{"id":"cfdf9dcc-3c88-4284-81c2-ba85b022096b","arxiv_id":"2501.02740","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A stochastic configuration CNN with reinforcement-learning kernel pruning classifies four fused magnesium furnace working conditions at 92.57% accuracy, but the proof of convergence and the interpretability advantage are not well supported.","lead":"This paper builds a convolutional neural network for fused magnesium furnace images by adding Gaussian-shaped kernels one by one with a stochastic configuration rule, then prunes kernels with reinforcement learning. It reports 92.57% test accuracy on 12,000 furnace images and proposes class-activation visualizations, but the convergence proof and the interpretability comparisons are not quantitatively solid.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The convergence proof's contraction factor can exceed 1 and Eq. (22) misstates the projection step, so the claimed global error-convergence guarantee is not established as written.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing weakness: the contraction premise in Eqs. (8)/(9) and Eq. (19) does not guarantee error decrease given the stated parameter ranges, and Eq. (22) contains a clear algebraic misstatement of the projection step. My independent reading of Section III.B.3 confirms both problems. This is not a manufactured concern: the proof is presented as a contribution ('ensuring the convergence of recognition errors in a hierarchical manner') and is used to justify the method's interpretability and its advantage over backpropagation-trained CNNs. If the proof fails, the theoretical foundation collapses, even though the reported 92.57% test accuracy could still be a valid empirical result. I considered whether the absence of quantitative IoU numbers for the interpretability claim is more load-bearing; it is a serious reporting gap, but it is fixable by adding measurements, whereas the proof error indicates the formal argument as written is incorrect. I also noted the parameter-count inconsistency between Table 1 (48.85 MB) and Table 2/Table 3 (12.854 MB) for the same 'proposed method'; this is secondary but reinforces the need for reproducible artifacts. The reader's REJECT verdict is therefore appropriate: the central theoretical guarantee is not established, and the empirical and interpretability claims are not backed by released code or the paper's own defined IoU metric. No change to the reader's verdict is needed.","tokens_in":18120,"tokens_out":3942,"duration_ms":37249,"concrete_test":"Re-derive Eq. (22) from Eqs. (20)–(21) using the standard orthogonal projection identity: for optimal scalar weight \\bar{o} = <e, a>/||a||^2, one has ||e - \\bar{o} a||^2 = ||e||^2 - <e,a>^2/||a||^2 = ||e||^2 - \\bar{o}<e,a>. Check whether the paper's expression ||e||^2 - Σ_q \\bar{O}_{1,q} is equivalent; it is not. Then evaluate the contraction factor (ξ+r)u_C at the stated parameter ranges, e.g., ξ=0.5, r=0.8, C=1, giving 1.3 > 1. If either check fails, the convergence proof is invalid and the claim that recognition errors converge hierarchically is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is that DCSCNs ensure hierarchical convergence of recognition errors without backpropagation (Abstract; Section III.B; Contribution 1). The proof rests on Eq. (19), which asserts ||e_l_Cl||^2 ≤ (ξ+r) u_Cl ||e_l_Cl-1||^2, with u_C = 1/C, ξ ∈ [0.5, 5], r ∈ [0.8, 1.5]. Since ξ+r ≥ 1.3 and u_1 = 1, the factor (ξ+r)u_C exceeds 1 for small C, so Eq. (19) and the chain in Eq. (23) do not force errors to shrink. Even if Eq. (19) is taken from reference [25], the present paper must establish it or cite it correctly; as written, the inequality is compatible with error growth. The proof also contains a concrete algebraic error in Eq. (22): it writes ||e - \\bar{O} A||^2 = ||e||^2 - Σ_q \\bar{O}_{1,q} ≤ ||e||^2, which is dimensionally inconsistent and not the orthogonal projection identity. The correct identity involves subtracting a positive multiple of ||A||^2 (or an inner-product term), not subtracting the output weight itself. Because this step is the bridge from layer l to layer l+1, the claimed monotone decrease of the error does not follow. This is an internal correctness flaw, not a disagreement with community consensus: the empirical accuracy may survive, but the paper's formal convergence guarantee and its stated advantage over backpropagation-based CNNs are unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes deep convolutional stochastic configuration networks (DCSCNs) for recognizing working conditions of fused magnesium furnaces from furnace images. The method generates Gaussian differential convolution kernels through a supervised stochastic configuration mechanism, constructs a deep convolutional network incrementally without backpropagation, claims a global convergence proof for the resulting errors, and defines a class-activation-map-based interpretability index (IoU) together with a DDPG-based reinforcement-learning pruning procedure driven by a joint reward of