{"id":"a58429e0-7a7c-4735-b9d8-7500f4933425","arxiv_id":"2502.00074","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An SNN made by swapping ReLU for LIF neurons does 4D radar 3D detection on K-Radar at about RTNH accuracy with an estimated 78% lower energy.","lead":"This paper converts an existing 4D radar object detection network, RTNH, into a spiking neural network and reports similar detection accuracy at an estimated 78% lower energy cost. It also introduces a multi-step inference scheme that feeds progressively lower-density point clouds through the network.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Energy reduction claim relies on unverifiable op counts and a cost model that omits memory traffic; Table I's factor-of-ten arithmetic error shows the accounting is unreliable, so the 78% headline is not established.","rationale":"The reader's verdict identified the ideal 45nm CMOS cost model and the factor-of-ten error as the weakest assumption. I agree that the cost model is load-bearing, but I extend the concern in two ways: the MAC/AC counts themselves cannot be produced by the stated tool (thop), and no source code for SpikingRTNH is released despite the claim that all codes are available. This makes the quantitative accounting unverifiable from the manuscript alone. The Factor-of-ten error, while it does not change the 78% ratio, indicates careless handling of the energy figures. I also note that the memory-access energy, which the reader mentioned, is particularly important for 3D convolutions on dense grid data, where weight and activation movement often dominate arithmetic cost. If memory traffic is included, the efficiency advantage of AC over MAC may shrink considerably. My concern is therefore a refinement of the reader's: the headline energy reduction may be an artifact of an incomplete model and non-reproducible op counts. The paper's other contributions, such as the first SNN for 4D radar and the BTI inference mechanism, remain plausible and are supported by the T=1 and T=3 performance numbers, but the energy claim as stated is not robust. I recommend keeping the conditional verdict, with requirements that the authors release the actual SNN code, justify the AC counting methodology, and provide a memory-aware energy estimate or a clear statement that the figures are ideal arithmetic-energy estimates only. This does not change the reader's verdict, so verdict_should_be is UNCHANGED.","tokens_in":9217,"tokens_out":12118,"duration_ms":117165,"concrete_test":"Obtain or reimplement the SpikingRTNH code and independently profile the exact architecture with both a dense MAC counter (thop) and a spike-aware AC counter that tracks active input positions per timestep. Verify that the resulting MAC/AC counts match Table I to within 10%. Then recompute total energy with a memory-aware model that adds per-operation weight and activation SRAM/DRAM access energies (e.g., Horowitz's figures of ~10 pJ per SRAM read and higher for DRAM) in addition to the 4.6 pJ/0.9 pJ arithmetic costs. If the AC counts cannot be reproduced or the corrected energy ratio (SNN vs RTNH) drops below 50%, the 78% claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of a 78% energy reduction is a product of two factors: the ratio of MACs to ACs and the assumed per-operation costs (4.6 pJ/MAC vs 0.9 pJ/AC). Both factors are questionable. First, the paper states that the thop library was used to measure computational requirements, but thop counts MACs for dense tensor operations and cannot produce AC counts that depend on spike timing and sparsity. The reported 48.6G AC and 137G AC counts therefore cannot be produced by the cited tool, and no code for the spiking implementation is released (the provided link is the K-Radar dataset repository, not SpikingRTNH). Second, the per-operation costs are ideal 45nm CMOS values from Horowitz that exclude memory access, data movement, and control logic. For a 3D convolutional network on a dense radar tensor, weight reads and activation movement typically dominate energy; a single SRAM read can cost tens of pJ, far exceeding the 4.6 pJ MAC or 0.9 pJ AC. If memory traffic is included, the gap between MAC and AC narrows substantially. Third, the absolute energy values in Table I contain a factor-of-ten error: 156G * 4.6 pJ = 0.718 J, not 7.16 J, and the BTI row similarly should read 0.158 J, not 1.58 J. While the ratio still gives roughly 78% after correction, the arithmetic error underscores