{"id":"758be354-31d5-4d6f-b42d-c23898c703ee","arxiv_id":"2506.07188","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"Feature-level reverse propagation can reconstruct intermediate features from labels via linear least squares and yield post-training accuracy comparable to backpropagation fine-tuning.","lead":"This paper proposes a post-training method that reconstructs intermediate neural network features from class labels by solving linear equations, then uses those reconstructed features to fine-tune selected layers while keeping others frozen. The authors claim superior generalization and efficiency over standard backpropagation-based fine-tuning, but their experiments show mostly comparable accuracy and higher memory use.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unquantified replacement of G[a*_l] by G[hat a_l] in Eq. 17 undermines the claim of a rigorous well-posed convolutional reconstruction, and the reported experiments do not substantiate the headline performance advantage.","rationale":"The reader's weakest-assumption analysis identifies exactly the step I also consider most load-bearing: the acknowledged replacement of G[a*_l] by G[hat a_l] in Section 3.2. The paper presents this as a minor engineering convenience, but it changes the optimization problem being solved: the exact constraint in Eq. 17 is linear in the unknown feature, and substituting the forward feature removes the dependence on the unknown without quantifying the resulting constraint violation. Since the reconstruction pipeline is greedy, any per-layer error can propagate; the paper's own Section 4.1 attributes worse performance at early layers to accumulated information loss, which is consistent with this concern rather than a cure for it. The experiments also fail to support the abstract's comparative claims: FR-PT is within noise of BP-PT in most configurations, worse on MNIST, and uses more GPU memory in every reported case, with no wall-clock training-time measurements to support 'computational efficiency.' The omission of target-propagation work additionally inflates the novelty claim, though that is secondary to the correctness issue. Because the central mathematical claim and the headline empirical claim are both left unsupported, the reader's REJECT verdict is appropriate; my read does not change it.","tokens_in":23132,"tokens_out":4972,"duration_ms":61241,"concrete_test":"Write a small script that, for a trained 3-layer CNN with 8->16->32 channels and 3x3 kernels, forms the exact convolution as a sparse Toeplitz/im2col matrix and solves Eq. 4/7 directly. Compare this ground-truth reconstruction with the FFT-based solution from Eqs. 19-20. Measure the relative L2 error in a*_l and the constraint residual ||conv(a*_approx)-z*||/||z*|| for each layer, and compute cumulative drift over 3-5 greedy layers. If residual grows with depth or exceeds 1e-2, the approximation is not negligible; if it remains at machine precision, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 derives Eq. 17 with the boundary correction G^x_{n,m}[a*_l], then states that in real computing G^x_{n,m}[a*_l] is replaced by G^x_{n,m}[hat a_l], making Eq. 17 'approximately correct.' This is the load-bearing step: a*_l is the unknown being solved for, so the constraint matrix in Eqs. 18-20 is evaluated at the forward feature rather than at the reconstructed feature. No error bound, Lipschitz estimate, or residual certificate is provided. The FFT-based solution therefore minimizes a modified objective and need not satisfy the actual network computation z*_{l+1} = conv(a*_l, K) + b_l. Because reconstruction is greedy and layer-by-layer, the error in G can accumulate, so the final target a*_lR may deviate systematically from any feature the frozen layers can actually produce. This casts doubt on the central claim that the reverse computation is 'rigorously reformulated' as well-posed linear systems, and it also weakens the validity of L_rec as a supervisory signal. The full-rank assumption on W_l is likewise asserted without validation for convolutional layers, where the frequency-domain matrix can be rank-deficient. The reported gains over BP-PT are small, within noise in many configurations, and absent on MNIST, while GPU memory is higher in every configuration; these tables do not independently rescue the theoretical concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical and decoupled post-training framework (FR-PT) for pretrained CNNs. Starting from a label, it first maps the label to a target output vector via an optimal embedding (L1 maximum assignment or L2 nearest embedding), then reconstructs intermediate feature maps backward through frozen layers by solving linear systems or least-squares problems for linear and convolutional operations, with pooling and nonlinear activations handled by heuristic reverse rules. The reconstructed feature at layer l_R is used together with the classification loss to post-train the module between l_S and l_R. The authors compare FR-PT with BP-based post-training on six image classification benchmarks and report accuracy and GPU memory, along with ablations of the embedding method and an analysis across training stages.","tokens_in":23397,"tokens_out":7352,"duration_ms":82081,"significance":"If the