{"id":"3cb6e7af-56be-4b88-888b-458a8da9b7d7","arxiv_id":"2411.14585","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"PointLCA-Net stores PointNet features in a dictionary and uses a spiking Locally Competitive Algorithm encoder-decoder to classify spatio-temporal event data, reporting up to 98.78% accuracy with lower estimated energy than prior spiking point-cloud networks.","lead":"This paper combines PointNet feature extraction with a spiking sparse-coding decoder, called PointLCA-Net, for recognizing event-based spatio-temporal signals on low-power hardware. It reports high accuracy on NMNIST, DVS128, and SHD with theoretically estimated energy savings, though the energy comparisons exclude part of the pipeline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 92% energy-saving claim excludes the PointNet/PointNet++ feature extractor that is part of the pipeline; adding its FLOPs may erase the advantage over SpikePoint.","rationale":"The reader's weakest assumption is that the LCA decoder's FLOP count and the RRAM crossbar energy fully represent the deployed system while PointNet is excluded. My stress test identifies the same load-bearing concern and sharpens it: the comparison in Table II is asymmetric because SpikePoint's energy includes the full spiking PointNet++ pipeline, whereas PointLCA-Net's reported energy omits the non-spiking PointNet feature extractor. This is not an internal inconsistency in the algorithm, but it directly undermines the headline 'approximately 92% energy reduction' claim as stated. The paper is otherwise coherent: the accuracy results are plausible, the method is reproducible in principle, and the LCA formulation is standard. The fix is not to reject the work but to require an end-to-end accounting of FLOPs and energy, or at least a stated assumption about the feature extractor's cost. Since the reader already assigned a CONDITIONAL verdict based on this issue, my read does not change the verdict; it confirms and concretizes the condition.","tokens_in":10094,"tokens_out":3413,"duration_ms":33593,"concrete_test":"Profile the actual PointNet and PointNet++ feature extractors on 1024-event DVS128 point clouds (e.g., with torchprofile) and add the measured FLOPs to Eq. 10 for K=100; recompute energy using the same 9.09e-14 J/FLOP figure. Then compare the resulting total against SpikePoint's 0.9 GFLOPs and 0.82 mJ dynamic plus 0.756 mJ static energy under the same data partition. If the total GFLOPs exceeds ~0.9 or total energy exceeds 1.576 mJ, the central savings claim fails. Ideally, also measure end-to-end energy on an RRAM crossbar or a neuromorphic simulator to validate the per-FLOP energy assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is energy efficiency, but Section IV.E states that the FLOP estimates 'exclude the FLOPs required for feature extraction by PointNet and PointNet++.' The PointLCA-Net pipeline still requires generating a 1024-dimensional point-cloud feature vector before the Exemplar LCA encoder can run, so Eq. 10 counts only LCA-decoder operations, not the full inference cost. Table II then compares the resulting 0.065 mJ for DVS128 against SpikePoint's 0.82 mJ dynamic plus 0.756 mJ static energy, which are full-pipeline numbers. This is not a like-for-like comparison: the omitted PointNet feature-extraction stage is a deep MLP over 1024 points (including T-Nets), and its FLOPs can be comparable to the reported 0.7 GFLOPs for the LCA stage. Adding a realistic PointNet cost would increase total energy by at least tens of percent and could materially shrink the claimed 92% reduction or even reverse it if PointNet++ is used. Additionally, applying Yao et al.'s 9.09e-14 J/FLOP, which was measured for MAC operations in a memristor CNN accelerator, to thresholding, neuron dynamics, and peripheral circuits is an unvalidated extrapolation. The energy-efficiency claim therefore rests on an incomplete accounting of the deployed system.