that the quantitative accounting was not carefully checked. Together, these issues mean the headline energy reduction, as stated, does not describe any real, deployable system and may be a substantial overestimate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SpikingRTNH, the first spiking neural network (SNN) for 3D object detection from 4D Radar data, obtained by replacing ReLU activations in the RTNH architecture with leaky integrate-and-fire (LIF) neurons. It also introduces biological top-down inference (BTI), which sequentially processes successively lower-density point clouds at successive time steps. On the K-Radar test set, SpikingRTNH with BTI is reported to achieve 51.1% AP3D and 57.0% APBEV, comparable to RTNH (50.7% and 56.5%), while reducing inference energy by 78% (1.58 J/frame vs 7.16 J/frame). The paper claims this demonstrates the viability of SNNs for energy-efficient 4D Radar perception.","tokens_in":9552,"tokens_out":5770,"duration_ms":51231,"significance":"If the energy-efficiency claim were established, the paper would make a meaningful contribution by showing that SNNs can match ANN performance on a challenging 4D Radar detection task while consuming substantially less energy. The detection results themselves are plausible and the BTI idea of progressive density reduction is simple but interesting. However, the central quantitative claim is not currently supported: the reported energy numbers contain a factor-of-ten arithmetic error, the operation counts are not reproducible from the released materials, and the cost model omits memory traffic and assumes idealized neuromorphic hardware. The paper also appears to select the BTI hyperparameters on the test set. These issues are fixable but require substantial revision; the paper does not currently establish its headline result.","major_comments":[{"comment":"The energy values in Table I and Section IV-B are a factor of 10 too high. For RTNH, 156G MACs at 4.6 pJ/MAC gives 0.718 J, not 7.16 J. For SpikingRTNH with BTI, 7.43G MACs at 4.6 pJ plus 137G ACs at 0.9 pJ gives 0.157 J, not 1.58 J. For the T=1 row, 2.48G MACs plus 48.6G ACs gives 0.055 J, not 0.551 J. The relative reductions (78% and 92%) are unchanged by this correction, but the absolute numbers are wrong, and this arithmetic error calls into question the care with which the energy analysis was performed.","section":"IV-B, Table I"},{"comment":"The paper states that the thop library was used to measure computational requirements, but thop counts dense MAC/FLOP operations and cannot produce spike-dependent AC counts that depend on the timing and sparsity of spikes. The reported AC counts (48.6G and 137G) are therefore not obtainable from the cited tool. Moreover, the link provided for 'All codes are available' points to the K-Radar dataset repository, not to a SpikingRTNH implementation, so the reported MAC/AC counts are not reproducible from the manuscript or the released code.","section":"IV-A, IV-B"},{"comment":"The energy comparison assigns 4.6 pJ per MAC and 0.9 pJ per AC using ideal 45nm CMOS estimates from Horowitz, but this model omits memory access, data movement, and control logic. For 3D convolutional layers operating on dense radar tensors, memory traffic typically dominates energy, and a single SRAM read can cost tens of pJ, far exceeding the per-operation costs used here. In addition, the actual PyTorch/GPU implementation executes MAC operations in the SNN simulation, so the claimed savings would only be realized on dedicated neuromorphic hardware, which is not described or analyzed. The 78% reduction is thus an idealized upper bound, not a property of any deployed or simulated system.","section":"IV-B, Energy Analysis"},{"comment":"The ablation study in Table II appears to be performed on the K-Radar test set, and the BTI hyperparameters r=80% and T=3 are selected based on these test-set results. Tuning on the test set can inflate the reported performance (51.1% AP3D, 57.0% APBEV). A separate validation split should be used for hyperparameter selection, with the test set reserved for a single final evaluation.","section":"IV-C, Table II"},{"comment":"The BTI procedure feeds different point clouds (with lower density at each time step) into the same network at successive time steps, and the final detection result is taken from the last time step's features. This is not temporal integration of a single input in the usual SNN sense; the LIF membrane potential is carried