reconstruction step were genuinely well-posed and exact, the framework would be an interesting alternative to end-to-end backpropagation for modular post-training, and the feature-deviation visualization could contribute to interpretability. The paper contains some correct components: the L1 optimal-embedding proof, the L2 nearest-embedding derivation via KKT conditions, the linear-layer equality-constrained QP and least-squares formulations, and a reasonably large set of experiments. However, the central claim of a rigorous reformulation is undermined by the acknowledged approximation in the convolutional reconstruction and by the heuristic reverse operations for pooling and ReLU. Moreover, the experimental results do not consistently support the stated superiority in generalization or computational efficiency, since FR-PT is often within one standard deviation of BP-PT, worse on MNIST, and uses more GPU memory in every reported configuration. The contribution is therefore not established as presented.","major_comments":[{"comment":"The load-bearing approximation is stated in the text immediately after Eq. (17): in real computing, G^x_{n,m}[a*_l] is replaced by G^x_{n,m}[\\hat a_l], making Eq. (17) only approximately correct. The unknown a*_l appears inside the boundary-correction term of the very constraint used to solve for a*_l, yet no Lipschitz estimate, error bound, or residual certificate is provided for this replacement. Because the reconstruction is greedy and layer-by-layer, the error can accumulate, so the final reconstructed feature a*_{l_R} may systematically deviate from any feature that the frozen subnetwork can actually produce. This directly contradicts the abstract's claim that feature-level reverse computation is 'rigorously reformulated' as well-posed linear systems. The authors should at minimum report consistency residuals ||z*_{l+1} - (conv(a*_l, K) + b_l)|| on the actual data, and should provide either a provable bound or an iterative correction scheme.","section":"Section 3.2, Eq. (17)"},{"comment":"The manuscript assumes without validation that 'the linear weight matrix W^l is always full-rank' and relies on this for uniqueness in Eq. (6) and for the block systems in Eqs. (19) and (20). The assumption is not checked for the actual networks. For linear layers, the required condition is full row rank in the case n_l >= n_{l+1} and full column rank in the case n_l < n_{l+1}; for convolutional layers, the frequency-domain matrix F(\\tilde K) can be rank-deficient (for example with symmetric kernels or certain channel counts), in which case the KKT matrix is singular and the claimed unique solution does not exist. The authors should report numerical ranks of the relevant matrices for the architectures used, or restrict the method to provably full-rank cases.","section":"Section 3, full-rank assumption and Eqs. (6), (19), (20)"},{"comment":"The experimental evidence does not support the abstract's claims of 'superior generalization performance and computational efficiency.' In many configurations the difference between FR-PT and BP-PT is within one standard deviation; on MNIST (Table 1) FR-PT is consistently below BP-PT; and GPU memory usage is higher for FR-PT in every listed configuration, with no wall-clock time reported. The paper should provide paired significance tests across the 10 runs, report a proper efficiency metric (e.g., end-to-end time and peak memory), and either demonstrate consistent statistically significant gains or substantially weaken the superiority claims.","section":"Tables 1-6"},{"comment":"The reverse computation for nonlinear activations and pooling is heuristic: ReLU is inverted by the identity map, and pooling layers are reconstructed by directly copying values. These steps are not posed as the well-posed optimization problems that the paper claims as its main contribution, and their approximation error is not analyzed. Since these heuristic inverses are used in the same greedy backward chain as the linear and convolutional reconstructions, the overall reconstructed feature is not the solution of a single well-posed problem. The authors should either incorporate these operations into the optimization framework with explicit error control or explicitly limit the rigor claim to the linear and convolutional subproblems.","section":"Section 3.4"}],"minor_comments":[{"comment":"In the Lagrangian definition, the term '\\hat a_l W^l' appears to be a typo for 'W^l a_l', and the multiplier term should be written with a transpose; the subsequent KKT matrix in Eq. (6) is consistent with the corrected form.","section":"Section 3.1, Eq. (6) and surrounding text"},{"comment":"The notation F(\\tilde K) and F(a_l) in Eqs. (19) and (20) is ambiguous because these objects are not defined for multi-channel kernels and feature maps; the authors should define the matrix with entries F(\\tilde K[n,m]) and the column vector with entries F(a_l[x,m]) at each frequency (u,v).","section":"Section 3.2, Eqs. (12)-(19)"},{"comment":"The symbol 'T b,n_m' in Eq. (14) is not defined; it should presumably be T^x_{n,m} after zero-padding.","section":"Section 3.2, Eq. (14)"},{"comment":"The loss coefficient alpha is set per dataset with values 0.1, 0.3, or 0.7, but no sensitivity analysis or selection criterion