{"id":"3591cbc2-15c3-4f2c-a64f-b5437d98275b","arxiv_id":"2505.05983","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Sparse event streams from a neuromorphic implant, filtered by a new EvFilter, can be decoded by lightweight ANN and SNN decoders with R2 up to 0.73 and 192-554x event reduction.","lead":"This paper tests whether brain implants can decode movement directly from sparse, event-based signals, bypassing conventional spike detection and sorting. It shows that a spiking neural network decoder achieves nearly the same accuracy as a larger LSTM while using 5 to 23 times less computation and memory.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"EvFilter hyperparameters (nth, tau, tref) are unreported and no validation protocol is shown; the 0.01-0.06 R2 advantage over GT spike times could be a test-set selection artifact.","rationale":"The paper's headline claim is an engineering one: a simulated NCNS event stream, after EvFilter, carries enough information to decode hand velocity, and this removes conventional spike detection/sorting. The evidence is Table I. The most load-bearing assumption is that the EvFilter settings are not chosen by looking at the test reaches, because the performance deltas are small and unsupported by error bars. The reader identified this same assumption. I agree with the reader's conditional verdict: the pipeline is plausible and the external Indy dataset is a fair source of raw signals, but the missing parameter reporting and lack of repeated-seed statistics leave the central quantitative claim insecure. A nested validation check directly tests the concern. I do not object to the architecture exploration or the resource-efficiency estimates as stated; those are secondary to the information-preservation claim, and the EvFilter selection issue is the controlling uncertainty.","tokens_in":8100,"tokens_out":10581,"duration_ms":114745,"concrete_test":"Run a nested validation protocol on each of the five recordings: for the ST-NN decoder, grid-search nth in {2,4,8,16}, tau in {1,5,10,20} ms, and tref in {0,1} ms using only training and validation reaches, freeze the settings that maximize validation R2, then evaluate on test reaches. Compare the mean and per-day test R2 of EvFilter/EvFilter-SPD against the GT-spike baseline computed with identical splits and seeds. If the EvFilter advantage disappears or falls within run-to-run noise, the claim that event-based features eliminate spike detection/sorting is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's empirical basis for 'eliminating the need for signal recovery, spike detection, or sorting' is the R2 advantage of EvFilter/EvFilter-SPD events over GT spike times in Table I. These differences are small (e.g., ST-NN: 0.7070 vs 0.6823; SNN: 0.6980 vs 0.6703; LSTM: 0.7272 vs 0.7233) and are reported without error bars or repeated seeds. Section II-A defines EvFilter with parameters nth, tau, and tref, but the paper never reports their values or the procedure by which they were chosen. If these parameters were selected by inspecting test-set R2 on the five recordings, the advantage over GT spike times is a selection artifact rather than evidence that NCNS events preserve decodable information. Since the main conclusion depends on this comparison, the central claim is not secured until parameter selection is shown to be validation-based and variance estimates are provided.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a decoding pipeline for neuromorphic implantable brain-machine interfaces (Neu-iBMI). It introduces EvFilter, a tunable event filter, and EvFilter-SPD, a spike-detector variant, which process events generated by the authors' earlier NCNS sensing scheme. The filtered events are fed to shallow ANN (NN and ST-NN), LSTM, and SNN decoders. On five recordings from a public non-human-primate reaching dataset, the pipeline reaches R2=0.73 with an LSTM decoder and R2=0.70 with an SNN decoder, while reducing event throughput by 192x and 554x compared with raw NCNS events. The authors claim that this eliminates the need for conventional signal recovery, spike detection, and sorting, and that the SNN decoder requires 5-23x fewer compute and memory resources than the compared ANN/LSTM decoders.","tokens_in":8340,"tokens_out":6444,"duration_ms":55117,"significance":"If the claims hold, the pipeline would be a meaningful step toward low-power, wireless neuromorphic BMI decoding. Strengths include the use of an external public dataset, an actual reaching/kinematic decoding task, hardware-oriented resource estimates via Neurobench, and a parameterizable filtering scheme that can trade off event reduction against decoding accuracy. The reported event-compression ratios (192x-554x) are impressive if reproducible. However, the central claims currently rest on small R2 differences over the