{"id":"35ffbd39-efbd-4291-bf33-a8074fc44eee","arxiv_id":"2412.04008","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A complex-valued S-LISTA network for harmonic retrieval is converted to a spiking neural network with a two-neuron Few Spikes scheme and deployed on SpiNNaker2, claiming but not cleanly supporting a five-fold power-efficiency gain.","lead":"This paper converts a learned sparse-recovery network for harmonic retrieval into a spiking neural network and runs it on the SpiNNaker2 neuromorphic chip, reporting much lower power draw than an embedded GPU. The headline five-fold power-efficiency gain is undermined by an internal inconsistency in the power measurements.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The FS neuron pair cannot represent the complex soft-threshold required by S-LISTA, so the claimed conversion fidelity—and the 'moderate performance loss' part of the headline—is unsupported.","rationale":"I agree with the reader's weakest assumption: the two-neuron FS representation is separable in Re(s) and Im(s), and this is the most load-bearing issue for the paper's central method. The complex soft-threshold of Eq. (2) is non-separable, so the conversion pipeline as written cannot faithfully implement S-LISTA. The paper's Fig. 2 does not test the actual activation class used in the experiments, so the empirical approximation claim is not established for the relevant function. The power-accounting inconsistency in Section IV (two SpiNNaker2 spikes at about 25 mW and 225 mW cannot average to the implied 254.35 mW) is also serious and should be corrected, but even a corrected power ratio would not rescue the conversion claim. The paper may describe a useful separable approximation, but its headline claim of converting complex-valued S-LISTA to SNNs is not supported. The REJECT verdict stands.","tokens_in":8396,"tokens_out":8064,"duration_ms":83323,"concrete_test":"Solve Eq. (15) for f(s) = s·max(1 - 1/|s|, 0) on B = {s in C : |Re(s)|, |Im(s)| <= 1} using the paper's FS pair (Eqs. 10–14) and the reported SGD settings (learning rate 1e-2, stop at 5e-4), for K = 10, 20, 30. If the optimized error stays above the paper's threshold and does not decrease to zero with K, the 'arbitrary complex activation' claim is refuted. A stronger analytic check: for any FS output, Re(f_hat(s)) is constant in Im(s), but for fixed Re(s) = 0.5, the true Re(S_1)(s) changes from 0 at Im(s) = 0 to about 0.053 at Im(s) = 1, which no g(Re(s)) can match.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Contribution 1 claims the two-neuron FS scheme (Eqs. 10–14) approximates arbitrary complex activations on a bounded domain. However, Eq. (14) has the form f_hat(s) = g(Re(s)) + j·h(Im(s)): the real neuron receives only Re(s) and the imaginary neuron only Im(s), with no cross-coupling term in Eqs. (10)–(13). The activation needed in S-LISTA's Eq. (6) for the l1 problem Eq. (2) is the complex soft-threshold S_alpha(s) = s·max(1 - alpha/|s|, 0), whose real part Re(s)·max(1 - alpha/|s|, 0) depends on Im(s) through |s| whenever |s| < alpha. Hence S_alpha is not in the representable class. Fig. 2 verifies only the real soft-threshold (Eq. 5 with real u) and f(s) = s^2; it never tests a non-separable complex activation. Unless unstated coupling or preprocessing is added, the conversion pipeline does not implement S-LISTA, and the measured 'moderate performance loss' is not a loss against the intended algorithm.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a conversion pipeline that maps the complex-valued, deep-unrolling S-LISTA algorithm for multidimensional harmonic retrieval into spiking neural networks (SNNs) using an extension of Few Spikes (FS) coding. A complex-valued two-neuron FS model is introduced, and a full ANN-to-SNN conversion is implemented on the SpiNNaker2 neuromorphic board. The authors report support-recovery performance on a 1-D harmonic retrieval task and power measurements comparing SpiNNaker2 with an NVIDIA Jetson Xavier, claiming an almost five-fold power efficiency gain at moderate performance loss.","tokens_in":8649,"tokens_out":5974,"duration_ms":60091,"significance":"If the claims were sound, the paper would contribute a useful step toward deploying complex-valued sparse-recovery algorithms on neuromorphic hardware, and the hardware demonstration on SpiNNaker2 is a genuine practical effort. The paper includes real measurements, a plausible pipeline for mapping trained convolutional layers to SNNs, and a clear comparison against FISTA. However, the central technical claims are undermined by a representational limitation in the proposed complex FS neuron and by inconsistencies in the power-efficiency numbers. The paper is therefore not currently publishable in its present form.","major_comments":[{"comment":"The claim in Contribution 1 that the two-neuron FS model can approximate arbitrary complex-valued activations on a bounded domain is not supported and appears to be false. In Eqs. (10)-(13), the real neuron's state evolves using only the real part of the input and the imaginary