is provided; the reported conclusions depend on this free parameter.","section":"Tables 1-6"},{"comment":"The claim that this is 'the first method to formalize feature-level reverse computation as well-posed optimization problems' is presented without discussing prior reverse-computation and synthetic-target methods in the related work; the wording should be tempered or supported by a comparative discussion.","section":"Section 1, Contributions"}],"recommendation":"reject","confidential_remarks":"The paper's main theoretical claim rests on an approximation that is acknowledged but unquantified, and the experimental results do not rescue it. I would be open to reconsidering a thoroughly revised version that provides an error analysis or an iterative fix for the G-term approximation, validates the rank assumptions, and reports paired significance tests together with a genuine efficiency comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline is this: the paper contains some correct and clean derivations, but the abstract overstates both the novelty and the results. The best parts are the optimal-embedding results (the L1 maximum-assignment proof and the L2 nearest-embedding closed form) and the linear-layer reconstruction, where the KKT system and least-squares formulation are correct. That is real formal work.\n\nWhat is actually new is the specific pipeline: post-train a module of a frozen pretrained CNN by reconstructing an output vector from the label, then inverting the frozen layers one by one to get a feature target, then training the module against that target plus the classification loss. Target propagation and deconv nets have done pieces of this, and the paper does not cite the target-propagation literature, so the \"first to formalize\" claim is inflated. But the packaged combination, and the feature-deviation visualization, is a reasonable proposal for decoupled post-training. The targets are self-generated from the network itself, which is inherent to any target-propagation method and is not the main problem.\n\nThe weak point is the convolutional reconstruction. The derivation of Eq. 17 needs G evaluated at the unknown a*_l; the authors then substitute G evaluated at the forward feature hat a_l and call it \"approximately correct.\" That is a real gap. There is no error bound, no Lipschitz estimate, no residual check, and because reconstruction is greedy, the error can accumulate. That directly contradicts the \"rigorously reformulated\" claim for the conv case. The full-rank assumption is also unvalidated, and the reverse computations for pooling (copy) and ReLU (identity) are crude, though honestly stated.\n\nThe experiments do not rescue the theory. On the six small benchmarks, FR-PT is usually within noise of BP-PT, loses on MNIST, and uses more GPU memory in every configuration. No wall-clock timing is reported, so \"computational efficiency\" is unsubstantiated. On the deeper modules of CIFAR-100 and Tiny ImageNet there are some consistent gains, but those are on very weak baselines and the BP-PT numbers degrade noticeably at later epochs.\n\nWho should read it: anyone working on target propagation or decoupled learning will find the nearest-embedding and linear-inversion math a useful building block, even if they do not adopt the full pipeline. The paper deserves a serious referee; it is not a desk reject. But it needs major revision: quantify the G approximation or drop the rigorous claim, add timing comparisons, include a target-propagation baseline, and tone down the novelty and performance claims.","headline":"A target-propagation-style post-training scheme with a few clean closed-form derivations, but the rigorous claim rests on an unquantified approximation and the experiments do not support the advertised gains.","tokens_in":23911,"tokens_out":3820,"would_cite":false,"duration_ms":42514,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","90C20","90C25","65T50"],"pacs":[],"model":"deepseek-v4-flash","headline":"A pretrained network can be post-trained by solving for the feature maps its frozen layers would need in order to classify correctly, replacing end-to-end backpropagation with systems of linear equations and least-squares problems.","keywords":["hierarchical decoupled post-training","feature map reconstruction","feature backpropagation","optimal embedding","explainable AI","FFT-based convolution reconstruction","image classification","autonomous driving"],"falsifier":"Reconstruct the feature maps for a trained network, then feed each reconstructed $a^*_l$ forward through the frozen layer it was solved against and compare with the target $z^*_{l+1}$: an exact MDP solution must match to numerical precision, so a residual that grows with depth directly measures the boundary-correction error. A reader could then run post-training twice, once with $G^x_{n,m}[\\hat a_l]$ and once with the boundary term solved to a fixed point, and check whether the accuracy gap between FR-PT and backpropagation-based post-training survives the change.","tokens_in":22903,"feed_emoji":"🧩","tokens_out":12242,"duration_ms":120946,"temperature":0.7,"pith_summary":"This paper claims that a pretrained neural network can be improved module by module without end-to-end backpropagation: first compute, from