{"id":"1932a1e8-da11-4b8b-bd8c-525b5e6d6788","arxiv_id":"1908.06180","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A multi-view version of the Broad Learning System fuses LFP and spike features and achieves the highest reported decoding accuracy on a monkey oculomotor decision dataset, with small but statistically significant gains over baselines.","lead":"The paper extends the Broad Learning System, a wide neural network, to combine two types of brain signals, local field potentials and spikes, recorded from monkeys making oculomotor choices. The combined model reaches about 48% accuracy in a four-way decision decoding task, several points above single-signal baselines, but the gains are small and the dataset and code are not released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Significance tests treat repeated random partitions as independent samples, likely inflating the evidence for MvBLS's superiority.","rationale":"The reader's weakest_assumption focuses on the decoding window including the saccade, which is an interpretive concern. However, the reader's rationale also lists the significance-test issue as one of three concerns. I identify the significance-test independence assumption as the most load-bearing because the central claim of MvBLS being 'more effective' is explicitly supported by Table IV's significant p-values; if those p-values are invalid, the headline comparison is unsupported regardless of the saccade-window interpretation. The saccade-window issue is real but less decisive for the algorithmic claim, since decoding the executed saccade is still behaviorally relevant. The verdict remains CONDITIONAL as the reader stated, so I mark it UNCHANGED: the paper should correct the statistical analysis or release session-level data/code before the significance claims can be accepted. My agreement with the reader is partial because we share the same overall verdict and several concerns, but our chosen weakest assumption differs; I find the statistical test validity more fundamental than the decoding-window interpretation.","tokens_in":19720,"tokens_out":4399,"duration_ms":44277,"concrete_test":"Recompute Dunn's multiple comparison using one aggregated accuracy per session per algorithm (the mean over the 30 repeated splits, yielding 45 paired samples per algorithm), or apply a paired bootstrap across sessions. If the MvBLS vs. Ridge p-value rises above 0.05, or if any of the significant comparisons in Table IV lose significance, the 'significant outperformance' claim fails. The paper should also explicitly state whether the original test used all 30×45 accuracies or session-level aggregates, since the test's validity hinges on that choice.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MvBLS 'significantly outperforms' all baselines (Table IV) rests on Dunn's non-parametric multiple comparison test applied to 'cross-validation accuracies' from Section III-D, where each of the 45 datasets was randomly partitioned 30 times. If the test uses all 30×45 = 1350 accuracies per algorithm, those observations are not independent: the same neural recordings contribute 30 correlated accuracies, since the random splits reuse the same trial data. Rank-based tests assume independent samples, so the effective sample size is far smaller than 1350, and the reported p-values (e.g., MvBLS vs. Ridge = 0.0019) are likely artificially small. The paper never states whether the test used all repeats or session-level aggregates; Section III-E says the Table II standard deviations were computed from '30 average accuracies of the 45 sessions,' implying a 30-sample summary, which would make significant differences much harder to detect. Without a valid significance test, the claim of superiority over Ridge (46.51% vs. 47.94%, a 1.43-point gap) and over BLS (45.65%) rests only on point estimates whose boxplots overlap considerably. The 'more effective' conclusion and the complementary-information claim are directly underwritten by these p-values, making this the most load-bearing weakness in the paper's argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the Broad Learning System (BLS) to a multi-view setting (MvBLS) by constructing feature nodes for each view separately, concatenating them to form enhancement nodes, and then jointly regressing to the label matrix. The method is applied to decode four oculomotor choice directions from LFPs and spikes recorded in the medial frontal cortex of two macaque monkeys over 45 sessions. The authors report that MvBLS achieves higher average accuracy than SVM, ridge regression, BLS, MvDA, and MvMDA, and that combining LFPs and spikes improves over either modality alone, which they interpret as evidence for complementary decision information. They also report parameter sensitivity, runtime comparisons, and additional experiments with single-unit versus multi-unit spikes.","tokens_in":19943,"tokens_out":7071,"duration_ms":67572,"significance":"If the empirical claims are reliable, the paper contributes a simple multi-view extension of BLS that is computationally efficient and performs well on a challenging neural decoding task. The evaluation is extensive, covering 45 sessions, 30 random partitions per session, multiple single-view and multi-view baselines, binary and four-class decoding, single-unit versus multi-unit comparisons, and runtime benchmarks. Notable strengths include the explicit use of non-parametric multiple-comparison tests, the session-level counts of improvement, and the parameter sensitivity analysis. However, the statistical analysis and the temporal definition of the decoding window need to be corrected before the central claims can be accepted.","major_comments":[{"comment":"The significance tests appear to treat the 30 random partitions per session as independent samples. Dunn's procedure is applied to 'cross-validation accuracies,' but all partitions of a session reuse the same trials, so the 30 accuracies per session are strongly correlated. The manuscript never states whether the test input consisted of 30×45 = 1350 per-algorithm accuracies, 45 session-mean accuracies, or 30 session-averaged accuracies; the description in Section III-E that Table II standard deviations were 'computed from 30 average accuracies of the 45 sessions' suggests the latter, which would drastically change the p-values in Table IV. As reported, p-values such as MvBLS versus Ridge = 0.0019 cannot be taken at face value. The authors must re-run the comparisons using a valid blocked or repeated-measures test (e.g., Friedman test on session-level accuracies with post-hoc paired comparisons), state