across different inputs, which is not formally justified. The interaction between the SNN temporal dynamics and the changing input is unclear, and the paper does not provide an ablation separating the effect of multi-step processing from the effect of density reduction.","section":"III-B, Section IV-C"}],"minor_comments":[{"comment":"The statement 'All codes are available at https://github.com/kaist-avelab/k-radar' is misleading, as the link points to the K-Radar dataset repository rather than to the proposed SpikingRTNH implementation; please either provide the actual code or state that it is not yet released.","section":"Abstract, Footnote"},{"comment":"The decay factor lambda is set to 0.25 in the text with a citation, but it is unclear whether this value is used in all experiments and whether it is optimized or fixed; please clarify.","section":"III-A, Eq. (7)"},{"comment":"The sentence 'SNNs activate neurons only when spike events occur' is an oversimplification, since the LIF membrane potential is updated at every time step and the reported energy model charges 0.9 pJ per AC regardless of sparsity; consider wording it more precisely.","section":"II-A"},{"comment":"The ablation table would be easier to interpret if it included a row for RTNH or a column indicating the relative energy reduction for each configuration, since the absolute MAC and AC counts alone do not make the efficiency trade-off transparent.","section":"IV-C, Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper's code availability statement is misleading, and the energy analysis needs substantial revision before publication. The topic is within the journal's scope, but the central claim of 78% energy reduction is not currently established. If the authors can correct the arithmetic, provide reproducible operation counts, and revisit the cost model with a more realistic treatment of memory and hardware, the paper could become a useful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is showing that a straightforward LIF-for-ReLU swap on RTNH yields a spiking network that matches its ANN counterpart on K-Radar at T=1 (48.1 vs 50.7 AP3D; 55.3 vs 56.5 APBEV). That is genuinely new: no prior SNN has been applied to 4D radar object detection. The BTI heuristic—processing point clouds from high to low density over multiple time steps—is simple but works: with T=3 and r=80% they get 51.1/57.0, slightly better than RTNH. The paper is clearly written and the ablation table is informative.\n\nBut the energy story is soft in several places. The factor-of-ten error in Table I is the most concrete: 156G MAC at 4.6 pJ/MAC is 0.72 J, not 7.16 J, and the SNN rows have the same issue. The 78% relative reduction survives the correction, but the absolute J/frame numbers are wrong, which makes one wonder how carefully the accounting was checked. Second, the AC counts cannot come from thop, which counts MACs; spike-dependent AC counts require actually running the spiking network and counting spikes. The paper does not describe how these numbers were obtained. Third, the cost model uses ideal 45nm arithmetic costs (4.6 pJ/MAC, 0.9 pJ/AC) and omits memory traffic. For a 3D convolutional network on dense radar tensors, weight reads and activation movement typically dominate; a single SRAM read is tens of pJ. Including memory would narrow the MAC/AC gap and probably shrink the 78% claim substantially. Also, the BTI hyperparameters (r and T) were selected on the test set—Table II is test-set performance—so the 51.1/57.0 result is mildly overfit. There is no ablation that isolates BTI from simply running T=3 on the same full-density input or on a random subsample, so we cannot tell what BTI itself adds.\n\nOn the positive side, the central feasibility claim—that an SNN can perform 4D radar detection at accuracy comparable to an equivalent ANN—is likely defensible. The T=1 results alone support it. The paper is not circular or incoherent; it is an engineering contribution with an overreaching energy analysis. I would send it to peer review, not desk reject, because it opens a plausible new direction (SNN-based radar perception) and the issues are fixable. The referee should demand corrected energy numbers, a clear explanation of how AC counts were measured, memory-aware or clearly labeled estimates, and a proper T=3 baseline without BTI. The paper will also need a real code link if the authors claim \"all codes are available.