GT-spike baseline, and the manuscript does not report the EvFilter hyperparameters, the hyperparameter selection procedure, or any variance statistics. Those omissions materially weaken the support for the main conclusions.","major_comments":[{"comment":"The EvFilter parameters nth, tau, and tref are defined in Section II-A but never reported, and the procedure by which they were chosen is not described. The central claim that EvFilter/EvFilter-SPD events preserve or improve decodable information over GT spike times rests on R2 differences of only 0.01-0.06 (e.g., ST-NN: 0.7070 EvFilter vs. 0.6823 GT; LSTM: 0.7331 EvFilter-SPD vs. 0.7233 GT). If these parameters were selected on the test recordings, the advantage could be a selection artifact. Please report the hyperparameter values and demonstrate that they were selected using only the validation splits (for example, by grid search on the validation reaches), and provide per-seed or per-session error bars.","section":"§II-A, Table I"},{"comment":"All R2 scores in Table I are single numbers without error bars, repeated seeds, or statistical tests. The paper ranks decoders and input representations based on differences as small as 0.01 (e.g., ST-NN EvFilter 0.7070 vs. GT 0.6823; LSTM EvFilter-SPD 0.7331 vs. EvFilter 0.7272; SNN EvFilter-SPD 0.7033 vs. EvFilter 0.6980). Single-run results cannot support these rankings. Please add repeated-seed or session-wise statistics, and where possible a significance test, to support the comparative claims.","section":"§III-C, Table I"},{"comment":"The comparisons with external baselines (rEFH-dynamic [29], 2D-SNN [30], UKF [29]) are not controlled. These systems were trained on different data selections and with different protocols; the text itself concedes that rEFH-dynamic was trained on a much larger selection of datasets. The statement that event-based features 'outperform other popular decoding algorithms' is therefore not established by this table. Please either reimplement those baselines under identical train/validation/test splits and training pipelines, or restrict the claim to the internal comparisons among the decoders studied in this paper.","section":"Table I, §III-C2"},{"comment":"The paper uses 'five recordings including the first and last day' from the NHP 'Indy' dataset, but it does not justify why these five recordings were selected. If the recordings were chosen after inspecting decoding performance, the aggregate R2 could be biased upward. Please state the inclusion criterion explicitly, or report results across all available sessions with per-session variability, so the reader can assess the generality of the findings.","section":"§III-A"},{"comment":"The abstract and introduction claim the pipeline eliminates 'the need for signal recovery, spike detection, or sorting,' yet Section II-A states that EvFilter-SPD is a spike detector realized as a special case of EvFilter. This is internally consistent only if 'eliminating' means there is no separate, conventional threshold-crossing and sorting stage with the detection integrated into the filter. As written, the claim is overstated and confusing. Please rewrite the claim to state precisely which conventional stages are removed and what role EvFilter-SPD plays.","section":"Abstract, §I, §II-A"}],"minor_comments":[{"comment":"The event-reduction factor is reported inconsistently: the contributions say '192−544×,' the abstract says '554×,' and Section III-D says '192× and 554× respectively.' Please correct this inconsistency (presumably 192x and 554x).","section":"Abstract, §I, §III-D"},{"comment":"The text says 'hidden layer 1 and 2 with 32 and 48 layers,' but the intended meaning is clearly 32 and 48 neurons. Please fix this wording.","section":"§II-B1"},{"comment":"There is an inconsistency in the dropout rate: Section II-B1 states 'dropout (50% probability),' while Section III-C1 says 'a dropout rate between 0.3−0.5.' Please clarify which dropout rates were used and for which models.","section":"§II-B1, §III-C1"},{"comment":"Table II is difficult to read because the columns are not visually separated (e.g., '0.6571 3643.952 0 4.866 20.856' runs together), and some rows appear to lack the Memory column. Please reformat the table with clear column alignment and verify that all entries are present.","section":"Table II"},{"comment":"The notation 'Sout[t−1]θ' in the membrane-potential update is ambiguous; please insert an explicit multiplication sign and define Sout consistently with the subsequent firing rule.","section":"Eq. (5)"},{"comment":"Reference [28] is cited as an arXiv preprint; if a peer-reviewed version has been published, please update the citation.","section":"References"},{"comment":"The abstract highlights 'R2=0.73' without noting that this value is achieved by the LSTM decoder on EvFilter-SPD input, not by the resource-efficient SNN decoder (R2≈0.70). Please clarify this in the abstract to avoid overstating the performance of the low-resource decoders.","section":"Abstract, Table I"},{"comment":"The SNN row is labeled 'SNN Stream' and has no T_bin entry; please enter '—' or 'streaming (no binning)' to make the table self-explanatory.