{"id":"9ba522db-b299-41ed-9e38-f5ab8406e406","arxiv_id":"2509.03103","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A complete CapsNet was deployed on a PYNQ-Z1 FPGA, reaching 1351 FPS (MNIST) and 934 FPS (F-MNIST) via look-ahead kernel pruning and hardware-friendly routing approximations.","lead":"This paper shows how to run a complete capsule network on a cheap FPGA board, increasing speed from 5 to 1351 images per second on MNIST. It combines a new kernel-pruning method with simplified routing math, which matters for putting such networks on small edge devices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Truncated Taylor softmax in Eq. 2 is only accurate near logit 0.5; paper never bounds routing logits b_ij, so the '<1% accuracy drop' claim for the FPGA model is unsupported—and Table I already contradicts it for F-MNIST.","rationale":"The central claim is an FPGA CapsNet that is simultaneously fast and accurate. The speedup is plausible if the pruned network is correct, but the accuracy side rests on two unsupported/contradicted claims: the Taylor exp/div approximations are asserted to have no accuracy drop without a logit-range analysis or hardware accuracy measurement, and Table I's F-MNIST entries already show >1 pp drop at the compression rate claimed. The reader's verdict of CONDITIONAL is appropriate: the paper should be accepted only if the authors provide a logit-range analysis and full reproduced accuracy of the deployed bitstream (and correct the F-MNIST claim). My concern does not move the verdict; it strengthens the conditions.","tokens_in":10803,"tokens_out":11644,"duration_ms":129554,"concrete_test":"Take the exact pruned weights used for the FPGA deployment and run the software CapsNet on MNIST and F-MNIST, logging all routing logits b_ij after each routing iteration. Then replace exp() and division in the softmax with the fixed-point Eq. 2/Eq. 3 approximations and re-measure test accuracy. If any b_ij falls below about -1.7 (where the truncated polynomial becomes negative) or the approximated model loses >1 percentage point on either dataset, the claim that the optimized routing runs 'without dropping accuracy' fails; the throughput numbers would then describe a different, less accurate classifier.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B replaces exp() with a 5th-order Taylor polynomial expanded around 0.5 (Eq. 2). This polynomial is not globally accurate: at x = -1.7 it already becomes negative while the true exp is 0.18, and at b_ij values beyond a couple of units the softmax coupling coefficients are systematically wrong. Dynamic routing updates b_ij by accumulated dot-product agreements, so negative and >1-magnitude logits are plausible, but the paper reports neither the observed b_ij range nor an end-to-end accuracy measurement of the FPGA implementation. Eq. 3 then implements division as exp(log(a)-log(b)) on the same flawed exp, adding further error and requiring positive operands, which is not guaranteed once the polynomial goes negative. The entire 270x/187x speedup is measured on this approximate design; if the approximation changes routing outputs, the 'no accuracy drop' assertion is unsupported. Independently, Table I shows F-MNIST baseline error 10.31% vs 11.82% at 1.37% survived weights, a 1.51-pp drop; the claimed 98.84% compression (~1.16% survived) would be worse, so the published data already contradict the '<1% drop' claim for F-MNIST.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FastCaps, a two-step methodology for deploying a full CapsNet on a low-cost FPGA (Xilinx PYNQ-Z1). First, LAKP (Look-Ahead Kernel Pruning) is introduced as a structured kernel-pruning method that uses look-ahead scores summed per kernel. Second, the dynamic routing algorithm is simplified by approximating exp() and div() with Taylor-series-based operations, reordering loops, and parallelizing operations with PE arrays. The paper reports throughput improvements from 5 FPS for the original CapsNet to 1351 FPS (MNIST) and 934 FPS (F-MNIST) for the pruned and optimized design, with effective compression rates of 99.26% and 98.84%, respectively, and claims an accuracy drop of less than 1%. It also presents LAKP results on VGG-19 and ResNet-18. The claimed contributions are a first full-fledged CapsNet FPGA accelerator, a novel pruning methodology, and a hardware-oriented