accuracy, interpretability, and parameter count. Experiments on a 12,000-image industrial dataset report a test accuracy of 92.57%, outperforming several comparison methods, with qualitative CAM visualizations claimed to locate the relevant furnace regions. The central theoretical claim of hierarchical global convergence, however, is not established as written: the contraction argument in Section III.B.3 contains an invalid inequality and an algebraic error, and the interpretability metric defined in Section III.C is never reported quantitatively.","tokens_in":18527,"tokens_out":8410,"duration_ms":76418,"significance":"If the convergence proof were correct and the interpretability index were properly evaluated, the paper would offer a useful contribution to industrial AI by combining a backpropagation-free randomized construction of deep CNNs with RL-based pruning and visual explainability. The empirical accuracy improvement on a real fused-magnesium-furnace dataset is credible and potentially valuable to practitioners. However, the paper's stated theoretical contribution is load-bearing and is not merely incomplete; the contraction inequality used to prove monotone convergence is numerically false under the reported parameter ranges, and the projection step in Eq. (22) is algebraically wrong. In addition, the interpretability evaluation is qualitative only, and the reported parameter counts are inconsistent across tables. The strengths are the clear application framing, the use of physically meaningful Gaussian-differential kernels, and the reproducible experimental setup; the weaknesses are in the formal guarantees and the quantitative validation of interpretability, which are central to the paper's claims.","major_comments":[{"comment":"The asserted inequality ||e^l_{C_l}||^2 ≤ (ξ+r) u_{C_l} ||e^l_{C_l-1}||^2 is not a contraction under the stated parameter ranges. With ξ∈[0.5,5], r∈[0.8,1.5], and u_C=1/C, the factor (ξ+r)u_C is at least 1.3 for C=1 and can be as large as 6.5, so the right-hand side can exceed the left-hand side even when the error has not improved. The chain in Eq. (23) therefore does not imply monotone decrease of the recognition error. Since Contribution 1 explicitly claims a proof of global convergence, the paper cannot discharge this step by citing [25]; it must either prove Eq. (19) under conditions that make the factor strictly less than 1, or restrict the parameter ranges accordingly.","section":"Section III.B.3, Eq. (19)"},{"comment":"The equality ||e^l_{C_l} − \\bar{O}^{l+1}_1 A^{l+1}_1||^2 = ||e^l_{C_l}||^2 − Σ_{q=1}^m \\bar{O}^{l+1}_{1,q} is algebraically wrong. The correct orthogonal projection identity gives ||e − \\bar{O} A||^2 = ||e||^2 − Σ_q ⟨e_q, A_q⟩^2 / ||A_q||^2, not subtraction of the output weight \\bar{O} itself. The written equality is dimensionally inconsistent, and the inequality ≤ ||e^l_{C_l}||^2 does not follow from it. This step is the bridge from layer l to layer l+1, so the claimed monotone convergence is not established.","section":"Section III.B.3, Eq. (22)"},{"comment":"The IoU-based interpretable trustworthiness index is defined in Eqs. (28)-(29) and used as a component of the RL reward in Eq. (30), but no numerical IoU values are reported anywhere in the experimental section. The interpretability claim (“accurately locate the target region of interest”) is supported only by qualitative CAM images. In addition, computing IoU requires a procedure for extracting the highlighted region d_i from the CAM (e.g., a thresholding rule), which is not specified; without that procedure the metric is not well-defined. Because IoU_val is also fed into the reward, a separate held-out evaluation of IoU would be needed to avoid circularity; the paper does not provide one.","section":"Section III.C and Section IV.C.2"},{"comment":"The “Proposed method” row in Table I reports P_A = 48.85 MB, while Tables II and III report P_A = 12.854 MB for the same method and the same test accuracy 92.57%. Section IV.C.1 attributes 94.53%/92.78% to the 8-layer model, which matches the “Without adaptive kernel pruning mechanisms” row of Table I rather than the “Proposed method” row. These discrepancies are material to the model-compression claim; the authors should state which configuration (before or after RL pruning, and with which kernel assignment) is used in each table, and ensure the numbers are consistent.","section":"Section IV.D, Tables I-III"},{"comment":"The kernel selection conditions are internally inconsistent. Eq. (8) uses p u_C b^2 ||e||^2, Eq. (9) uses ξ u_C b^2 ||e||^2, but the convergence score in Eq. (17) uses (ξ+r)u_C ||e||^2 with no b^2 and with a different constant structure. The paper never defines how p, b, ξ, r, and u_C jointly guarantee the inequality in Eq. (19). Without a precise link between the selection criterion and the asserted contraction inequality, the hierarchical convergence claim is not supported.","section":"Section III.B.1 and III.B.2"}],"minor_comments":[{"comment":"The caption of Fig. 11 contains the typo “convolutinal”, and the conclusion contains “practical species for fused magnesium furnace working condition recognition”, which should read “practical support for ...”