the top layer downward, what the intermediate feature maps would have to be for the frozen upper layers to classify each image correctly, then train the lower modules to reproduce those reconstructed features. The reverse computation is formalized as convex optimization — a system of linear equations when the layer's feature dimension shrinks in the forward direction (the underdetermined case) and a least-squares problem when it grows (the overdetermined case) — with convolutions solved frequency-wise in the Fourier domain and the scalar label mapped to an output vector by a provably optimal embedding. If the claim holds, any contiguous block of a pretrained network can be post-trained in isolation using label-derived feature targets, and the deviation between forward and reconstructed features becomes a measurable per-layer diagnostic. The experiments on six image classification benchmarks support the claim for blocks near the output, where the reconstructed targets are reliable, and the paper attributes its weaker early-layer performance to accumulated reverse-computation error.","feed_headline":"Post-train CNNs by solving linear systems, not backprop","feed_subtitle":"Labels become feature targets so any single module can train alone, beating backprop on converged networks.","key_machinery":"Two principles carry the argument. The minimal deviation principle (MDP) applies when the reconstruction equation is underdetermined: among the infinitely many features that would reproduce the next layer's target exactly, choose the one closest to the original forward feature under the L2 norm; this is a strictly convex quadratic program whose KKT conditions collapse into the linear system of Eq. 6. The computing consistency principle (CCP) applies when the equation is overdetermined: choose the feature that minimizes the reconstruction residual, a linear least-squares problem (Eq. 7). For convolutional layers the same dichotomy is solved in the Fourier domain, where flipping the kernel and zero-padding turn the network convolution into a pointwise product, so each spatial frequency becomes an independent small linear system built from FFTs — this is how the paper avoids dense linear algebra on full feature maps. The optimal embedding closes the loop at the top: the scalar label is replaced by the projection of the forward output vector onto the set of vectors whose argmax is the label, with the L1 norm solved by maximum assignment (provably optimal) and the L2 norm by a KKT active-set algorithm (nearest embedding).","core_discovery":"The central claim is that feature-level reverse computation — recovering the activations that would turn a desired label into the network's output — can be posed as well-posed optimization problems rather than heuristic inversion. At each linear layer the forward map $W_l a_l + b_l = z_{l+1}$ is either underdetermined (when $n_l \\ge n_{l+1}$) or overdetermined (when $n_l < n_{l+1}$). In the underdetermined case the paper selects the feature minimizing $\\|a_l - \\hat a_l\\|_2$ subject to exact reconstruction — the minimal deviation principle, whose KKT conditions collapse into a linear system; in the overdetermined case it selects the feature minimizing the reconstruction residual $\\|W_l a_l + b_l - z^*_{l+1}\\|_2$ — the computing consistency principle, a linear least-squares problem. For convolutional layers the same dichotomy is solved frequency-wise using the FFT and the convolution theorem, with a boundary-correction term evaluated at the forward feature rather than the unknown reconstructed one. At the output layer a scalar label is embedded as the closest vector, under the L1 or L2 norm, whose argmax equals the label — the maximum-assignment solution for L1 and a KKT active-set nearest-embedding solution for L2. Modules between layers $l_S$ and $l_R$ are then trained by ordinary backpropagation on a weighted sum of the classification loss and the deviation between forward and reconstructed features at layer $l_R$.","pith_inferences":["A likely transfer path the paper leaves untested: if the boundary-correction approximation in the Fourier-domain reconstruction can be made exact or bounded, the same MDP/CCP dichotomy should extend to ResNets and Transformers, where attention and normalization would be the genuinely hard operations because they have no natural \"closest feature consistent with the next layer\" analogue.","The greedy layer-by-layer scheme is the most plausible source of the early-layer accuracy loss; solving for an entire block of features jointly through one least-squares problem in the Fourier domain would directly test whether greedy choice, rather than the well-posed formulation, is the bottleneck.","A consequence the paper leaves implicit: the reconstruction pipeline doubles as a reachability test — if the optimal embedding demands an output vector far from the forward output, or if the deep reconstructed feature drifts outside what the frozen layers can produce, the network is being asked to represent a label its current trunk cannot express, and the deviation itself is a usable failure sign"],"forward_implications":["Any contiguous block of a pretrained CNN can be post-trained in isolation: the label-derived reconstructed feature at the block's output supplies the