{"id":"0535e95c-7bfa-48dd-8703-b054b0fba986","arxiv_id":"2412.19828","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A simulated quantum neural network (quINR) reportedly improves compression efficiency by up to 1.2 dB over COIN and JPEG2000 on one LiDAR image and one Kodak image, but without reproducible details.","lead":"This paper proposes quINR, a hybrid quantum-classical neural network for image compression, and reports that it beats COIN and JPEG2000 on two test images. A generalist might read it to see whether quantum neural networks can improve neural compression, but the evidence is thin and the bit-rate accounting is unspecified.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed rate-distortion gain is not falsifiable because bpp is never defined and parameter quantization/bit-width accounting is unspecified; if quINR bpp excludes the input linear layer or uses fewer bits per parameter than COIN, the 1.2 dB advantage can be an artifact.","rationale":"The reader's verdict is REJECT with high correctness risk, and the weakest assumption identified is exactly the undefined bit-rate accounting. My stress-test agrees: the central claim is explicitly rate-distortion ('up to 1.2 dB gain'), so the bpp axis is not a peripheral detail. The manuscript never defines bpp, never states the bit-width of stored parameters, and never specifies whether the input linear layer and other classical components are included in the quINR bit count. Without this, the plotted comparison between quINR and COIN is unverifiable: a different accounting choice could move quINR's curve horizontally by a large factor and erase the reported advantage. The other weaknesses noted by the reader—two images without error bars, the 8-bit JPEG2000 restriction, and missing code/data—are real but secondary; the missing bpp definition is the load-bearing flaw because it attacks the very quantity being compared. A single concrete check, recomputing the rate-distortion curves under a symmetric, explicitly stated bpp formula, would settle whether the concern lands. Since this check cannot be performed from the paper as written, the REJECT verdict stands without requiring any adjustment from my read.","tokens_in":5462,"tokens_out":4522,"duration_ms":44696,"concrete_test":"Obtain the exact bpp formula used for Figs. 3–4 and recompute the Kodak curve with one shared accounting rule: bpp = (number of stored scalar parameters × bits_per_parameter) / (768 × 512), where stored scalar parameters include W, b, all QNN rotation angles, and any output parameters, and bits_per_parameter is the same (e.g., 16) for quINR and COIN under the same quantization scheme. If the published quINR bpp omitted W and b or used a lower bit-width than COIN, re-plot with corrected counts; if the 1.2 dB crossover disappears, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a rate-distortion comparison, so the bitrate accounting is load-bearing. The paper never defines bpp nor states how the optimized parameters ψ are quantized or stored. From Eq. (1) and Fig. 1(b), ψ includes at least the input linear layer parameters W and b, the QNN rotation angles in the entangling layers, and any output-stage parameters, yet Figs. 3–4 plot quINR and COIN curves with no statement of which parameters are counted, at what bit-width, and whether quantization or entropy coding is applied identically to both methods. Because the x-axis is the only basis for the claimed 'up to 1.2 dB gain,' any asymmetry—for example, charging COIN for all weights at 32 bits while charging quINR only for a subset of rotation angles at lower precision, or omitting W and b from the quINR bit count—would shift the quINR curve and could eliminate the claimed advantage. The absence of code and data and the use of only two images compound this, but the unspecified bpp is the decisive gap: without it, the reported Pareto frontiers cannot be reproduced or meaningfully compared.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces quINR, a hybrid quantum-classical implicit neural representation for signal compression. Coordinates are passed through a linear sinusoidal layer, then a folded-angle quantum embedding and a parameterized entangling circuit, and finally read out via probability measurements and a QReLU output stage. The authors train quINR and COIN on one KITTI LiDAR range image and one Kodak color image, plot PSNR versus bpp against JPEG2000 and COIN, and claim up to 1.2 dB gain in rate-distortion performance.","tokens_in":5737,"tokens_out":5166,"duration_ms":42862,"significance":"The idea of using a quantum neural network as the backbone of an INR is creative and, if the claimed gains were substantiated, would point toward a parameter-efficiency advantage for quantum circuits in compression. The manuscript also has the virtue of explicitly acknowledging the limited color-image performance. However, the central quantitative claim is not verifiable from