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The manuscript leans heavily on the authors' own NCNS sensing scheme ([8], [9]) and earlier decoder exploration ([28]); this is acceptable in principle because the evaluation uses an external public dataset, but the dependence on unreported filter parameters and the lack of variance estimates make the central empirical claim fragile. The main fixes needed are reporting the EvFilter parameter values and selection protocol, adding error bars or session-wise statistics, and tightening the baseline comparisons. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the demonstration that filtered NCNS event streams—not recovered signals, not sorted spikes—can drive hand-velocity decoding in an NHP reaching task, with R2 up to 0.73 and 192–554x event reduction. That is a concrete step for wireless implantable BMIs, where power and bandwidth are real constraints. The resource analysis via Neurobench is also a plus: the SNN decoder's 5–23x savings in computes and memory are the kind of number that matters for on-implant feasibility. Credit where due: the evaluation uses an external public dataset, the decoder comparison is reasonably broad, and the claims in the abstract match the tables.\n\nWhere I'd push back: the stress-test note is on target. Section II-A defines EvFilter with parameters nth, tau, and tref, but the paper never reports their values or the selection procedure. The R2 advantage of EvFilter-SPD over GT spike times is small—roughly 0.01 to 0.04 depending on decoder—and without repeated seeds or error bars, that advantage could be a test-set selection artifact. The improvements are consistent across decoders, which makes me think the effect is real, but consistency across five sessions with unknown parameter tuning is not the same as statistical evidence. Also, the baseline comparisons to rEFH-dynamic and 2D-SNN come from other papers with different training splits; they are context, not controlled baselines. The five recording sessions are unexplained, and with only five, a single noisy session could drive the numbers.\n\nThe self-citation to the NCNS sensing scheme ([8], [9]) is a mild circularity concern, but the external dataset and the fact that the pipeline is evaluated end-to-end mitigate it. I don't see a load-bearing flaw in the central argument; the weakness is in the reporting and validation protocol. The paper would be much stronger with parameter values, a validation-based selection procedure, error bars, and ideally released code or data.\n\nThis deserves peer review. It is a solid engineering result with a clear purpose, and the gaps are fixable. A serious referee should ask for the missing hyperparameters and variance estimates before the claims about eliminating spike detection are accepted. For a reading group, it is worth a look if you work on neuromorphic sensing or low-power BMI; otherwise it is a skip.\n\nBottom line: send it to review, require the validation details, and it could become a citeable benchmark for NCNS-based decoding.","headline":"A useful engineering exploration of decoding NCNS event streams for motor BMI, but missing EvFilter hyperparameters and error bars leave the small advantage over ground-truth spikes unsecured.","tokens_in":8872,"tokens_out":1674,"would_cite":true,"duration_ms":19625,"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":"A tunable event filter lets a neuromorphic brain-machine interface decode hand velocity from a 554x sparser event stream, reaching R2=0.733 without separate spike sorting.","keywords":["neuromorphic implantable brain-machine interface","event-based neural sensing","neural decoding","spiking neural networks","event filter","hand kinematics decoding","neuromorphic compression","low-power implant"],"falsifier":"Run the same decoders on held-out recording sessions while choosing $n_{\\text{th}}$, $\\tau$, and $t_{\\text{ref}}$ using only training reaches; if the resulting test $R^2$ with EvFilter-SPD no longer exceeds the test $R^2$ from binned ground-truth spike times, the paper's claim that the pipeline can replace spike detection fails.","tokens_in":7937,"feed_emoji":"🧠","tokens_out":10327,"duration_ms":95008,"temperature":0.7,"pith_summary":"This paper is trying to establish that a neuromorphic brain-machine interface can decode hand motion directly from its sparse event stream, without reconstructing the neural signal and without the conventional separate steps of spike detection and sorting. The authors add a tunable event filter to