{"id":"f5a0b8c8-967e-40f0-aaa8-d6332edd7868","arxiv_id":"2505.05659","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Vector-valued EfficientNets with a new vector-valued depthwise convolution report 99.46% mean accuracy on ALL-IDB2 with about 70% fewer parameters than EfficientNetV2-B0.","lead":"The authors build vector-valued versions of EfficientNet, in which arithmetic on ordinary numbers is replaced by arithmetic on 4-dimensional hypercomplex numbers, cutting parameters by about 70% and reporting 99.46% mean accuracy on a 260-image leukemia benchmark.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing real-valued HSV control confounds input encoding with architecture; the claim that V-EfficientNets outperform EfficientNet is not yet established.","rationale":"The reader's weakest assumption identifies exactly the confound I would flag, so my agreement is 'agree'. The paper has a real technical contribution: it gives a careful construction of vector-valued depthwise convolution via Kronecker products, introduces a vectorization factor lambda, and releases source code. Those contributions do not depend on the empirical headline. What does not yet stand is the claim that vectorization is what makes V-EfficientNets outperform EfficientNetV2-B0, because Table II varies input encoding together with algebra for all strong results and omits a real-valued HSV control. The central claim would collapse if HSV alone explains the gain. This concern is not a rejection of the architecture; it is a missing control that can be supplied in a straightforward experiment. I therefore keep the reader's CONDITIONAL verdict rather than moving to acceptance or rejection. The recommended disposition is unchanged: the state-of-the-art claim should be accepted only after the color-encoding confound is resolved and significance is checked on paired runs.","tokens_in":12363,"tokens_out":7519,"duration_ms":91532,"concrete_test":"Add two arms to the Table II protocol: (1) real-valued EfficientNetV2-B0 on HSV-encoded ALL-IDB2 with the identical 50/50 split, 128x128 resizing, augmentation, RMSProp schedule, and 400-epoch/10-run setup; (2) hyperbolic-quaternion V-EfficientNetV2-B0 on RGB under the same setup. If the real-HSV arm reaches roughly 99.46% or improves over real-RGB by the same ~1.2 points, the reported advantage is an HSV preprocessing artifact. If real-HSV stays near 98.2% and hyperbolic-RGB falls well below 99.46%, the vectorization/HSV interaction may be genuine. Report paired 95% confidence intervals for the differences between real-RGB and real-HSV and between real-HSV and hyperbolic-HSV across the 10 runs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing empirical claim is in the abstract and Section V: the V-EfficientNetV2-B0 based on hyperbolic quaternions with HSV encoding reaches 99.46% accuracy on ALL-IDB2, outperforming the real-valued EfficientNetV2-B0 baseline at 98.23% while using about 70% fewer parameters. But Table II changes two variables at once: the real baseline is reported only for RGB, while the two best V-EfficientNet results use HSV. No real-valued HSV baseline is reported, so the 1.23-point advantage could be entirely due to HSV preprocessing rather than to vector-valued algebra. The partial evidence already hints at this: quaternion RGB (98.38 ± 0.80) and coquaternion HSV (98.08 ± 2.07) do not beat the baseline; tessarines HSV (98.23 ± 1.50) ties; only the two HSV vector models exceed it. Moreover, no paired significance tests are reported, and the coquaternion and tessarine results show overlapping variability. The abstract's statement that V-EfficientNets are 'significantly reducing parameters while outperforming state-of-the-art models, including the original EfficientNet' therefore depends on an uncontrolled comparison. Cross-paper comparisons to prior state-of-the-art accuracy (e.g., ALLNet, 98.46%) also assume equivalent evaluation protocols, but the internal color-encoding confound is the more immediate problem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces V-EfficientNets, an adaptation of EfficientNetV2 to vector-valued non-associative algebras. The authors formalize vector-valued depthwise convolution via a real-valued implementation based on Kronecker products, introduce an architecture vectorization factor lambda in [1/d, 1], and evaluate the resulting models on the ALL-IDB2 acute lymphoblastic leukemia image classification task. They report an average accuracy of 99.46% for a hyperbolic-quaternion HSV model with about 70% fewer trainable parameters than the real-valued