\"\n\nFor anyone working on SNN perception or radar detection, this is a useful data point. I would not cite the energy numbers as they stand, but I would cite the T=1 accuracy result if I needed a compact SNN-on-radar baseline.","headline":"First SNN for 4D radar detection, but the headline 78% energy saving rests on a factor-of-ten arithmetic error and a cost model that omits memory traffic, so the energy claims are not established.","tokens_in":10121,"tokens_out":2734,"would_cite":false,"duration_ms":27177,"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":"SpikingRTNH claims to be the first spiking neural network for 4D Radar 3D object detection, achieving detection accuracy comparable to the ANN baseline while cutting estimated per-frame energy by 78%.","keywords":["spiking neural network","4D radar","3D object detection","LIF neuron","energy efficiency","autonomous driving","top-down inference"],"falsifier":"Recompute the energy comparison using the paper's own unit costs and a deployment model that includes memory traffic: 156G MACs at 4.6 pJ equals 0.72 J, not 7.16 J, and if the SNN's real measured energy on event-driven hardware is not roughly one-fifth of the ANN's total energy, the 78% claim is refuted.","tokens_in":8987,"feed_emoji":"📡","tokens_out":8500,"duration_ms":64894,"temperature":0.7,"pith_summary":"The paper aims to establish that spiking neural networks (SNNs) can replace conventional artificial neural networks for 4D Radar 3D object detection in autonomous vehicles, cutting energy use without sacrificing accuracy. It converts the RTNH radar detector into SpikingRTNH by swapping every ReLU activation for a leaky integrate-and-fire (LIF) spiking neuron, and adds a biological top-down inference (BTI) schedule that feeds successively sparser point clouds across inference time steps. On the K-Radar test set, SpikingRTNH with BTI reaches 51.1% AP 3D and 57.0% AP BEV, closely matching the ANN baseline of 50.7% and 56.5%, while the paper estimates per-frame energy drops from 7.16 J to 1.58 J, a 78% reduction. A single-time-step version uses 92% less energy at a small accuracy cost. If correct, this makes spike-based radar perception an energy-efficient option for continuous on-vehicle sensing.","feed_headline":"Cut radar-detection energy 78% with spiking neurons.","feed_subtitle":"First spiking network for 4D radar detection: 78% lower energy, nearly equal accuracy.","key_machinery":"The central machinery is the leaky integrate-and-fire (LIF) neuron, a spiking unit that accumulates incoming currents into a membrane potential and fires an output spike when the potential crosses a threshold, leaking its potential over time. Replacing every ReLU activation in RTNH with LIF neurons converts each multiply-accumulate (MAC) into an accumulate (AC) operation, which the paper prices at 0.9 pJ versus 4.6 pJ per MAC in 45-nm CMOS. Because the spike function's Heaviside step is non-differentiable, training uses a scaled tanh surrogate gradient and a single simulation time step, while inference runs multiple time steps. The second mechanism is biological top-down inference (BTI), which at each time step keeps the top r% of points by radar power, feeding the network a sparser, lower-noise point cloud; the final detection comes from the last time step's feature maps.","core_discovery":"SpikingRTNH is the first spiking neural network for 3D object detection from 4D Radar data. The central discovery is that a direct ReLU-to-LIF conversion of the RTNH architecture, trained with a surrogate-gradient approximation of the spike derivative and run at inference over multiple time steps with progressively sparser point-cloud inputs, produces detection accuracy comparable to the original ANN while drastically lowering the estimated energy cost per frame. With BTI (r = 80% density retention, T = 3 time steps), the network reports 51.1% AP 3D and 57.0% AP BEV on the K-Radar test set versus 50.7% and 56.5% for RTNH, and the energy estimate falls from 7.16 J/frame to 1.58 J/frame, a 78% reduction; without BTI (T = 1) it falls to 0.551 J/frame, a 92% reduction, at 48.1% AP 3D. The paper interprets these numbers as demonstrating that spike-based computing is viable for dense 4D Radar perception in autonomous driving, including under adverse weather.","pith_inferences":["The conversion recipe and density-schedule idea could plausibly transfer to other point-cloud detectors beyond RTNH, such as pillar- or voxel-based networks on LiDAR, although the paper only demonstrates