. These should be corrected.","section":"Section IV.C.2 and Section V"},{"comment":"The symbol γ is used both for the DDPG discount factor in Eq. (32) and for the learning rate in the sentence “the learning rate γ is 0.005”, while Eqs. (34)-(35) use τ for the soft-update rate. Distinct symbols should be used for these three quantities.","section":"Section IV.C.3 and Eqs. (31)-(35)"},{"comment":"The notation C is overloaded: C indexes convolutional kernels (C=1,...,C_l), C_l denotes the number of kernels in layer l, and C_{l-1} denotes the number of input channels. In Eq. (26) the feature map dimensions are written as H×W×C_l, while the text also uses C_l for the number of channels. Please disambiguate the two uses.","section":"Section III.B"},{"comment":"The proof claims lim_{l→+∞} ||F − F^l|| = 0, but the construction stops at finite L_max and C_l,max. The asymptotic statement should be replaced by a finite-sample error bound consistent with the stopping criterion, or the precise sense of the convergence claim should be stated.","section":"Section III.B.1"}],"recommendation":"reject","confidential_remarks":"The paper shares substantial material with the authors' prior publication [25], which is cited as the source of Eq. (19); the relationship between this submission and [25] should be clarified, particularly whether the convergence claim and the interpretability index are genuinely new contributions of the present manuscript. Given the algebraic error in Eq. (22), the non-contractive factor in Eq. (19), the missing IoU results, and the inconsistent parameter counts across Tables I-III, the central claims are not currently supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the genuinely new piece here is the RL-based adaptive pruning with the joint reward in Eq. (30). The DCSCN construction, the Gaussian differential kernels, the channel independence scores, and the CAM were all in the same group's [25]. So the paper is an incremental but meaningful application of that machinery to a real industrial monitoring problem, with pruning as the new contribution. Credit where due: they use real furnace images, run ablations, compare against a reasonable set of baselines, and the 92.57% test accuracy is plausible.\n\nThe soft spots are real, and the reader's main concern lands. Eq. (22) is algebraically wrong: ||e - \\bar{O}A||^2 is not ||e||^2 minus the sum of output weights; the correct projection identity would subtract a term involving the norm of A. That breaks the claimed monotone error decrease. Eq. (19) is also not a contraction as written: with ξ∈[0.5,5], r∈[0.8,1.5], and u_1=1, the factor (ξ+r)u_C exceeds 1 for small C, so the chain in Eq. (23) doesn't force convergence. The 'global convergence' claim is therefore unsupported.\n\nTwo more issues. The interpretable trustworthiness index IoU is defined in Eq. (28), but the paper never reports IoU numbers on test data. Instead, IoU_val is inserted into the RL reward in Eq. (30), so the interpretability advantage is partially circular and never independently measured. Also, the parameter count for the proposed method is 48.85 MB in Table 1 but 12.854 MB in Table 2. That's a large inconsistency that needs an explanation.\n\nNo code or data is released, so the empirical claim is not independently checkable.\n\nMy take: this is a plausible engineering result with a broken theoretical wrapper. It deserves a serious referee because the empirical question is real and the pruning module is a reasonable extension, but as written the paper overclaims. I'd recommend sending it to review with the expectation of major revision: fix or remove the convergence proof, report the IoU numbers on a held-out set, reconcile the parameter counts, and release the code/data. I would not cite it in its current form.","headline":"A plausible industrial monitoring paper whose new RL-pruning module is interesting, but the convergence proof is broken and key reported numbers are inconsistent.","tokens_in":19035,"tokens_out":3489,"would_cite":false,"duration_ms":31948,"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 stochastic-configuration deep network recognizes fused magnesium furnace working conditions from images with 92.57% test accuracy and heatmap-based explanations, without backpropagation.","keywords":["fused magnesium furnace","working condition recognition","deep convolutional stochastic configuration networks","Gaussian differential convolution kernel","class activation mapping","reinforcement learning pruning","interpretability","industrial image classification"],"falsifier":"Record the residual norm after each accepted kernel during the incremental DCSCN construction on the reported 12,000-image dataset; if any accepted kernel violates the paper's contraction inequality or the per-layer error does not decrease monotonically under the stated parameter ranges, the claimed convergence guarantee is refuted. A behavioral