EfficientNetV2-B0 baseline, claiming state-of-the-art performance. The full text includes the algebraic framework, layer implementations, experimental details, and a link to source code.","tokens_in":12628,"tokens_out":20985,"duration_ms":204892,"significance":"If the claims are established, the paper makes a useful contribution: a general vector-valued depthwise convolution construction, a parameter-efficient vectorized EfficientNet variant, and a strong accuracy result on a medical benchmark with code provided. The algebraic derivation of the real-valued filter bank is internally consistent, and the vectorization factor is a clean design choice with no fitted constants. The reported parameter reduction is structurally plausible. However, the empirical superiority claim is not yet supported because the main comparison confounds architecture with color encoding, and the depthwise-convolution derivation is insufficiently specified.","major_comments":[{"comment":"The reported advantage of the best vector-valued model over the baseline is confounded with color encoding. The real-valued EfficientNetV2-B0 baseline is reported only with RGB input (98.23 ± 1.09), while the two models that exceed it, quaternion HSV and hyperbolic quaternion HSV, both use HSV encoding. No real-valued HSV baseline is given, so the 1.23-percentage-point gap could be entirely due to HSV preprocessing rather than the vector-valued algebra. The abstract's claim that V-EfficientNets 'outperform state-of-the-art models, including the original EfficientNet' therefore needs an additional control experiment: a real-valued EfficientNetV2-B0 trained on HSV input under the same protocol.","section":"Section V, Table II"},{"comment":"No significance testing is reported for the 10 repeated runs, despite the text using the words 'significantly outperformed' and 'solid margin.' The standard deviations overlap substantially (e.g., quaternion HSV 99.23 ± 0.91, hyperbolic quaternion HSV 99.46 ± 0.60, coquaternion HSV 98.08 ± 2.07), so the observed differences may not be statistically reliable. Paired comparisons across the 10 runs (e.g., Wilcoxon signed-rank test or a paired bootstrap) should be reported for each pair of models, especially for hyperbolic quaternion HSV versus real-valued RGB and versus quaternion HSV.","section":"Section V, Table II and following paragraph"},{"comment":"The conversion of the augmented real-valued depthwise output J(a) to the desired real representation J(R) is not specified precisely. The text says to sum channels 'row-wise in the Fig. 3' and then 'permutating appropriately' the channels of the resulting image, but no explicit index map or linear transformation is given. For a general non-associative algebra, a simple summation over one index followed by a permutation does not generally produce the components of the algebra product: for example, with quaternions the real component of the product requires the signed combination I0*F0 - I1*F1 - I2*F2 - I3*F3, which cannot be obtained by adding the four augmented channels that contain the positive products I1*F1, I2*F2, I3*F3. The paper should provide the explicit linear combination or permutation rule in terms of the structure constants πijk, or otherwise verify that the described operation is correct. Since the vector-valued depthwise convolution is the paper's main architectural novelty and is used in all experiments, this omission is load-bearing for the reproducibility and correctness of the method.","section":"Section III-B, Eqs. (9)-(12) and Figures 3-4"}],"minor_comments":[{"comment":"The table caption lists 'Color encoding' but the rows for vector models include the algebra and the encoding; consider adding a column for the algebra or a clearer separation of the two factors.","section":"Section V, Table II caption and text"},{"comment":"The word 'significantly' is used to describe accuracy differences without statistical support; please reserve it for significance tests or rephrase.","section":"Abstract and Section V"},{"comment":"The notation J(a) and the tilde variant J~(R) are not formally defined before use; a short definition or a labeled equation would improve readability.","section":"Section III-B, around Eq. (10)-(12)"},{"comment":"The range of λ is stated as [1/d, 1], but the description of the two extremes (λ=1/d and λ=1) could be clarified by giving the explicit number of filters in a layer for each case; a small table or example would help.","section":"Section IV-B, lambda definition"},{"comment":"The comparison to previously reported accuracies (e.g., ALLNet at 98.46%) assumes