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PointLCA-Net, a two-stage pipeline for spatio-temporal signal recognition on edge devices: a pre-trained PointNet/PointNet++ extracts a 1024-dimensional global feature vector from event-based point clouds, and an Exemplar LCA encoder-decoder then performs sparse coding and classification using a dictionary of stored training features. The method is evaluated on NMNIST, DVS128, and SHD, reporting top-1 accuracies up to 98.78%, 93.41%, and 78.46% respectively. The central claim is that PointLCA-Net achieves high accuracy while reducing energy consumption by approximately 92% compared to other spiking neural networks applied to point clouds, based on FLOP estimates and an RRAM crossbar energy model.","tokens_in":10360,"tokens_out":2252,"duration_ms":21865,"significance":"The idea of decoupling feature extraction from a neuromorphic sparse-coding classifier is conceptually attractive and the paper is, to its credit, explicit about the experimental setup, hyperparameters, and the fact that energy numbers are theoretical estimates. The work also extends the event-cloud formulation to NMNIST and SHD for the first time, which is a useful data point. The accuracy results are on external benchmarks and the core LCA-based classification pipeline is reproducible in principle from the equations given. However, the headline energy-efficiency claim depends on an incomplete accounting of the deployed system, so the significance of the paper as a hardware-energy contribution cannot be assessed until that accounting is corrected.","major_comments":[{"comment":"The energy comparison against SpikePoint is not like-for-like. Section IV.E explicitly states that the FLOP estimates 'exclude the FLOPs required for feature extraction by PointNet and PointNet++,' yet Table II compares the resulting 0.065 mJ for DVS128 against SpikePoint's full-pipeline dynamic (0.82 mJ) and static (0.756 mJ) energy. The abstract's claim of 'reducing energy consumption by approximately 92%' is therefore not supported for the full system. The feature-extraction stage (PointNet with input and feature T-Nets, MLP(64,64,64,128,1024) over 1024 points) performs a substantial number of MACs that must either be added to the PointLCA-Net total or explicitly excluded from the comparison with a clear statement that the comparison is LCA-stage-only. As written, this is a load-bearing omission for the central energy-efficiency claim.","section":"IV.E, Table II"},{"comment":"The sparsity estimate \\hat{M}, which drives the inference FLOP count, is inherited from reference [28] with the statement 'we have verified that this estimate is accurate.' No verification data for the three datasets used here (NMNIST, DVS128, SHD) is provided. Since the '99.54% reduction in computational effort' and all derived energy numbers scale linearly with \\hat{M}, the paper should report measured active-neuron fractions for each dataset and each feature extractor (PointNet vs PointNet++). If the fraction differs from 0.4%, the workload and energy conclusions change proportionally.","section":"IV.E, Eq. (10), Table I"},{"comment":"The energy per FLOP of 9.09e-14 J is taken from Yao et al. [37], which was measured for MAC operations in a memristor CNN accelerator. PointLCA-Net's LCA decoder includes thresholding (Eq. 6), membrane potential updates (Eq. 2), and peripheral input/output circuits (Section IV.F), not only MACs. Applying the same per-FLOP figure to all these operations is an unvalidated extrapolation. The paper should either model the hardware mapping more carefully (e.g., distinguishing analog MAC energy from digital control/peripheral energy) or clearly state that the reported energy is a MAC-only lower bound. Without this, the 92% reduction claim is not robust.","section":"IV.F, Section V"}],"minor_comments":[{"comment":"The equation is labeled 'F LOP s(Inf erenec)' — the word 'Inference' is misspelled; please correct the typo.","section":"IV.E"},{"comment":"The footnotes 'c' and 'd' state that the PointNet/PointNet++ [4] and SpikePoint accuracies use different class counts or data partitioning. Please move these qualifications into the main text or the table caption so that the comparison is not misread as being on identical test conditions.","section":"Table II"},{"comment":"The hyperparameters in Table III were 'selected to achieve an accuracy near 100% when tested on the training data.' This criterion risks overfitting to the training set; please report validation-set-based selection or at least discuss the sensitivity of the reported test accuracies to the threshold λ and the number of time steps K.","section":"V, Table III"},{"comment":"The description of NMNIST as 'each consisting of 300 time samples' is imprecise; NMNIST samples have varying event counts and time durations. Please rephrase to match the dataset documentation.","section":"IV.A"}],"recommendation":"major_revision","confidential_remarks":"The paper's core classification results are likely sound, but the energy-efficiency claim — the main advertised contribution — rests on a comparison that systematically omits the PointNet feature-extraction stage. This is fixable within the manuscript's scope by adding the excluded FLOPs/energy or by rewriting the claim as