routing optimization.","tokens_in":11161,"tokens_out":8239,"duration_ms":86351,"significance":"If the accuracy claims are validated, this is a practically useful contribution: it demonstrates that a full CapsNet can be mapped to a low-cost FPGA with very high throughput and high compression, using a pruning method that produces regular sparse structures compatible with hardware. The concrete hardware measurements (throughput, resource utilization, latency) are a strength. The paper also provides empirical comparisons of LAKP against magnitude-based kernel pruning on three architectures and multiple datasets, which supports the pruning methodology. However, the paper does not provide end-to-end accuracy measurements of the deployed FPGA design, and one of the stated accuracy claims is contradicted by the paper's own Table I. These issues make the central claim (high-speed FPGA CapsNet with <1% accuracy drop) currently unsubstantiated.","major_comments":[{"comment":"The conclusion states LAKP achieves 99.26% and 98.84% compression 'with an accuracy drop of less than 1%'. For F-MNIST, Table I gives a baseline test error of 10.31%; at 1.37% survived weights the LAKP error is 11.82%, i.e. +1.51 percentage points. The claimed 98.84% compression corresponds to 1.16% survived weights; interpolating between the 1.37% and 0.25% rows yields an error around 12.4%, about +2.1 pp. The <1% claim is therefore contradicted by the paper's own data. Please either restrict the claim to MNIST, report the actual drop for each dataset, or adjust pruning rates so the error increase is within the stated bound.","section":"Table I / Section V"},{"comment":"The 'without dropping accuracy' assertion for the Taylor approximation of exp() is unsupported. The manuscript does not bound the range of routing logits b_ij observed during inference, nor does it report the approximation error of Eq. (2) over that range. The polynomial is not globally accurate: for negative logits with magnitude greater than about 1.7 it can return negative values, which breaks softmax and makes Eq. (3) ill-defined because log of a negative operand is not real. Dynamic routing updates b_ij by accumulating dot-product agreements, so such logits are plausible. As written, Eq. (2) also is not the Taylor expansion around a=0.5: the coefficients do not match and the powers should be of (x-a), not x. Please provide the empirical distribution of b_ij, an error bound for Eq. (2) on that range, and end-to-end accuracy of the FPGA implementation.","section":"Section III-B, Eq. (2)-(3)"},{"comment":"No accuracy measurement of the deployed hardware is reported. The paper states that the 16-bit quantization 'did not lead to a reduction in accuracy', but no accuracy numbers for the FPGA model are given. Because the deployed model combines pruning, 16-bit quantization, and the approximate softmax/squash of Section III-B, the central claim of '<1% accuracy drop' cannot be checked without the accuracy of the complete FPGA design. Please include test accuracy of the full bitstream on MNIST and F-MNIST, compared to the baseline software CapsNet.","section":"Section IV-B"},{"comment":"The paper claims to be the first to accelerate a full-fledged CapsNet on FPGA, but the related work includes [30], 'Enabling Capsule Networks at the Edge through Approximate Softmax and Squash Operations' (ISLPED 2022), which is directly related to the routing simplification proposed here. A quantitative comparison with [30] (and with any other FPGA CapsNet works) is needed to substantiate the novelty claim. If [30] already implements a full CapsNet on FPGA, the 'first' statement must be qualified.","section":"Section I and Related Work"}],"minor_comments":[{"comment":"The acronym is inconsistently written as 'LKAP' in one sentence ('We propose a Look-Ahead Kernel Pruning (LKAP)') and as 'LAKP' elsewhere. Also, in Algorithm 1, 'fksi as a si-th smallest element' should read 'the s_i-th smallest element'.","section":"Section III-A"},{"comment":"The sentence 'where a = 0.5' is confusing because the polynomial is written in powers of x, not (x - a). Please clarify the expansion point and write the polynomial in the standard form (x - a).","section":"Eq. (2)"},{"comment":"Typo: 'CapNet' should be 'CapsNet'.","section":"Table III caption"},{"comment":"The unit 'FPJ' appears twice; likely should be 'pJ' or 