counterpart: test the pruned 8-layer model on a fresh labeled set of furnace videos from another plant and measure both accuracy and class activation map IoU against expert annotations.","tokens_in":17934,"feed_emoji":"🔥","tokens_out":8322,"duration_ms":76319,"temperature":0.7,"pith_summary":"The paper proposes a way to recognize the working condition of a fused magnesium furnace—normal, underburn, overheating, or abnormal exhaust—from furnace-mouth images, and claims it is both more accurate and more explainable than existing deep-learning approaches. The central idea is to build a deep convolutional network without backpropagation: convolutional kernels are generated one by one as Gaussian-difference filters, and a supervised rule accepts only kernels that reduce the current recognition error. The authors report that the resulting model reaches 92.57% test accuracy, that its class activation maps highlight the correct region of the furnace, and that a reinforcement-learning step prunes redundant kernels to make the model compact. If this holds, furnace operators could trust an automatically generated visual explanation rather than relying only on manual inspection.","feed_headline":"Furnace-state recognizer hits 92.57% and shows where it looks","feed_subtitle":"Classifies fused magnesium furnace images, explains verdicts with heatmaps, and prunes itself via reinforcement learning.","key_machinery":"The load-bearing object is the Gaussian differential convolution kernel, defined as $$\\psi(x,y)=\\frac{1}{2\\pi}\\left($e^{{-\\frac{x^2+y^2}}${2\\$xi^{2}$}}-\\frac{1}{r}$e^{{-\\frac{x^2+y^2}}${$2r^{2}$\\$xi^{2}$}}\\right),$$ a center-surround filter that sharpens edges and texture at the furnace mouth; $\\xi$ controls filter width and $r$ controls edge scale. The construction mechanism is a supervised configuration loop: candidate kernels are drawn uniformly over $\\xi\\in[0.5,5]$ and $r\\in[0.8,1.5]$, and a candidate is kept only if its convergence score is positive, which is meant to guarantee a contraction of the residual error. The network grows incrementally—one kernel at a time, then one layer at a time—with output weights fitted by least squares, avoiding backpropagation. Interpretability comes from the channel feature independence coefficient $FC^l_\\varrho$, which weights feature maps in a class activation map $L_q$; trustworthiness is scored by the IoU of the highlighted region with expert annotation. Finally, a DDPG reinforcement learner chooses per-layer pruning ratios under a joint reward $R = \\mathrm{ACC}_{\\mathrm{val}} + \\mathrm{IoU}_{\\mathrm{val}} - \\beta P_A$.","core_discovery":"On the paper's own terms, the discovery is that stochastic configuration networks—randomized learners whose hidden parameters are selected by a supervisory criterion—can be extended to deep convolutional networks for industrial image recognition. The paper argues that Gaussian differential convolution kernels, generated randomly within a data-correlated range and accepted only when they satisfy a residual-based inequality, let a multi-layer convolutional network converge in error layer by layer without gradient descent. The same construction yields feature maps whose channel independence scores can be combined with class scores into class activation maps, giving an interpretable heatmap of where the network looks. The paper further claims that pruning kernels with a DDPG reinforcement learner, guided by a reward combining accuracy, interpretability trustworthiness (IoU against expert annotations), and parameter count, improves focus and keeps accuracy high. Tested on 12,000 augmented images of four furnace conditions, the method's 92.57% test accuracy is claimed to beat CNNs, Bayesian networks, Swin Transformer, CNN+LSTM, SCNs, 2DSCNs, and DeepSCNs.","pith_inferences":["Our reading: the convergence proof is the part to test first. With $\\xi\\in[0.5,5]$ and $r\\in[0.8,1.5]$, $\\xi+r$ is often larger than 1, so the contraction factor in the paper's inequality need not be below 1; the claimed monotone decrease of the residual thus depends on a premise the stated parameter ranges do not guarantee.","The paper reports single-run recognition results; a repeated-seed or cross-factory evaluation would tell whether the 2–4 point accuracy gap over CNN, Swin Transformer, and CNN+LSTM is stable rather than an artifact of one split.","The interpretability claim could be stress-tested by having operators identify the anomaly region from the class activation map alone; the IoU metric already approximates this, but a human study would connect the quantitative score to real usability.","Because the Gaussian-difference kernel family is fixed in advance, the method's \"physical meaning\" is inherited from the filter design; one could ask whether allowing the kernel family to be learned while keeping the stochastic configuration rule would retain interpretability and improve accuracy on larger datasets."],"forward_implications":["If the central claim is right, furnace operators can receive both a class label and a heatmap