LCA-stage-only, but it is exactly the kind of load-bearing issue that requires major revision. I would also encourage the editor to ask for measured sparsity fractions and a clearer statement of the hardware model's validity before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does one genuinely new thing: it pairs PointNet feature extraction with the Exemplar LCA decoder and shows the combination works on NMNIST, DVS128, and SHD. That is a valid proof of concept. The paper is clearly written, the FLOP counts in Table I check out numerically, and the accuracy numbers are plausible given the dataset splits. I also credit the author for explicitly saying in Section IV.E that the FLOP estimates exclude PointNet feature extraction.\n\nWhere the paper falls down is the headline claim. The 93.41% on DVS128 with roughly 92% energy reduction sounds impressive, but it compares the LCA stage alone (0.065 mJ) against SpikePoint's full dynamic plus static energy (0.82 + 0.756 mJ). That is not apples to apples. PointNet is a deep MLP with T-Nets over 1024 points; its FLOPs are comparable to the 0.7 GFLOPs reported for LCA. Adding a realistic PointNet energy estimate would inflate the total substantially, and the advantage over SpikePoint might shrink or reverse. The paper's own numbers undercut the energy claim once the feature extractor is included.\n\nThe energy model itself is also an extrapolation. Taking Yao et al.'s 9.09e-14 J/FLOP, which was measured for MAC operations in a memristor CNN, and applying it to thresholding, neuron dynamics, and peripheral circuits is not validated. And the 0.4% active-neuron fraction is inherited from the author's prior work—fine if reproducible, but the paper says 'we have verified' without showing the verification.\n\nThese are real soft spots, but they are fixable. The method itself is not broken. The accuracy results come from external benchmarks, the equations are internally consistent, and the paper is transparent about the exclusion. I would happily read a revised version that either measures end-to-end energy or provides a full-pipeline estimate including PointNet, along with a clear statement about the RRAM energy assumptions.\n\nWho is this for? People working on sparse coding or neuromorphic edge inference for event cameras. They will find the combination interesting but will need to redo the energy math. It is not a field-reshaping paper; it is a solid workshop-level contribution with a legitimate question to explore.\n\nRecommendation: send it to peer review, not desk reject. The central idea is testable and the energy accounting can be corrected. A serious referee should push for a like-for-like energy comparison and request code or at least error bars.","headline":"Clean proof of concept, but the 92% energy claim is not like-for-like because it excludes the feature extractor.","tokens_in":10914,"tokens_out":2829,"would_cite":false,"duration_ms":26414,"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":"PointLCA-Net claims 93.41% on DVS128 with about 92% lower estimated energy than point-cloud SNNs.","keywords":["Exemplar LCA","PointNets","spatio-temporal signals","neuromorphic computing","event-based vision","energy efficiency","DVS128","sparse coding"],"falsifier":"Compute the FLOP count of the PointNet/PointNet++ feature extractor on the 1024-event DVS128 point clouds and add it to the 0.7 GFLOPs attributed to the LCA decoder; if that added cost is comparable to or larger than the decoder cost, the claimed 92% energy reduction over SpikePoint does not hold for an end-to-end deployment.","tokens_in":9845,"feed_emoji":"⚡","tokens_out":10630,"duration_ms":88668,"temperature":0.7,"pith_summary":"PointLCA-Net is a proof of concept that event-camera and cochlea-derived spatio-temporal signals, converted to (x, y, t) point clouds, can be recognized without training a deep spiking network. The paper's recipe is two-stage: a pre-trained PointNet or PointNet++ extracts a 1024-dimensional feature vector from each event cloud, and a single-layer Exemplar LCA encoder-decoder classifies those frozen features using sparse, competitive neuron dynamics. That matters because edge devices run under tight power and memory budgets, so shifting the cost into one-time feature extraction and sparse inference is an attractive trade. The strongest reported result is 93.41% top-1 accuracy on the 11-class DVS128 gesture benchmark with PointNet++ features, at an estimated 0.065 mJ for the LCA stage—the basis for the claimed roughly 92% energy reduction compared with other spiking point-cloud networks. The paper is explicit that the