'fJ' for energy. Please correct.","section":"Fig. 1"},{"comment":"Figure 5 is described as comparing 'various pruning techniques', but the axes and the quantities plotted are not defined. Please add axis labels and clarify whether the y-axis is test error and the x-axis is survived rate.","section":"Section IV-A / Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the hardware measurements are valuable. The main concern is that the accuracy claim, a key part of the central contribution, is not validated by the data presented and appears to be contradicted for F-MNIST. The authors should be asked to provide end-to-end FPGA accuracy numbers and a proper error analysis of the approximate softmax. If the F-MNIST drop is in fact greater than 1%, the claim should be revised accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look: this is a genuine full-pipeline CapsNet implementation on a low-cost FPGA, with concrete measured throughput (1351/934 FPS on MNIST/F-MNIST) and real resource numbers. The authors also show their structured LAKP pruning consistently beats magnitude-based kernel pruning across CapsNet, VGG-19, and ResNet-18, in some cases by large margins. That part is solid and reproducible given the tables. The engineering work is real, and I don't doubt the speedup numbers.\n\nWhere I share the reader's unease: the abstract claims 'less than 1% accuracy drop' for F-MNIST, but Table I shows the LAKP-pruned model at 1.37% survived weights has an error of 11.82% vs. a 10.31% baseline, a 1.51-point rise, and the abstract's 98.84% compression corresponds to ~1.16% survived, which would land at or below that row. So the claim contradicts their own data. That needs a rewrite or a corrected measurement.\n\nThe bigger soft spot is Section III-B. Equation 2 gives a 5-term Taylor polynomial for exp() expanded around 0.5. The paper asserts 'without dropping accuracy' but never bounds the range of routing logits b_ij encountered during inference. Dynamic routing accumulates dot-product agreements, so logits can plausibly leave the region where this polynomial is accurate; the polynomial itself goes negative around x ≈ -1.7. And Eq. 3 also assumes positive operands for the log approach. If the authors can't show the observed b_ij range and an end-to-end accuracy measurement on the FPGA for both datasets, the 'no accuracy drop' claim for the optimized design is unsupported. This is fixable: report the logit distribution, or use the exact exp/div in the FPGA model and compare accuracy.\n\nMinor: the 'first full-fledged CapsNet on FPGA' claim needs a baseline table versus prior CapsNet accelerators, especially [30] and [16], which are cited but never compared. Also, the LAKP novelty is modest—it's summing the existing Lookahead score over kernels—but it produces consistent gains, so I won't overstate that.\n\nOverall: the central deployment claim—that a pruned and optimized CapsNet can run at 1300 FPS on a PYNQ-Z1—appears solid because it's a measured quantity. The accuracy part is oversold and the approximations under-specified. This deserves a serious referee: the engineering result matters to the CapsNet hardware community, and the flaws are correctable with extra evidence rather than fatal.\n\nRecommendation: send to peer review, but require the authors to fix the F-MNIST accuracy claim, add a comparison with prior FPGA/ASIC CapsNet implementations, and provide end-to-end accuracy for the approximate routing on the FPGA.","headline":"A credible CapsNet FPGA acceleration result with real measured speedups, but the accuracy-drop claim is inconsistent and the Taylor softmax approximation needs bounding or end-to-end accuracy evidence.","tokens_in":11640,"tokens_out":970,"would_cite":true,"duration_ms":13673,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a two-step method to put a complete Capsule Network on a low-cost FPGA and reports the first full-fledged CapsNet deployment of its kind, reaching 1351 FPS on MNIST and 934 FPS on F-MNIST.","keywords":["Capsule Network","FPGA accelerator","structured kernel pruning","look-ahead score","dynamic routing","Taylor approximation","edge inference","model compression"],"falsifier":"Instrument the pruned MNIST and F-MNIST models to record the actual routing logits, then compare exact softmax output against the five-term Taylor softmax. If any test sample yields a