showing which part of the furnace mouth drove the decision, making each alarm checkable by a human.","Because convolutional kernels are configured incrementally rather than trained by backpropagation, the method avoids learning-rate sensitivity and weight-initialization dependence that complicate conventional CNN training on small industrial datasets.","The RL pruning step selects kernel subsets that preserve accuracy and interpretability while cutting parameters and training time, so the deployed model is smaller and faster than the unpruned network.","The same supervised kernel-configuration strategy can build deep convolutional recognizers for other process industries where labeled abnormal samples are scarce and explainable decisions are required."],"supporting_citations":[{"why":"supplies the stochastic configuration network framework, including the supervisory random-parameter selection and universal approximation property that the kernel generation strategy extends.","marker":"[29]"},{"why":"supplies the deep configuration architecture where hidden nodes in each layer connect to the output, which the DCSCN inherits for hierarchical error reduction.","marker":"[18]"},{"why":"supplies the prior DCSCN construction for fused magnesium furnaces, including the Gaussian differential kernel definition and the inequality that the convergence proof assumes.","marker":"[25]"},{"why":"supplies the DDPG reinforcement-learning algorithm used to adaptively prune convolutional kernels.","marker":"[33]"},{"why":"supplies a 2D convolutional stochastic configuration network baseline that the proposed method is compared against and outperforms.","marker":"[34]"},{"why":"supplies the convolutional neural network baseline used in the performance comparison.","marker":"[35]"}],"fun_headline_variants":["AI furnace grader shows its work with heatmaps, 92.57%","No backprop: stochastic convolutional net interprets furnace, 92.57%","RL-pruned net reads furnace images, shows why, 92.57%","Furnace AI: stochastic conv nets give heatmap verdicts, 92.57%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergence guarantee assumes that every newly accepted Gaussian-difference kernel shrinks the training error by a fixed factor; with the paper's own ranges $\\xi\\in[0.5,5]$ and $r\\in[0.8,1.5]$, that factor is not always forced below one, so the error-decrease premise can fail.","fun_headline_variants_meta":{"raw":{"variants":["AI furnace grader shows its work with heatmaps, 92.57%","No backprop: stochastic convolutional net interprets furnace, 92.57%","RL-pruned net reads furnace images, shows why, 92.57%","Furnace AI: stochastic conv nets give heatmap verdicts, 92.57%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000472,"raw_usage":{"total_tokens":2351,"prompt_tokens":953,"completion_tokens":1398,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":1310}},"tokens_in":569,"tokens_out":1398,"duration_ms":9222,"temperature":1.0,"reasoning_tokens":1310,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:06:02.201877+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the residual norm after each accepted kernel during the incremental DCSCN construction on the reported 12,000-image dataset; if any accepted kernel violates the paper's contraction inequality or the per-layer error does not decrease monotonically under the stated parameter ranges, the claimed convergence guarantee is refuted. A behavioral counterpart: test the pruned 8-layer model on a fresh labeled set of furnace videos from another plant and measure both accuracy and class activation map IoU against expert annotations.","supporting_citations":[{"cited_title":"Stochastic configuration networks: Fundamentals and algorithms,","cited_arxiv_id":null,"evidence_quote":"supplies the stochastic configuration network framework, including the supervisory random-parameter selection and universal approximation property that the kernel generation strategy extends."},{"cited_title":"Deep stochastic configuration networks with uni- versal approximation property,","cited_arxiv_id":null,"evidence_quote":"supplies the deep configuration architecture where hidden nodes in each layer connect to the output, which the DCSCN inherits for hierarchical error reduction."},{"cited_title":"Research on working condition recognition method of fmf based on deep convolutional stochastic configuration network,","cited_arxiv_id":null,"evidence_quote":"supplies the prior DCSCN construction for fused magnesium furnaces, including the Gaussian differential kernel definition and the inequality that the convergence proof assumes."},{"cited_title":"Human-level control through deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"supplies the DDPG reinforcement-learning algorithm used to adaptively prune convolutional kernels."},{"cited_title":"2D convolutional stochastic configuration net- works,","cited_arxiv_id":null,"evidence_quote":"supplies a 2D convolutional stochastic configuration network baseline that the proposed method is compared against and outperforms."}],"review_version":1}