energy figures are theoretical and that the PointNet feature-extraction stage is excluded from the FLOP and joule accounting, so the headline saving applies to the LCA decoder rather than the whole pipeline.","feed_headline":"PointLCA-Net hits 93.41% on DVS128 with ~92% less energy","feed_subtitle":"A spiking LCA decoder over frozen PointNet features cuts estimated inference cost for event-camera recognition.","key_machinery":"The central object is the Exemplar LCA-Decoder, a single-layer spiking encoder-decoder whose dictionary columns are the PointNet feature vectors $\\varphi_i$. Each LIF neuron integrates an excitatory input $b_i = S\\varphi_i$ and is suppressed by other active neurons through the Gramian matrix $G = \\varphi^T\\varphi$; a threshold $\\lambda$ in Eq. (6) keeps the activation sparse, and the paper reports that about 0.4% of neurons are active at each time step. Classification uses either the maximum activation or the maximum sum of activations per class, so no trained readout layer is required. This machinery turns the deep PointNet extractor into a one-time stored dictionary and makes the per-test computation a crossbar-friendly vector-matrix product with sparse neuron activity.","core_discovery":"The paper's central claim is that a dictionary of PointNet/PointNet++ features can serve as the synaptic memory of an Exemplar LCA encoder-decoder, making spatio-temporal event recognition both accurate and sparse. Training consists of extracting features once from the training point clouds and storing each feature vector as a column of a memristor crossbar; no dictionary learning or decoder backpropagation is needed. At inference, LCA dynamics—excitatory drive, Gramian inhibition, and thresholding—select a tiny fraction of active neurons, and decoding by maximum sum of per-class activations yields 98.78% on NMNIST, 78.46% on SHD, and 93.41% on DVS128 with PointNet++ features. The paper further claims that LCA sparsity cuts inference computation by an average of 99.54%, and that mapping the crossbar to an RRAM array at $9.09\\times10^{-14}$ J/FLOP gives an estimated 0.065 mJ for DVS128 inference, about 24 times lower than the reported SpikePoint energy.","pith_inferences":["The paper leaves implicit that the ~92% energy saving is an LCA-stage figure: adding the excluded PointNet/PointNet++ FLOPs would determine whether the saving survives at the system level, and that calculation is an immediate next step.","The DVS128 comparison mixes protocols—PointLCA-Net uses all 11 classes with a fixed 28,606/7,408 split, while the cited 90.20% PointNet baseline used 10 classes—so a matched 10-class run would isolate how much of the gain is architectural rather than evaluational.","The paper sketches dictionary compression (PCA, SVM, discriminative feature selection) as future work; since the dictionary holds every training point, a natural stress test is to measure accuracy and FLOPs as the dictionary is pruned on a larger continuous event stream, where the full-dictionary approach would not scale."],"forward_implications":["Across three event-based datasets (NMNIST, DVS128, SHD), the same two-stage pipeline achieves high top-1 accuracy without any backpropagation in the spiking stage.","LCA sparsity reduces average inference computation by 99.54%, and cutting the integration interval from 100 to 10 time steps lowers inference FLOPs by roughly 80%.","With PointNet++ features and the maximum-sum decoder, DVS128 accuracy reaches 93.41% at an estimated 0.7 GFLOPs and 0.065 mJ per inference, giving the claimed order-of-magnitude energy advantage over the point-cloud SNN baseline.","Because the dictionary is assembled from stored features rather than learned, the same algorithm can be applied uniformly to a new spatio-temporal dataset by changing only the feature-extraction front end."],"supporting_citations":[{"why":"supplies the PointNet architecture that extracts the 1024-dimensional global feature vectors used as dictionary entries.","marker":"[1]"},{"why":"supplies PointNet++, the hierarchical feature extractor that yields the best reported DVS128 accuracy.","marker":"[2]"},{"why":"introduces the space-time event-cloud representation and the PointNet/PointNet++ DVS128 baselines that PointLCA-Net extends.","marker":"[4]"},{"why":"defines the DVS128 gesture dataset, including its 11 classes and the training/test split used with PointLCA-Net.","marker":"[7]"},{"why":"provides the Exemplar LCA-Decoder algorithm and the FLOP formulas used for the training and inference estimates.","marker":"[28]"},{"why":"provides the Locally Competitive Algorithm that