different argmax digit, or if the measured accuracy gap between the two versions exceeds the claimed <1 percent, the no-drop claim would be refuted.","tokens_in":10706,"feed_emoji":"⚡","tokens_out":4952,"duration_ms":54282,"temperature":0.7,"pith_summary":"This paper proposes a two-step method to put a complete Capsule Network on a low-cost FPGA and reports the first full-fledged CapsNet deployment of its kind. The first step is a structured pruning method, LAKP, that removes whole kernels by scoring them with look-ahead information from neighboring layers, shrinking the routing stage dramatically. The second step replaces the routing algorithm's exp() and division operations with Taylor-based arithmetic and reorders loops so operations run in parallel. On MNIST and F-MNIST the pruned, optimized design runs at 1351 and 934 FPS, 270x and 187x faster than the original CapsNet, while keeping accuracy loss under 1 percent and compressing the network by about 99 percent. If true, this gives edge-device designers a working recipe for CapsNet inference at high speed on cheap hardware.","feed_headline":"Full CapsNet on a low-cost FPGA hits 1351 FPS","feed_subtitle":"Structured kernel pruning and a Taylor-based softmax shorten MNIST inference from 0.19 s to 0.74 ms per image.","key_machinery":"Two mechanisms carry the argument. (1) Look-Ahead Kernel Pruning (LAKP) scores each convolution kernel by summing per-parameter look-ahead scores—each weight's magnitude times the Frobenius norms of the adjacent-layer weight slices it connects to—and removes the lowest-scoring kernels layer by layer. This structured pruning collapses the PrimaryCaps capsule count from 1152 to 252 (MNIST) and 432 (F-MNIST), reducing routing weight parameters by 1280x. (2) The simplified routing datapath replaces exp() with a fixed five-term Taylor polynomial around a=0.5 and computes division a/b as exp(log(a)-log(b)), then reorders loops and uses a 10-PE array to parallelize the agreement and fully-connected","core_discovery":"The central claim is that a full CapsNet—not just the routing loop—can fit on an FPGA and run fast if you attack both the parameter count and the routing math. CapsNets are usually considered too large and too irregular for FPGAs; the paper's counter is a two-part pipeline. LAKP turns simple magnitude-based kernel pruning into structured pruning by summing a look-ahead score per kernel, cutting the digit-capsule routing computation drastically and reducing routing weight parameters by up to 1280x. Then the routing algorithm's exp() and div() are replaced with Taylor approximations and multiplications, cutting softmax latency by 85 percent, after which loop reordering makes the agreement and","pith_inferences":["The claimed <1 percent accuracy drop couples the pruning step with the Taylor softmax approximation; a natural follow-up is to measure routing-logit ranges on other datasets, since a fixed five-term expansion may need more terms or a range check when routing dynamics differ from MNIST/F-MNIST.","The reported 270x and 187x speedups combine two effects—pruning and math simplification—and the paper already separates them in throughput (82/48 FPS after pruning, 1351/934 after optimization); a reader should note that accuracy is reported after pruning, while the additional routing approximation is asserted not to change it.","Because the routing simplification is independent of training, the same accelerator design could be re-targeted to higher-resolution inputs or deeper capsule stacks, with the practical limit set by BRAM and DSP headroom on larger FPGAs.","On low-cost boards, throughputs in the hundreds-to-thousands of FPS range would make CapsNet competitive with lightweight CNNs for small-image classification workloads, an implication the authors leave implicit."],"forward_implications":["Pruning the layer before a capsule layer directly cuts the number of capsules, so any CapsNet with a convolution front-end can inherit this acceleration without changing the routing algorithm's semantics.","The routing optimizations are modular: other capsule networks deployed on FPGA or ASIC could adopt the Taylor-based exp/div blocks and reordered loops to gain an 85 percent reduction in softmax latency without a full redesign.","Because LAKP prunes whole kernels, the accelerator stores only kernel indices—about 0.1 percent of the