that the evaluation protocols are equivalent, yet the manuscript uses a particular 50/50 split, 128×128 resizing, and augmentation. A sentence acknowledging this limitation or providing a comparison under a shared protocol would strengthen the state-of-the-art claim.","section":"Section V, cross-paper comparison"}],"recommendation":"major_revision","confidential_remarks":"The paper leans heavily on the authors' own prior framework (refs. 9, 12, 28, 35) for both motivation and baseline choices, which is acceptable but should be acknowledged more explicitly as a potential source of confirmation bias in the experimental design. The code availability is a strength; the missing real-valued HSV control and the underspecified depthwise-convolution reduction are, in my view, fixable within a revision and do not warrant rejection at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper contributes a real and useful construction—vector-valued depthwise convolution and the lambda vectorization factor—but the headline accuracy claim rests on a confounded comparison, and the state-of-the-art claim is not established on current evidence.\n\nThe math in Sections II–III is careful and, as far as I can tell, correct: emulating a vector-valued depthwise convolution with real-valued depthwise convolution plus a channel permutation is a clean derivation, and the code is on GitHub, which is more than many papers in this area do. The vectorization factor lambda is a nice knob, giving a principled middle ground between parameter-compressed and parameter-expanded vector-valued networks. The choice to base on EfficientNetV2-B0 is natural, and the ~70% parameter reduction is arithmetic, not magic. The paper cites its own prior framework (refs 9, 10, 11, 35) honestly, and the lineage is clear.\n\nSoft spots are in Section V, exactly as the stress-test note says. Table II changes two variables at once: the real baseline uses RGB, while the best vector-valued results use HSV. No real-valued HSV baseline is reported, so the 1.23-point advantage of hyperbolic-quaternion HSV over real RGB could be mostly preprocessing. The internal evidence already hints at this: quaternion RGB (98.38) barely beats real RGB (98.23), coquaternion HSV (98.08) is worse, and tessarine HSV (98.23) ties. Only the two HSV quaternion/hyperbolic variants exceed the baseline. Without a real HSV control or paired significance tests across the 10 runs, the word \"significantly\" in Section V is doing work it shouldn't. The comparison to ALLNet's 98.46% also assumes equal evaluation protocols, which is not established.\n\nIs the central claim dead? No. The construction is valid, and a real HSV baseline would likely close part of the gap but might not eliminate it. This is a well-executed incremental extension for a niche community—hypercomplex/vector-valued deep learning and small-data medical imaging. It won't reorganize the field, but it is a solid building block.\n\nWho is this for: researchers in hypercomplex neural networks, and practitioners who want a compact EfficientNet variant for small medical image sets. I would send it to a serious referee at a specialty venue (ICASSP/ICIP or a neural-networks track), not desk reject, with the explicit request: add the real-valued HSV baseline, run paired significance tests, and if feasible evaluate on a second dataset. Current form is a conditional accept at best. If the authors fix the confound, it would be citable.","headline":"Useful vector-valued depthwise-convolution construction and a sensible vectorization factor, but the headline accuracy claim rests on a confounded RGB-versus-HSV comparison and needs one control experiment before it is citable.","tokens_in":13112,"tokens_out":2348,"would_cite":false,"duration_ms":26146,"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":"Vector-valued EfficientNets claim 99.46% accuracy on leukemia images with about 70% fewer parameters.","keywords":["vector-valued neural networks","hypercomplex-valued neural networks","EfficientNet","vector-valued depthwise convolution","non-associative algebras","medical image classification","acute lymphoblastic leukemia","ALL-IDB2"],"falsifier":"Retrain the real-valued EfficientNetV2-B0 on HSV-encoded images with the same 50/50 split, augmentation, and 10-run protocol; if its mean accuracy reaches or exceeds the 99.46% reported for the hyperbolic-quaternion model, the vector-valued architecture's advantage is explained by color encoding rather than by the algebraic mechanism.","tokens_in":12174,"feed_emoji":"🔬","tokens_out":6014,"duration_ms":59360,"temperature":0.7,"pith_summary":"This