neuron's state using only the imaginary part; the output in Eq. (14) is thus of the form f_hat(s) = g(Re(s)) + j h(Im(s)). This class cannot represent the complex soft-threshold used in S-LISTA, S_alpha(s) = s * max(1 - alpha/|s|, 0), whose real part depends on Im(s) through |s|. The approximation results in Fig. 2 test only real-valued functions, so the conversion fidelity for the actual S-LISTA activation is not established. This is a load-bearing issue because the paper's entire conversion pipeline is justified by this approximation claim.","section":"Section III-A, Eqs. (10)-(14)"},{"comment":"The reported power numbers are internally inconsistent. The text states that the Jetson mean is 1022.35 mW and that SpiNNaker2 consumes almost 768 mW less, which implies a SpiNNaker2 mean of approximately 254.35 mW. Yet the described PDF 'consist[s] of two spikes, one centered around 25 mW, the other one around 225 mW.' If the distribution were concentrated at these two modes, its mean could not exceed 225 mW unless there is additional unquantified mass above 225 mW. Moreover, the ratio 1022.35 / 254.35 is approximately 4.02, not 'almost 5×'; even comparing against 225 mW gives only about 4.54×. The 'almost 5×' claim in the abstract and conclusions is therefore not supported by the reported measurements.","section":"Section IV, Fig. 4 and power comparison"},{"comment":"The support recovery error (SRE) definition is ill-posed as written. The expression inside the cardinality braces is a set of pairs (l, l') with l in supp(b_hat_d) and l' in supp(b_d), but the intended quantity appears to be the number of mismatched support indices normalized by the true support size, e.g., |supp(b_hat_d) Δ supp(b_d)| / |supp(b_d)|. As written, the metric would behave very differently and could inflate or distort the reported performance. This makes the quantitative support-recovery results in Fig. 3 difficult to interpret.","section":"Section IV, Eq. (17)"}],"minor_comments":[{"comment":"There is a mismatch between the text and the figure caption: Section III-B says FS Id is obtained for f(s) = s, while Fig. 2 is captioned 'FS neuron approximating Eq. (5) (α = 1) and f(s) = s^2.' Please clarify which functions were actually approximated and whether they were evaluated on real or complex inputs.","section":"Section III-B and Fig. 2"},{"comment":"The phrase 'employing a bigger β' is ambiguous: β is earlier defined as the learning rate in line 'a learning rate of β = 0,009', and it is unclear how a larger learning rate mitigates approximation artifacts. The authors likely mean a larger number of timesteps K or a different hyperparameter; this should be stated explicitly.","section":"Section IV, first paragraph"},{"comment":"The title and abstract emphasize 'multidimensional' harmonic retrieval, but the experimental evaluation is restricted to M = 1 dimension (N = 64, L = 128). The authors do note the memory limitation, but the mismatch between the claimed scope and the demonstrated scope should be acknowledged more clearly in the abstract or conclusions.","section":"General"}],"recommendation":"reject","confidential_remarks":"The paper addresses an interesting and timely problem, and the hardware implementation effort is commendable. However, the central conversion-fidelity claim rests on a representational assumption that the proposed complex FS neuron cannot satisfy for the actual S-LISTA activation. The power-efficiency headline is also not supported by the numbers in the paper. These are not merely presentation issues; they affect the validity of the two main contributions. If the authors can introduce a genuinely coupled complex FS mechanism or restrict the claims to separable activations, and if the power measurements are corrected and reported consistently, a substantially revised version might be reconsidered. For the current manuscript, I cannot recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nRead the Andrei et al. preprint on converting deep-unrolled harmonic retrieval (S-LISTA) to SNNs on SpiNNaker2. The honest take: this is a genuine engineering attempt with a novel combination—extending Few Spikes conversion to complex-valued layers and running it on actual neuromorphic hardware with power measurements. The on-chip comparison against a Jetson Xavier, the support recovery results, and the discussion of quantization effects are concrete and useful. If the conversion worked as claimed, the efficiency result would matter for low-power radar and 6G sensing.\n\nThe soft spots are load-bearing. The complex FS neuron from Eqs. (10)–(14) is separable: the output is f_hat(s) = g(Re(s)) + j·h(Im(s)). The complex soft-threshold used in S-LISTA is not separable—its real part depends on Im(s) through |s|. So the \"arbitrary complex-valued activations\" claim in Contribution 1 is false, and the conversion as described does not implement S-LISTA's activation. Fig. 2 does not resolve this: it tests a real soft-threshold and a scalar square, not the coupled complex soft-threshold. The paper needs either a precise statement of the representable class plus an approximation error bound for the actual activation, or a modified neuron model with cross-coupling.