surviving weight count—rather than per-weight indices, keeping on-chip memory small.","LAKP transfers beyond CapsNet: on VGG-19 and ResNet-18 across CIFAR-10 and GTSRB it consistently outperforms magnitude-based kernel pruning, with the largest gains in the high-sparsity regime.","With 16-bit quantization and the simplified routing math, the design reportedly keeps accuracy loss under 1 percent, making the pruned-and-optimized model a practical candidate for edge deployments."],"supporting_citations":[{"why":"Defines the original CapsNet architecture and dynamic routing algorithm that this paper accelerates, providing the baseline model and the routing steps being optimized.","marker":"[4]"},{"why":"Supplies the magnitude-based structured kernel pruning baseline (KP) that LAKP is compared against and that motivates the need for look-ahead scoring.","marker":"[14]"},{"why":"Provides the look-ahead score formula for individual parameters, which LAKP adapts into a sum over kernels for structured pruning.","marker":"[15]"},{"why":"A prior capsule-pruning work that prunes 95 percent of primary capsules; the paper uses it as the comparison point for pruning effectiveness in CapsNets.","marker":"[24]"},{"why":"Shows prior work on approximate softmax and squash operations for CapsNets at the edge, providing context for the routing-algorithm simplifications proposed here.","marker":"[30]"}],"fun_headline_variants":["Full CapsNet on FPGA hits 1351 FPS via LAKP pruning","CapsNet on low-cost FPGA: 1351 FPS after 99% prune","FastCaps: Kernel pruning and Taylor routing boost CapsNet to 1351 FPS","FPGA runs full CapsNet at 1351 FPS with 99% compression"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The routing softmax stays accurate when the fixed five-term Taylor expansion of exp() covers the range of routing logits seen during inference; the paper reports no accuracy drop but never bounds that range.","fun_headline_variants_meta":{"raw":{"variants":["Full CapsNet on FPGA hits 1351 FPS via LAKP pruning","CapsNet on low-cost FPGA: 1351 FPS after 99% prune","FastCaps: Kernel pruning and Taylor routing boost CapsNet to 1351 FPS","FPGA runs full CapsNet at 1351 FPS with 99% compression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000222,"raw_usage":{"total_tokens":1359,"prompt_tokens":884,"completion_tokens":475,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":384}},"tokens_in":628,"tokens_out":475,"duration_ms":5062,"temperature":1.0,"reasoning_tokens":384,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:06:36.227895+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Instrument the pruned MNIST and F-MNIST models to record the actual routing logits, then compare exact softmax output against the five-term Taylor softmax. If any test sample yields a different argmax digit, or if the measured accuracy gap between the two versions exceeds the claimed <1 percent, the no-drop claim would be refuted.","supporting_citations":[{"cited_title":"Dynamic Routing Between Capsules,","cited_arxiv_id":null,"evidence_quote":"Defines the original CapsNet architecture and dynamic routing algorithm that this paper accelerates, providing the baseline model and the routing steps being optimized."},{"cited_title":"Exploring the Regularity of Sparse Structure in Convolutional Neural Networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the magnitude-based structured kernel pruning baseline (KP) that LAKP is compared against and that motivates the need for look-ahead scoring."},{"cited_title":"Lookahead: A Far-sighted Al- ternative of Magnitude-based Pruning,","cited_arxiv_id":null,"evidence_quote":"Provides the look-ahead score formula for individual parameters, which LAKP adapts into a sum over kernels for structured pruning."},{"cited_title":"PrunedCaps: A Case For Pri- mary Capsules Discrimination,","cited_arxiv_id":null,"evidence_quote":"A prior capsule-pruning work that prunes 95 percent of primary capsules; the paper uses it as the comparison point for pruning effectiveness in CapsNets."},{"cited_title":"Enabling Capsule Networks at the Edge through Approximate Softmax and Squash Operations,","cited_arxiv_id":null,"evidence_quote":"Shows prior work on approximate softmax and squash operations for CapsNets at the edge, providing context for the routing-algorithm simplifications proposed here."}],"review_version":1}