their compression-based neural sensing scheme; the filter suppresses background events by 192x, and when used as a spike detector by 554x, before feeding an ANN, a temporally segmented ANN, or a spiking neural network decoder. On a non-human primate reaching task, the best decoder reaches $R^2 = 0.733$ and the much cheaper SNN decoder reaches $R^2 \\approx 0.70$ with 5-23x fewer computations and memory than NN and LSTM decoders. If the claim holds, wireless implants with many channels could avoid the power- and bandwidth-heavy preprocessing of conventional iBMIs and still control prosthetics or cursors.","feed_headline":"Filtered events decode hand motion at R2=0.733, no spike sorting","feed_subtitle":"A simple event filter lets ANN and spiking decoders skip spike sorting, cutting compute and memory up to 23x.","key_machinery":"The load-bearing object is the EvFilter, a channel-wise temporal-neighborhood filter adapted from dynamic-vision-sensor background-activity filtering. It passes an incoming event on a channel only if at least $n_{\\text{th}}$ events have occurred within a temporal window $\\tau$ on that channel, with a refractory period $t_{\\text{ref}}$ that can be set to zero or to about 1 ms; with a large $t_{\\text{ref}}$ the filter becomes the spike detector EvFilter-SPD. This single mechanism does two jobs: it suppresses background events (192x reduction) and, in the SPD setting, reduces the stream to roughly one event per spike (554x reduction), while in both settings preserving or improving the kinematic information relative to ground-truth spikes. On the decoder side, the SNN decoder's binary input $\\bar{x}_{\\text{SNN}}(t) \\in \\{0,1\\}^{N_{\\text{ch}}}$ replaces multiply-accumulate with accumulation, which is what turns the event sparsity into 5-23x resource savings.","core_discovery":"The central claim is that the event stream generated by the neuromorphic compression sensing scheme, after passing through the EvFilter, carries enough information for hand-velocity decoding on its own, and that this information is actually easier to decode than the dataset's ground-truth spike times. In the authors' experiments, every decoder that was tested with both input types performed better when fed EvFilter or EvFilter-SPD output than when fed ground-truth spike times; the best mean score is $R^2 = 0.733$ for an LSTM with EvFilter-SPD. The SNN decoder reaches $R^2 = 0.703$ using binary per-channel inputs and accumulate-only operations, giving 5-23x compute and memory savings over NN and LSTM decoders. The paper concludes that the full conventional pipeline of signal recovery, spike detection, and spike sorting can be omitted, and that the filtered event stream can be decoded on-implant or on a low-power wearable.","pith_inferences":["If the EvFilter's advantage is general, the same simple 'pass only events with several neighbors in the recent past' rule should transfer to other event-based neural sensors, such as peripheral-nerve or surface electrodes; a direct test is to reuse the default filter settings on a second recording session or species and compare against binned spikes.","The fact that EvFilter-SPD output beats ground-truth spike times suggests some of the retained events are not true spikes by the reference sorter's criteria but still carry kinematic information; decoding from only those 'extra' events would isolate and test that interpretation.","The resource comparisons count 32-bit operations and memory; actual implant benefit depends on filter hardware cost and wireless transmit energy, so an end-to-end chip measurement that includes the filter is the natural next step and could either confirm or shrink the reported 5-23x savings.","For closed-loop control, latency and smoothness matter as much as mean $R^2$; a natural extension is to test the SNN decoder in an online brain-controlled cursor task with continuous streams and no reach-segmented resets."],"forward_implications":["A neuromorphic implant could transmit only filtered events, cutting the data volume by 192-554x while preserving motor-decoding accuracy, relaxing the main bandwidth constraint on wireless high-channel-count implants.","A shallow SNN decoder with binary inputs and accumulate-only operations is sufficient for hand-velocity decoding on this task, needing 5-23x fewer computations and memory than NN or LSTM decoders, which makes on-implant decoding plausible.","The ST-NN decoder with segmented time-bins recovers most of the LSTM's performance at 2.5x lower resources, showing that a small amount of temporal structure in the input features can substitute for a recurrent network.","When the sensor is event-based, conventional signal recovery, spike detection, and spike sorting can be removed from the iBMI pipeline; the filter/detector is co-designed with the decoder rather than a separate