produces sparse neuron activations in the encoder.","marker":"[30]"},{"why":"supplies the NMNIST spiking digit dataset used for one of the three evaluations.","marker":"[32]"},{"why":"supplies the Spiking Heidelberg Digits dataset used for the audio event-stream evaluation.","marker":"[33]"},{"why":"provides the SpikePoint accuracy, GFLOPs, and dynamic/static energy numbers used as the main point-cloud SNN comparison.","marker":"[35]"},{"why":"supplies the measured RRAM crossbar energy per FLOP used to convert the estimated FLOPs into joules.","marker":"[37]"}],"fun_headline_variants":["PointLCA-Net: sparse spiking recognition cuts energy 92%","Edge AI: PointNet + LCA on memristors hits 98.78% NMNIST","LCA spiking decoder over PointNet features: 99.5% less compute","PointLCA-Net slashes inference energy ~24x on event cameras","Sparse spiking LCA on RRAM: 93.41% DVS128 at 0.065 mJ"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the FLOP and energy estimates for the LCA decoder—the 0.4% active-neuron fraction and the $9.09\\times10^{-14}$ J/FLOP RRAM figure—represent the deployed system's real cost, even though the paper explicitly excludes the PointNet feature-extraction stage from all accounting.","fun_headline_variants_meta":{"raw":{"variants":["PointLCA-Net: sparse spiking recognition cuts energy 92%","Edge AI: PointNet + LCA on memristors hits 98.78% NMNIST","LCA spiking decoder over PointNet features: 99.5% less compute","PointLCA-Net slashes inference energy ~24x on event cameras","Sparse spiking LCA on RRAM: 93.41% DVS128 at 0.065 mJ"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000292,"raw_usage":{"total_tokens":1746,"prompt_tokens":1031,"completion_tokens":715,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":596}},"tokens_in":647,"tokens_out":715,"duration_ms":6497,"temperature":1.0,"reasoning_tokens":596,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:07:50.745629+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the FLOP count of the PointNet/PointNet++ feature extractor on the 1024-event DVS128 point clouds and add it to the 0.7 GFLOPs attributed to the LCA decoder; if that added cost is comparable to or larger than the decoder cost, the claimed 92% energy reduction over SpikePoint does not hold for an end-to-end deployment.","supporting_citations":[{"cited_title":"Sparse coding via thresholding and local competition in neural circuits,","cited_arxiv_id":null,"evidence_quote":"provides the Locally Competitive Algorithm that produces sparse neuron activations in the encoder."},{"cited_title":"Converting Static Image Datasets to Spiking Neuromorphic Datasets Using Sac- cades,","cited_arxiv_id":null,"evidence_quote":"supplies the NMNIST spiking digit dataset used for one of the three evaluations."},{"cited_title":"PointNet: Deep learning on point sets for 3D classification and segmentation,","cited_arxiv_id":null,"evidence_quote":"supplies the PointNet architecture that extracts the 1024-dimensional global feature vectors used as dictionary entries."},{"cited_title":"PointNet++: Deep hierarchical feature learning on point sets in a metric space,","cited_arxiv_id":null,"evidence_quote":"supplies PointNet++, the hierarchical feature extractor that yields the best reported DVS128 accuracy."},{"cited_title":"Space-time event clouds for gesture recognition: From RGB cameras to event cameras,","cited_arxiv_id":null,"evidence_quote":"introduces the space-time event-cloud representation and the PointNet/PointNet++ DVS128 baselines that PointLCA-Net extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the DVS128 gesture dataset, including its 11 classes and the training/test split used with PointLCA-Net."},{"cited_title":"D-SELD: Dataset- Scalable Exemplar LCA-Decoder","cited_arxiv_id":null,"evidence_quote":"provides the Exemplar LCA-Decoder algorithm and the FLOP formulas used for the training and inference estimates."},{"cited_title":"The Heidelberg Spiking Data Sets for the Systematic Evaluation of Spiking Neural Net- works,","cited_arxiv_id":null,"evidence_quote":"supplies the Spiking Heidelberg Digits dataset used for the audio event-stream evaluation."},{"cited_title":"SpikePoint: An Efficient Point-based Spiking Neural Network for Event Cameras Action Recognition,","cited_arxiv_id":null,"evidence_quote":"provides the SpikePoint accuracy, GFLOPs, and dynamic/static energy numbers used as the main point-cloud SNN comparison."},{"cited_title":"Joshua, and Qian, He","cited_arxiv_id":null,"evidence_quote":"supplies the measured RRAM crossbar energy per FLOP used to convert the estimated FLOPs into joules."}],"review_version":1}