\n\nThe power numbers are also internally inconsistent. A PDF with spikes at ~25 and ~225 mW cannot have a mean near 254 mW (which is what 1022.35 − 768 implies) without a long tail that the text doesn't mention. And 1022.35/254.35 is about 4.0×, not \"almost 5×.\" That overstatement should be fixed by reporting the actual ratio and the distribution's tail.\n\nScope is also narrow: M = 1 only, despite the multidimensional title, and no code or data are released. That limits reproducibility but isn't fatal in itself.\n\nMy verdict: the paper deserves a serious referee, because the problem is relevant and the engineering work is real, but as written it would need major revision. The representability issue alone undermines the headline claim. I'd send it to review with the expectation of heavy revision, and I'd want the activation class and the power statistics corrected before publication.\n\nBest","headline":"A real engineering demo of complex FS-to-SNN conversion on SpiNNaker2, undermined by an overclaimed activation model and an inconsistent power ratio.","tokens_in":9266,"tokens_out":3415,"would_cite":false,"duration_ms":30651,"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":"The paper claims that complex-valued S-LISTA networks for multidimensional harmonic retrieval can be converted to spiking neural networks that run on SpiNNaker2 at nearly five times the power efficiency of an embedded GPU, with moderate…","keywords":["multidimensional harmonic retrieval","deep unrolling","spiking neural networks","ANN-to-SNN conversion","Few Spikes coding","complex-valued neural networks","SpiNNaker2","sparse recovery"],"falsifier":"Measure the approximation error of the FS pair on a coupled complex function such as $s \\cdot \\max(1 - \\alpha/|s|, 0)$ on a bounded disk; if the error does not fall toward zero as the number of timesteps $K$ grows and the sampling of the domain refines, the claim that the neuron model approximates arbitrary complex-valued activations fails, and the conversion fidelity for S-LISTA's soft-threshold is not established.","tokens_in":8121,"feed_emoji":"🧠","tokens_out":4717,"duration_ms":39127,"temperature":0.7,"pith_summary":"The paper tries to show that a deep-unrolling sparse-recovery algorithm for multidimensional harmonic retrieval (MHR) can be moved from GPUs onto neuromorphic hardware without losing its practical usefulness. It converts the complex-valued S-LISTA network into a spiking neural network (SNN) by extending Few Spikes coding to complex-valued activations, then maps the SNN onto the SpiNNaker2 board. Measured against an NVIDIA Jetson Xavier, the SpiNNaker2 implementation consumes about 768 mW less on average, an almost five-fold power-efficiency gain, while support-recovery error stays comparable to FISTA and close to S-LISTA for few sources. This matters because energy-hungry GPU inference is a bottleneck for radar and wireless sensing systems that need single-snapshot harmonic estimation.","feed_headline":"Spiking neural net runs harmonic retrieval at a fifth of GPU power","feed_subtitle":"Mapping S-LISTA onto SpiNNaker2 keeps support recovery near FISTA while cutting power by 80 percent.","key_machinery":"The machinery is a complex-valued extension of Few Spikes coding: two FS neurons, one driven by the real part and one by the imaginary part, whose membrane dynamics release a serialized binary code over K timesteps, and whose weighted sum (Eq. 14) approximates the desired activation. Each activation in S-LISTA—the identity encoder and the complex soft-threshold for each layer—is replaced by its FS approximation trained by gradient descent on the squared error (Eq. 15). The SpiNNaker2 mapping uses pipelined receive-send cycles and 4-bit quantized projections for the convolution weights.","core_discovery":"The central claim is that the Few Spikes (FS) neuron model, extended to a pair of coupled real-and-imaginary neurons, can approximate complex-valued activations well enough on a bounded domain that an entire complex-valued convolutional network solving MHR can be converted to an SNN and run on SpiNNaker2. The converted network achieves support recovery similar to S-LISTA for one to a few sources and beats FISTA, while drawing roughly one fifth of the GPU's power. The authors attribute residual performance loss to 4-bit weight quantization and to the residual errors in the FS neuron approximations.","pith_inferences":["The authors do not draw this inference, but if the FS pair can only represent functions that are sums of a function of the real part and a function of the imaginary part, then coupled complex activations such as the standard complex soft-threshold cannot be represented exactly; a corrected rule would need cross-coupled membrane dynamics or an amplitude-phase decomposition.","The paper's own observation that quantization is the main bottleneck points to a testable extension: retraining with quantization-aware training, or