it for one radar architecture.","On true event-driven neuromorphic hardware, the real energy saving might be larger than the paper's 78% estimate, because silent neurons in sparse lower-density clouds consume near-zero dynamic energy; the paper's ideal-CMOS accounting does not count such event-driven effects.","A direct test of BTI's mechanism would sweep T beyond 3 and r below 50%: if accuracy keeps rising with T, the benefit is temporal integration; if it saturates or drops, the benefit is specifically noise pruning at the 80% retention level."],"forward_implications":["A radar perception network can run on spike-based hardware with roughly one-fifth the per-frame energy of the ANN version, making continuous 4D Radar processing more feasible for onboard deployment.","The energy-accuracy trade-off is controlled by the number of inference time steps T and the density retention ratio r: T = 1 trades 2.6 points of AP 3D for a 92% energy cut, and BTI at T = 3 restores the accuracy for a 78% cut.","BTI's density schedule improves over single-step SNN inference by 3.0% AP 3D, indicating that temporally processing progressively sparser radar clouds is itself a source of accuracy gain.","The same ReLU-to-LIF conversion recipe transfers from the camera and LiDAR SNN literature to 4D Radar tensor inputs, broadening the input modalities for which spike-based detection is reported."],"supporting_citations":[{"why":"Supplies the RTNH architecture and K-Radar dataset that SpikingRTNH converts and is compared against.","marker":"[2]"},{"why":"Provides the direct-training method, discrete LIF formulation, and single-time-step training used for the conversion.","marker":"[16]"},{"why":"Supplies the surrogate-gradient function and hyperparameters (beta=5.0) and the multi-step inference approach.","marker":"[34]"},{"why":"Gives the 45-nm CMOS per-operation energy values (4.6 pJ per MAC, 0.9 pJ per AC) used to compute the 78% saving.","marker":"[21]"},{"why":"Supports the premise that spike-based accumulation is more energy-efficient than ANN MAC operations.","marker":"[11]"}],"fun_headline_variants":["First spiking neural network for 4D radar detection cuts energy 78%","4D radar object detection with spiking network: 78% energy savings","SNN for radar matches ANN accuracy at 78% lower energy","Spiking neurons bring 4D radar detection down to 1.58 J per frame"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's 78% energy saving rests on the assumption that every operation in the spiking network costs an idealized 0.9 pJ accumulation and every operation in the ANN costs 4.6 pJ, with no additional energy for memory access, data movement, or control logic in either network.","fun_headline_variants_meta":{"raw":{"variants":["First spiking neural network for 4D radar detection cuts energy 78%","4D radar object detection with spiking network: 78% energy savings","SNN for radar matches ANN accuracy at 78% lower energy","Spiking neurons bring 4D radar detection down to 1.58 J per frame"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000665,"raw_usage":{"total_tokens":3068,"prompt_tokens":1008,"completion_tokens":2060,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":1975}},"tokens_in":624,"tokens_out":2060,"duration_ms":30889,"temperature":1.0,"reasoning_tokens":1975,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T21:55:27.285204+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the energy comparison using the paper's own unit costs and a deployment model that includes memory traffic: 156G MACs at 4.6 pJ equals 0.72 J, not 7.16 J, and if the SNN's real measured energy on event-driven hardware is not roughly one-fifth of the ANN's total energy, the 78% claim is refuted.","supporting_citations":[{"cited_title":"Direct training for spiking neural networks: Faster, larger, better,","cited_arxiv_id":null,"evidence_quote":"Provides the direct-training method, discrete LIF formulation, and single-time-step training used for the conversion."},{"cited_title":"Spik- ing pointnet: Spiking neural networks for point clouds,","cited_arxiv_id":null,"evidence_quote":"Supplies the surrogate-gradient function and hyperparameters (beta=5.0) and the multi-step inference approach."},{"cited_title":"Are snns really more energy-efficient than anns? an in-depth hardware-aware study,","cited_arxiv_id":null,"evidence_quote":"Supports the premise that spike-based accumulation is more energy-efficient than ANN MAC operations."}],"review_version":1}