022)"},{"why":"Supplies Space Carving, the supervision used to create 3D segmentation labels.","marker":"Kutulakos & Seitz (1999)"}],"fun_headline_variants":["GA-Planes: convex training for implicit neural volumes","Implicit volumes break free from nonconvex training","Convex optimization enters implicit volume fitting","GA-Planes: global optimality for volume reconstruction","First implicit volume model trainable convexly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that randomly freezing a sample of ReLU activation patterns yields a convex program whose global optimum stays close to the true optimum of the original GA-Planes problem, and, for the semiconvex variant, that the factored model meets the theoretical condition that makes every local minimum globally optimal—a condition the paper asserts but does not prove.","fun_headline_variants_meta":{"raw":{"variants":["GA-Planes: convex training for implicit neural volumes","Implicit volumes break free from nonconvex training","Convex optimization enters implicit volume fitting","GA-Planes: global optimality for volume reconstruction","First implicit volume model trainable convexly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000686,"raw_usage":{"total_tokens":3134,"prompt_tokens":993,"completion_tokens":2141,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":2070}},"tokens_in":609,"tokens_out":2141,"duration_ms":16098,"temperature":1.0,"reasoning_tokens":2070,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:19:50.330192+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One concrete check is to fit a small 2D image with the convex GA-Planes model while exhaustively enumerating all possible ReLU activation patterns for that problem, then compare that true global optimum with the optimum found by the paper's randomly subsampled convex program; if the two objectives diverge substantially, the surrogate optimum is not the original problem's optimum. For the semiconvex model, one can run many random initializations of the frozen gating grids on a fixed segmentation task and test whether every run converges to the same objective value; a spread of final values would contradict the claim that all local optima are global.","supporting_citations":[],"review_version":1}