using more than 4 bits for projections, should close much of the gap between the SNN and the original CNN.","The same conversion recipe could be applied to other complex-valued unrolled networks, such as learned ISTA variants for two- and three-dimensional MHR, once the memory constraints of larger neuromorphic boards are lifted.","The two-mode power distribution on SpiNNaker2, caused by dynamic voltage and frequency scaling, suggests that comparing only average power may overstate the benefit in continuously loaded scenarios and understate it in bursty inference workloads."],"forward_implications":["If the conversion is sound, single-snapshot MHR estimation can be deployed on neuromorphic hardware with near-GPU accuracy and roughly one fifth the power draw.","The same pipeline extends S-LISTA-style deep-unrolling networks beyond image classification to complex-valued regression tasks in signal processing.","Because the converted SNN outperforms 100-iteration FISTA with only five layers, the conversion preserves the convergence benefit of deep unrolling while reducing execution cost.","For one to five sources, the converted SNN's support recovery remains comparable to S-LISTA and degrades gracefully as the number of sources grows.","The measured power-efficiency advantage makes neuromorphic deployment a credible option for bandwidth- and energy-constrained sensing systems."],"supporting_citations":[{"why":"Supplies the Few Spikes coding scheme that the complex-valued neuron model extends.","marker":"[13]"},{"why":"Defines S-LISTA, the deep-unrolling network whose layers and activations are converted to SNNs.","marker":"[14]"},{"why":"Describes SpiNNaker2, the neuromorphic board used for mapping and power measurement.","marker":"[4]"},{"why":"Provides FISTA and the soft-thresholding recursion that S-LISTA unrolls and that the SNN must approximate.","marker":"[16]"},{"why":"Shows DNN-to-SNN conversion restricted to real-valued ReLU networks, motivating the complex-valued extension.","marker":"[9]"},{"why":"Supplies the tensor/Khatri-Rao system model for multidimensional harmonic retrieval used to generate data.","marker":"[15]"},{"why":"Motivates deep unrolling as the interpretable low-complexity framework to which S-LISTA belongs.","marker":"[3]"}],"fun_headline_variants":["Spiking harmonic retrieval at 1/5 GPU power","Deep-unrolled MHR goes neuromorphic","SNN solves MHR near CNN accuracy, at fifth power","Few Spikes neuron model powers harmonic retrieval on SpiNNaker2","Complex-valued SNN beats FISTA at 5x less power"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two-neuron FS pair can approximate any complex activation that appears in the converted network; the pair is a sum of a function of the real part and a function of the imaginary part, so any activation whose real and imaginary outputs each depend on both input components, such as the standard complex soft-threshold, cannot be represented exactly.","fun_headline_variants_meta":{"raw":{"variants":["Spiking harmonic retrieval at 1/5 GPU power","Deep-unrolled MHR goes neuromorphic","SNN solves MHR near CNN accuracy, at fifth power","Few Spikes neuron model powers harmonic retrieval on SpiNNaker2","Complex-valued SNN beats FISTA at 5x less power"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1331,"prompt_tokens":884,"completion_tokens":447,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":362}},"tokens_in":500,"tokens_out":447,"duration_ms":5318,"temperature":1.0,"reasoning_tokens":362,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:52:17.585857+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the approximation error of the FS pair on a coupled complex function such as $s \\cdot \\max(1 - \\alpha/|s|, 0)$ on a bounded disk; if the error does not fall toward zero as the number of timesteps $K$ grows and the sampling of the domain refines, the claim that the neuron model approximates arbitrary complex-valued activations fails, and the conversion fidelity for S-LISTA's soft-threshold is not established.","supporting_citations":[{"cited_title":"Optimized Spiking Neurons Can Classify Images With High Accuracy Through Temporal Coding With Two Spikes,","cited_arxiv_id":null,"evidence_quote":"Supplies the Few Spikes coding scheme that the complex-valued neuron model extends."},{"cited_title":"Structured lista for multidi- mensional harmonic retrieval,","cited_arxiv_id":null,"evidence_quote":"Defines S-LISTA, the deep-unrolling network whose layers and activations are converted to SNNs."},{"cited_title":"Con- version of Continuous-Valued Deep Networks to Efficient Event-Driven Networks for Image Classification,","cited_arxiv_id":null,"evidence_quote":"Shows DNN-to-SNN conversion restricted to real-valued ReLU networks, motivating the complex-valued extension."},{"cited_title":"Tensor algebra and multidimensional harmonic retrieval in signal processing for mimo radar,","cited_arxiv_id":null,"evidence_quote":"Supplies the tensor/Khatri-Rao system model for multidimensional harmonic retrieval used to generate data."}],"review_version":1}