the text because the bit-rate accounting is unspecified, the JPEG2000 comparison on the range image is unfair, and the evaluation rests on just two images with no statistical support. As written, the paper does not establish its advertised improvement.","major_comments":[{"comment":"The x-axis is 'Bits per Pixel (bpp)', but the manuscript nowhere defines bpp, nor does it state how the optimized parameter set ψ (which includes W and b from Eq. (2), the QNN rotation angles, and any output-stage parameters) is quantized, entropy-coded, or otherwise counted. Because the rate-distortion claim (up to 1.2 dB gain) rests entirely on this axis, any asymmetry in the parameter accounting between quINR and COIN—such as charging COIN for all weights at 32 bits while charging quINR for only a subset of rotation angles at lower precision—could shift the reported curves and eliminate the advantage. The bpp accounting must be specified and applied identically to all methods before the central claim can be assessed.","section":"Performance Comparison (Figs. 3 and 4)"},{"comment":"The RI comparison of Fig. 3 is biased against JPEG2000 because the manuscript states that JPEG2000 requires conversion to 8-bit precision in advance, whereas the LiDAR range image stores distance values ρ with floating-point precision. Since the two methods are not operating on the same target precision, the reported PSNR curve for JPEG2000 is not directly comparable, and the conclusion that quINR 'achieves better image quality than other baselines' on RI is not supported.","section":"Experiments, Baseline"},{"comment":"The evaluation uses exactly one range image and one color image, with no error bars or statistical significance measures, and no code or trained models are released. The training details (optimizer, learning rate, number of iterations, the hyperparameter sweep used for the Pareto frontier) are also omitted. Given that the claimed improvement is up to 1.2 dB, which is within the typical variation across images and random seeds, these results cannot be considered a reliable demonstration of a rate-distortion advantage.","section":"Experiments, Settings and Performance Comparison"}],"minor_comments":[{"comment":"The word 'expressibity' should be 'expressibility'.","section":"Related Work"},{"comment":"The reference to Parigi et al. has a missing year (listed as '????') and an incomplete bibliographic entry.","section":"References"},{"comment":"The paper cites COIN++ (Dupont et al., 2022) but does not include it as a baseline; because COIN++ is a stronger INR codec than COIN, the claim of outperforming 'classic INR-based coding methods' is conditional at best.","section":"Baseline"},{"comment":"The wording of the conclusion ('demonstrated good PSNR performance') overstates the results; the color-image performance in Fig. 4 is below JPEG2000 over much of the plotted range, and the abstract more cautiously says 'could improve'.","section":"Conclusion"}],"recommendation":"reject","confidential_remarks":"This manuscript reads more like a workshop extended abstract than a journal paper. The central idea is plausible, but the missing definition of the rate metric is a fundamental flaw for a compression paper. The unfair JPEG2000 comparison and the lack of any reproducibility artifacts further weaken the evidence. I recommend rejection of the current version; a future submission with a transparent bit accounting, unbiased comparisons, and a more systematic evaluation could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible proof-of-concept for using a parameterized quantum circuit as the implicit neural representation in compression, but the headline result—up to 1.2 dB gain over COIN and JPEG2000—rests on an undefined bitrate accounting. The paper never says which parameters count toward the bitstream or at what precision, so the rate-distortion curves in Figs. 3–4 cannot be reproduced or compared. That’s the thing to know.\n\nWhat’s actually new: the folded-angle embedding, which packs an arbitrary-length embedding vector onto a fixed number of qubits by alternating RX and RZ rotations, is a small but reasonable engineering contribution. And as far as I can tell, this is the first paper to evaluate a QNN-based INR in a rate-distortion setting. The authors are also honest in the conclusion that the color-image results are limited. The writing is clear and the architecture description is understandable.\n\nThe soft spots are real and load-bearing. First, bpp is never defined. From the text and Fig. 1(b), the parameter set includes the input linear layer weights and biases, the QNN rotation angles, and any output-stage parameters, but Figs. 3–4 plot quINR and COIN without saying how each is quantized or entropy-coded. If quINR is charged for fewer parameters or fewer bits per parameter than COIN, the 1.2 dB gain can be an artifact. The stress-test note is right about this. Second, the experiments use one LiDAR range image and one Kodak image, with no error bars and no code or data. Third, JPEG2000 is forced to 8-bit for the range image, which biases that comparison. These are not cosmetic; they make the central claim unverifiable as written.