supervision, so the rest of the network stays frozen during the entire update.","The reverse pass is deterministic and convex, so surrogate feature targets for an entire dataset can be precomputed once and reused across repeated post-training runs without re-entering the network.","On converged baselines, a single epoch of FR-PT can surpass ten epochs of backpropagation-based post-training, indicating that once a network has converged, the gains come from correcting the prediction process rather than from further iterative optimization.","Feature reconstruction succeeds where channel counts decrease in the forward direction (the exactly consistent MDP regime) and degrades where channels increase (the approximate CCP regime) or pooling discards information, which is why the experimental sweet spot is post-training blocks near the output.","The absolute difference between forward and reconstructed feature maps shrinks as a network's accuracy grows, giving a per-layer, label-informed visualization of training maturity."],"supporting_citations":[{"why":"Supplies the convex optimization machinery — duality, Lagrange multipliers, KKT conditions, Slater's condition — that turns the minimal deviation and computing consistency principles into solvable linear systems and least-squares problems.","marker":"[5]"},{"why":"Introduces decoupled neural interfaces with synthetic gradients, the line of work this paper extends by replacing gradient-level decoupling with label-derived feature targets.","marker":"[9]"},{"why":"Provides decoupled greedy learning of CNNs, the greedy layer-wise training scheme whose layer-by-layer reconstruction strategy mirrors this paper's reverse pass.","marker":"[3]"},{"why":"Presents fully decoupled training with delayed gradients, the independent-module-update baseline that FR-PT is designed to improve upon in post-training settings.","marker":"[40]"}],"fun_headline_variants":["Train CNNs by solving linear systems, not backprop","Replace backprop with feature-level least squares","Hierarchical post-training: reconstruct features, solve linear equations","Feature reverse propagation: linear systems replace backprop","Decoupled training via feature reconstruction and linear systems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"In Section 3.2 the Fourier-domain reconstruction replaces the boundary-correction term $G^x_{n,m}[a^*_l]$ — which should be evaluated at the unknown reconstructed feature — with $G^x_{n,m}[\\hat a_l]$, evaluated at the known forward feature, and the paper calls the resulting equation 'approximately correct' without bounding the error; because the reconstruction proceeds greedily layer by layer, that error is free to accumulate.","fun_headline_variants_meta":{"raw":{"variants":["Train CNNs by solving linear systems, not backprop","Replace backprop with feature-level least squares","Hierarchical post-training: reconstruct features, solve linear equations","Feature reverse propagation: linear systems replace backprop","Decoupled training via feature reconstruction and linear systems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1398,"prompt_tokens":1035,"completion_tokens":363,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":303}},"tokens_in":651,"tokens_out":363,"duration_ms":4510,"temperature":1.0,"reasoning_tokens":303,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:40:09.509176+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reconstruct the feature maps for a trained network, then feed each reconstructed $a^*_l$ forward through the frozen layer it was solved against and compare with the target $z^*_{l+1}$: an exact MDP solution must match to numerical precision, so a residual that grows with depth directly measures the boundary-correction error. A reader could then run post-training twice, once with $G^x_{n,m}[\\hat a_l]$ and once with the boundary term solved to a fixed point, and check whether the accuracy gap between FR-PT and backpropagation-based post-training survives the change.","supporting_citations":[{"cited_title":"Cambridge University Press, 2004","cited_arxiv_id":null,"evidence_quote":"Supplies the convex optimization machinery — duality, Lagrange multipliers, KKT conditions, Slater's condition — that turns the minimal deviation and computing consistency principles into solvable linear systems and least-squares problems."},{"cited_title":"Decoupled neural interfaces using synthetic gradients, 2017","cited_arxiv_id":null,"evidence_quote":"Introduces decoupled neural interfaces with synthetic gradients, the line of work this paper extends by replacing gradient-level decoupling with label-derived feature targets."},{"cited_title":"Decoupled greedy learning of cnns","cited_arxiv_id":null,"evidence_quote":"Provides decoupled greedy learning of CNNs, the greedy layer-wise training scheme whose layer-by-layer reconstruction strategy mirrors this paper's reverse pass."},{"cited_title":"Fully decoupled neural network learning using delayed gradients.IEEE transactions on neural networks and learning systems, 33(10):6013–6020, 2021","cited_arxiv_id":null,"evidence_quote":"Presents fully decoupled training with delayed gradients, the independent-module-update baseline that FR-PT is designed to improve upon in post-training settings."}],"review_version":1}