neural-signal preprocessing stage.","Because event-filtered inputs also improve linear and Kalman decoders, the benefit is not specific to deep networks; the filter itself is the main source of the quality gain."],"supporting_citations":[{"why":"Supplies the NCNS event-generation architecture that turns raw neural recordings into the sparse event stream used throughout the paper.","marker":"[8]"},{"why":"Extends the NCNS scheme and provides the simulation pipeline used to convert raw recordings into neural events.","marker":"[9]"},{"why":"Introduces hybrid event-frame spike detection for neuromorphic iBMIs, the prior approach the EvFilter builds on.","marker":"[24]"},{"why":"Provides the dynamic-vision-sensor background-activity filter principle that the EvFilter adapts to neural events.","marker":"[25]"},{"why":"Supplies the ANN-versus-SNN decoder comparison and hidden-layer sizes that inform the NN and ST-NN architectures.","marker":"[28]"},{"why":"Provides the non-human primate reaching dataset with raw recordings and ground-truth spike times used for training and baseline comparison.","marker":"[29]"},{"why":"Provides the benchmarking harness used to estimate computes, memory, and model size for the decoder resource comparisons.","marker":"[30]"}],"fun_headline_variants":["Filtered events beat spike times, decode hand motion at R²=0.733","554X fewer events: EvFilter enables on-implant BMI decoding","Skip spike sorting: filtered event stream hits R²=0.733","SNN decoder cuts compute 23X, no spike sorting required"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the way neural recordings are converted into sparse events, and the filter settings used in that conversion, preserve the information needed to decode movement without being chosen by looking at the test data; otherwise the reported advantage over ground-truth spike times is an artifact.","fun_headline_variants_meta":{"raw":{"variants":["Filtered events beat spike times, decode hand motion at R²=0.733","554X fewer events: EvFilter enables on-implant BMI decoding","Skip spike sorting: filtered event stream hits R²=0.733","SNN decoder cuts compute 23X, no spike sorting required"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000364,"raw_usage":{"total_tokens":1951,"prompt_tokens":927,"completion_tokens":1024,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":944}},"tokens_in":543,"tokens_out":1024,"duration_ms":9869,"temperature":1.0,"reasoning_tokens":944,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:50:51.297714+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same decoders on held-out recording sessions while choosing $n_{\\text{th}}$, $\\tau$, and $t_{\\text{ref}}$ using only training reaches; if the resulting test $R^2$ with EvFilter-SPD no longer exceeds the test $R^2$ from binned ground-truth spike times, the paper's claim that the pipeline can replace spike detection fails.","supporting_citations":[{"cited_title":"Architectural exploration of neuromorphic compression based neural sensing for next-gen wireless implantable-BMI,","cited_arxiv_id":null,"evidence_quote":"Supplies the NCNS event-generation architecture that turns raw neural recordings into the sparse event stream used throughout the paper."},{"cited_title":"Towards neuromorphic compression based neural sensing for next-generation wireless iBMI,","cited_arxiv_id":null,"evidence_quote":"Extends the NCNS scheme and provides the simulation pipeline used to convert raw recordings into neural events."},{"cited_title":"Hybrid event-frame neural spike de- tector for neuromorphic implantable BMI,","cited_arxiv_id":null,"evidence_quote":"Introduces hybrid event-frame spike detection for neuromorphic iBMIs, the prior approach the EvFilter builds on."},{"cited_title":"A 128× 128 120 db 15 µs latency asynchronous temporal contrast vision sensor,","cited_arxiv_id":null,"evidence_quote":"Provides the dynamic-vision-sensor background-activity filter principle that the EvFilter adapts to neural events."},{"cited_title":"Combining SNNs with Filtering for Efficient Neural Decoding in Implantable Brain-Machine Interfaces","cited_arxiv_id":"2312.15889","evidence_quote":"Supplies the ANN-versus-SNN decoder comparison and hidden-layer sizes that inform the NN and ST-NN architectures."},{"cited_title":"Nonhuman primate reaching with multichannel sensorimotor cortex electrophysiology,","cited_arxiv_id":null,"evidence_quote":"Provides the non-human primate reaching dataset with raw recordings and ground-truth spike times used for training and baseline comparison."},{"cited_title":"Neurobench: Advancing neuromorphic computing through collaborative, fair and representative benchmarking,","cited_arxiv_id":null,"evidence_quote":"Provides the benchmarking harness used to estimate computes, memory, and model size for the decoder resource comparisons."}],"review_version":1}