\n\nThat said, the idea is not wrong. A hybrid quantum-classical INR for compression is a legitimate research direction, and the folded-angle embedding is worth exploring. The paper just doesn’t provide enough evidence yet. I’d treat it as a workshop-quality contribution that needs serious follow-up: define the bitstream, release the code, run on a real benchmark suite, and compare against COIN++ rather than just COIN.\n\nFor peer review: yes, I’d send it out, not desk-reject. The topic is timely and the approach is new enough to merit referee time, even though my own verdict would be reject unless the bpp accounting is fixed and the evaluation is expanded. If the authors can show the gain survives consistent bitrate accounting on more than two images, the result becomes meaningful.\n\nConfidence: moderate. The missing bpp is the kind of thing that might be in the appendices of a longer version, but as this paper stands, it’s an open technical hole.","headline":"Plausible quantum-INR compression idea, but the rate-distortion claim is unverifiable because bpp is never defined.","tokens_in":6275,"tokens_out":2636,"would_cite":false,"duration_ms":22081,"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":"quINR, a hybrid quantum-classical INR codec, outperforms JPEG2000 and COIN by up to 1.2 dB in rate-distortion performance on tested images.","keywords":["quantum neural network","implicit neural representation","image compression","rate-distortion","LiDAR range image","hybrid quantum-classical","data re-uploading","folded-angle embedding"],"falsifier":"Re-run the rate-distortion comparison with a fixed quantization scheme applied identically to COIN and quINR parameters, and include the bit cost of the circuit description and measurement settings; if quINR no longer reaches the reported PSNR at the same bits per pixel, the claimed gain is an accounting artifact rather than a coding gain.","tokens_in":5256,"feed_emoji":"🖼️","tokens_out":6723,"duration_ms":44574,"temperature":0.7,"pith_summary":"Quantum implicit neural compression (quINR) is a hybrid quantum-classical codec that overfits a small neural network to map image coordinates to pixel values, then transmits only the trained parameters. The authors claim that, because the quantum layers pack information more densely than classical layers, the network can reproduce high-frequency details with fewer parameters, improving rate-distortion performance. On a LiDAR range image and one Kodak color image, they report reconstructions up to 1.2 dB higher PSNR than JPEG2000 and the classical INR baseline COIN at similar bit rates. If the gain is real, quantum INR would be a parameter-efficient alternative for lossy image and sensor-data compression.","feed_headline":"Hybrid quantum codec lifts image compression up to 1.2 dB","feed_subtitle":"A compact quantum-classical network maps pixel coordinates to colors, beating classic codecs on tested images.","key_machinery":"The load-bearing object is the quINR architecture, a hybrid quantum-classical network whose QNN layers carry most of the representational load. Its folded-angle embedding packs an M-dimensional embedding vector into alternating RX and RZ rotations on a small number of qubits, avoiding the qubit-count limit of standard angle embedding; its entangling layers use single-qubit rotations plus two-qubit controlled-Z rotations, shuffled with a data re-uploading trick; and its output layer measures quantum-state probabilities and applies a quantum ReLU. The argument is that the exponentially large Hilbert space of the circuit lets a small set of classical parameters encode high-frequency pixel variation that a classical MLP of the same size cannot.","core_discovery":"The paper's central claim is that replacing a layer of a classical implicit neural network with a parameterized quantum circuit produces a coordinate-to-value mapping that reconstructs images more accurately at the same bit rate. The proposed quINR architecture folds the coordinate embedding into alternating RX/RZ rotations on a small register of qubits, iterates entangling layers with data re-uploading, and reads out probabilities through a quantum ReLU; the classical parameters are trained by mean squared error minimization and then transmitted as the compressed representation. In the reported experiments, quINR's PSNR-versus-bits-per-pixel frontier lies above COIN and JPEG2000 on a LiDAR range image, and on the Kodak color image it beats both baselines in the low-to-medium rate regime with up to 1.2 dB gain. The authors present this as evidence that quantum expressivity can be converted into compression efficiency, while acknowledging that color-image performance