paper tries to establish that the EfficientNet architecture, which normally operates on real-valued feature maps, can be rebuilt over a d-dimensional non-associative algebra so that each neuron processes entire vectors as single entities. The authors construct vector-valued versions of convolutional and depthwise-convolutional layers, wrap them into MBConv and Fused-MBConv blocks, and introduce a vectorization factor that controls how many real features each vector replaces. On the ALL-IDB2 leukemia image dataset, the smallest such model based on hyperbolic quaternions reports 99.46% average accuracy, compared with 98.23% for the real-valued EfficientNetV2-B0 baseline and 98.46% for the previous state of the art, while cutting trainable parameters from 5.86 million to 1.72 million. If correct, this shows that moving from scalar to vector-valued algebra is a practical way to make compact CNNs more accurate on small medical datasets.","feed_headline":"Vector-valued EfficientNet hits 99.46% on leukemia images","feed_subtitle":"A quarter-size hyperbolic-quaternion model beats full EfficientNetV2-B0 and prior state of the art on ALL-IDB2.","key_machinery":"The load-bearing mechanism is the representation of vector-valued operations by real-valued tensor operations, centered on the multiplication-table tensor P whose entries define the product of basis elements in a non-associative algebra. Any vector product can be evaluated through a matrix-vector product built from P, which lets vector-valued dense, convolutional, and the newly introduced vector-valued depthwise convolutional layers be implemented with existing real-valued operations. A bank of vector-valued filters is assembled by summing Kronecker products of real filter banks with slices of P, so one vector-valued filter reuses related parameters across d channels. The architecture vectorization factor then controls whether each vector packs d real features, giving a parameter reduction by 1/d, or each real feature becomes a vector, upscaling by d, with the 1/4 setting used in the experiments. This construction is what turns the scalar EfficientNetV2-B0 into V-EfficientNet while preserving the MBConv and Fused-MBConv block structure.","core_discovery":"The central claim is that vector-valued EfficientNets, meaning EfficientNetV2-B0 with its scalar arithmetic replaced by multiplication in a four-dimensional hypercomplex algebra, can outperform both the real-valued baseline and prior specialized models on ALL-IDB2 while using roughly 70% fewer parameters. The winning model, based on hyperbolic quaternions and trained on HSV-encoded images, achieves 99.46% average accuracy over 10 runs, whereas the real-valued RGB baseline scores 98.23%. The paper further claims this is the highest mean accuracy reported for ALL-IDB2 to date, and presents these numbers as evidence that vector-valued networks exploit inter-channel relationships that real-valued networks must learn, with the choice of algebra acting as a meaningful design decision.","pith_inferences":["If the real-valued baseline were retrained on HSV input under the same augmentation and split, the reported gap might shrink: the winning model uses HSV while the baseline uses RGB, so the current comparison does not fully isolate the algebraic contribution.","The vector-valued depthwise convolution is a general primitive that could be dropped into other depthwise-heavy architectures, such as MobileNet-style backbones, not just EfficientNet.","With only 260 images and 10 runs, the separation between the best and second-best models may not be statistically significant; confidence intervals or a significance test over more seeds would settle it.","The algebraic structure of hyperbolic quaternions, which are non-associative and non-commutative, appears to matter for accuracy, so algebra selection could potentially be guided by geometric or group-theoretic properties of the data rather than brute force."],"forward_implications":["Replacing real-valued depthwise and regular convolutions with vector-valued counterparts in EfficientNetV2-B0 reduces trainable parameters by approximately 70% at the 1/4 vectorization factor, while classification accuracy on ALL-IDB2 rises from 98.23% to 99.46% for the hyperbolic-quaternion model.","The same vectorization recipe extends to other EfficientNet variants and scales, since it only swaps the underlying arithmetic and keeps the block-level architecture intact.","The experiments isolate the choice of hypercomplex algebra as a performance lever: quaternions, coquaternions, tessarines, and hyperbolic quaternions deliver different accuracies on the same task, with hyperbolic quaternions performing best.","Vector-valued depthwise convolution, introduced here, is what allows the MBConv blocks of EfficientNet to be vectorized, so the method is not limited to dense or regular convolutional layers."],"supporting_citations":[{"why":"Defines EfficientNet and compound scaling, supplying the underlying architecture philosophy that the paper extends.","marker":"[6]"},{"why":"Defines EfficientNetV2-B0, the real-valued backbone that is vectorized and the baseline compared against.","marker":"[7]"},{"why":"Shows how hypercomplex dense layers can be implemented with real-valued layers, providing the basis for vector-valued dense layers.","marker":"[10]"},{"why":"Introduces parameterized hypercomplex convolutions and reports parameter efficiency, providing the basis for vector-valued convolutional layers and the 1/d reduction claim.","marker":"[11]"},{"why":"Introduces MobileNetV2's inverted residual blocks with depthwise convolution, which underpins the MBConv blocks used in the architecture.","marker":"[26]"},{"why":"Prior hypercomplex CNN work on ALL-IDB2 that supplies the RGB and HSV encoding setup and color-encoding precedent.","marker":"[35]"},{"why":"Reports the previous highest accuracy on ALL-IDB2 that the paper claims to surpass.","marker":"[39]"},{"why":"Supplies the 50-50 train-test split protocol used for evaluation.","marker":"[41]"}],"fun_headline_variants":["V-EfficientNets: 99.46% on leukemia with 70% fewer parameters","Hyperbolic quaternion EfficientNet tops leukemia benchmark","Vector-valued EfficientNets slim down, beat state of the art","99.46% leukemia detection: vector-valued EfficientNet wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim assumes that the 1.2-point accuracy gap between the winning vector-valued model and the real baseline comes from the vector-valued algebra, not from the fact that the winning model was trained on HSV images while the baseline used RGB.","fun_headline_variants_meta":{"raw":{"variants":["V-EfficientNets: 99.46% on leukemia with 70% fewer parameters","Hyperbolic quaternion EfficientNet tops leukemia benchmark","Vector-valued EfficientNets slim down, beat state of the art","99.46% leukemia detection: vector-valued EfficientNet wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000269,"raw_usage":{"total_tokens":1581,"prompt_tokens":864,"completion_tokens":717,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":641}},"tokens_in":480,"tokens_out":717,"duration_ms":6972,"temperature":1.0,"reasoning_tokens":641,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:00:12.616450+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the real-valued EfficientNetV2-B0 on HSV-encoded images with the same 50/50 split, augmentation, and 10-run protocol; if its mean accuracy reaches or exceeds the 99.46% reported for the hyperbolic-quaternion model, the vector-valued architecture's advantage is explained by color encoding rather than by the algebraic mechanism.","supporting_citations":[{"cited_title":"EfficientNet: Rethinking model scaling for con- volutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Defines EfficientNet and compound scaling, supplying the underlying architecture philosophy that the paper extends."},{"cited_title":"Efficientnetv2: Smaller models and faster training,","cited_arxiv_id":null,"evidence_quote":"Defines EfficientNetV2-B0, the real-valued backbone that is vectorized and the baseline compared against."},{"cited_title":"PHNNs: lightweight neural networks via parameterized hypercomplex convolutions,","cited_arxiv_id":null,"evidence_quote":"Introduces parameterized hypercomplex convolutions and reports parameter efficiency, providing the basis for vector-valued convolutional layers and the 1/d reduction claim."},{"cited_title":"Acute Lymphoblastic Leukemia Detection Using Hypercomplex-Valued Convolutional Neural Networks,","cited_arxiv_id":null,"evidence_quote":"Prior hypercomplex CNN work on ALL-IDB2 that supplies the RGB and HSV encoding setup and color-encoding precedent."},{"cited_title":"ALLNet: acute lymphoblastic leukemia detection using lightweight convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Reports the previous highest accuracy on ALL-IDB2 that the paper claims to surpass."},{"cited_title":"Histopathological transfer learning for acute lymphoblastic leukemia detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the 50-50 train-test split protocol used for evaluation."}],"review_version":1}