is limited and needs further work.","pith_inferences":["The reported gain depends on an unstated bit-accounting rule; a fair comparison would need the bit cost of quantized classical parameters plus any circuit-description overhead.","Because the QNN is simulated classically in the experiments, the compression results do not yet show that real quantum hardware can deliver the same gain under noise.","Folded-angle embedding is a general method for packing many classical features into few qubits and could be reused in other hybrid quantum models outside compression.","A natural next experiment would test quINR on all 24 Kodak images and against COIN++ to see whether the single-image result represents a systematic advantage."],"forward_implications":["quINR shifts some of the representational burden from classical weights to the quantum circuit, so compression of LiDAR-like range data could need fewer stored parameters at equal quality.","The reported gains appear in the low-to-medium bit-rate range, which is exactly where INR codecs are expected to compete with classical codecs.","Because the same training loop maps arbitrary coordinates to values, the scheme could be applied to other signal types, including video frames and 3D point clouds.","The method's main promise is not replacing JPEG2000 at high quality but enabling extremely small model sizes for sensor data."],"supporting_citations":[{"why":"Supplies the COIN baseline that quINR must beat and defines the INR-based image compression setting.","marker":"Dupont et al. 2021"},{"why":"Extends COIN across modalities and anchors the INR compression framework the paper builds on.","marker":"Dupont et al. 2022"},{"why":"Provides the parameterized entangling circuit used in the QNN layers and the expressibility argument.","marker":"Sim, Johnson, and Aspuru-Guzik 2019"},{"why":"Supplies the data re-uploading trick and universal approximation property that quINR exploits.","marker":"Pérez-Salinas et al. 2020"},{"why":"Supplies the QReLU activation used in the output layer of quINR.","marker":"Parisi et al. 2022"},{"why":"Introduces the quantum neural network training paradigm that quINR adapts.","marker":"Farhi and Neven 2018"},{"why":"Source of the LiDAR range-image data used in the main experiment.","marker":"Geiger et al. 2013"},{"why":"Source of the Kodak color image used in the second experiment.","marker":"Eastman Kodak Company 1999"}],"fun_headline_variants":["Quantum INR codec gains 1.2 dB over classic codecs","Small quantum circuit improves INR rate-distortion","Quantum neural network boosts image compression efficiency","Quantum expressivity lifts implicit neural compression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison counts bits per pixel for quINR and COIN in the same way, including how trained parameters are quantized and whether the circuit description costs extra bits.","fun_headline_variants_meta":{"raw":{"variants":["Quantum INR codec gains 1.2 dB over classic codecs","Small quantum circuit improves INR rate-distortion","Quantum neural network boosts image compression efficiency","Quantum expressivity lifts implicit neural compression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000901,"raw_usage":{"total_tokens":3824,"prompt_tokens":834,"completion_tokens":2990,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":2931}},"tokens_in":450,"tokens_out":2990,"duration_ms":18856,"temperature":1.0,"reasoning_tokens":2931,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:50:58.283838+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the rate-distortion comparison with a fixed quantization scheme applied identically to COIN and quINR parameters, and include the bit cost of the circuit description and measurement settings; if quINR no longer reaches the reported PSNR at the same bits per pixel, the claimed gain is an accounting artifact rather than a coding gain.","supporting_citations":[{"cited_title":"W.; and an an, A","cited_arxiv_id":null,"evidence_quote":"Supplies the COIN baseline that quINR must beat and defines the INR-based image compression setting."},{"cited_title":"W.; and Doucet, A","cited_arxiv_id":null,"evidence_quote":"Extends COIN across modalities and anchors the INR compression framework the paper builds on."},{"cited_title":"D.; and Aspuru-Guzik, A","cited_arxiv_id":null,"evidence_quote":"Provides the parameterized entangling circuit used in the QNN layers and the expressibility argument."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the data re-uploading trick and universal approximation property that quINR exploits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the QReLU activation used in the output layer of quINR."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the LiDAR range-image data used in the main experiment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the Kodak color image used in the second experiment."}],"review_version":1}