the exact test unit, and justify the independence assumption.","section":"Section III-E and Table IV"},{"comment":"The decoding window [0,400) ms after target onset includes the saccade, because reaction times are 100–300 ms. The neural features therefore contain movement-execution and post-saccadic feedback activity. The claim of decoding 'oculomotor decision' or 'choice intention' is not uniquely supported, since the classifiers may be decoding the direction of the executed saccade. This confound also weakens the complementary-information interpretation in Section III-G, as the improvement from combining LFPs and spikes could reflect complementary motor-related signals. The authors should provide control analyses, for example time-resolved decoding restricted to pre-saccadic windows or a window ending before the minimum reaction time, to show that the accuracy advantage of MvBLS and the multimodal gain are present before movement onset.","section":"Section III-C and III-G"},{"comment":"The claim that 'combining LFP and spike features always improved the decoding performance' is made without any significance test, and the table itself shows the same LFP condition with three slightly different mean accuracies (40.57±0.70, 40.61±0.60, 40.41±0.61), which the text attributes to different train/validation/test partitions. This means the conditions are not matched on identical partitions, so the pairwise differences between conditions could be partly due to partition variability. The authors should use the same random partitions across all conditions and perform paired statistical tests before asserting that fusion 'always' improves decoding.","section":"Section IV-B and Table VI"}],"minor_comments":[{"comment":"The sentence 'rejected if p ≥ α/2' is presumably a typo; with an FDR-corrected Dunn's test one rejects when the adjusted p-value is below α. Please correct.","section":"Section III-E"},{"comment":"The word 'preforming' should be 'performing'.","section":"Section III-B"},{"comment":"Table VI duplicates the first three rows of Table II for the LFPs, Spikes, and LFPs+Spikes conditions; consider removing the duplication and referencing Table II to avoid confusion.","section":"Table VI"},{"comment":"The description of how the standard deviations in Table II were computed is ambiguous; please clarify whether the 30 averages are across sessions per partition or across partitions per session.","section":"Section III-E"},{"comment":"The statement that most prior studies using combined LFPs and spikes 'have not shown significant improvements' is a strong generalization that would benefit from a systematic comparison or a quantitative summary.","section":"Section IV-G"},{"comment":"The DCCAE comparison used modified hyperparameters and only eight runs per session; this weaker comparison should be described as preliminary.","section":"Section IV-E"},{"comment":"The criterion for selecting the 10 displayed datasets is not stated; please specify how they were chosen.","section":"Figure 9(b)"},{"comment":"The paper does not report a per-monkey breakdown; because the two monkeys contributed unequal numbers of sessions (33 vs. 12), the authors should confirm that the main results hold within each monkey.","section":"General"},{"comment":"The manuscript does not provide code or data; for a paper whose main contribution is an empirical decoding comparison, a reproducibility statement or a link to an implementation would be valuable.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The statistical issue in the main comparison is serious but fixable with a reanalysis using session-level paired tests and a clear statement of the test unit. The temporal-window confound requires additional analyses, but the existing data should allow a pre-saccadic window check. I recommend requesting a major revision focused on these two points, and I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent, incremental machine learning paper. The MvBLS extension is simple—each view gets its own feature nodes, then they're concatenated and fed into enhancement nodes—and the empirical claim that it beats a strong Ridge baseline on a real monkey dataset is plausible. But the statistical evidence for that claim is muddled enough to be the load-bearing soft spot.\n\nWhat the paper does well: it's the first application of multi-view learning to invasive non-human primate LFP/spike decoding, which is a fair first-use contribution. The comparison set is thorough (SVM, Ridge, BLS, plus two other multi-view methods), the 45-session repeated-partition protocol is sensible, and the parameter sensitivity and runtime analyses are genuinely useful. The complementary-information argument gets support from the across-session win counts, not just point estimates, which is the right kind of evidence.\n\nThe real problem is the significance testing. The paper says Dunn's test was run on \"cross-validation accuracies,\" but never states whether those are the 30 per-session values (1350 total) or 30 session-averaged values. If it's the former, those observations are not independent—the same trials are reused across the 30 random splits—so the p-values in Table IV are likely inflated. If it's the latter, the test is more defensible but the paper needs to say so. This matters because the headline claim that MvBLS \"significantly\" outperforms Ridge (47.94 vs 46.51) rests entirely on those p-values. The point estimates alone, with overlapping boxplots and a 1.4-point gap, are not convincing without a valid test. This is a fixable issue: report the test on session-level paired differences or use a proper repeated-measures test.\n\nSecondary concerns: the [0,400) ms window includes the executed saccade (reaction times 100–300 ms), so the \"decision\" decoding may partly reflect the movement itself. This weakens the interpretation but not the algorithmic comparison. And there's no code or data release, so the exact numbers can't be verified. None of this destroys the paper, but together they mean the central superiority claim is not as solid as the abstract suggests.\n\nVerdict: send to peer review—the architecture, experiments, and application are worth serious referee time. The revision should clarify the significance test, ideally share code, and at least discuss the saccade-window issue. I'd bring it to a reading group focused on neural decoding methods, though I wouldn't build anything on it until the stats are resolved.","headline":"Plausible incremental ML extension—MvBLS combines per-view feature nodes—but the significance tests are under-specified and the decoding window includes the saccade.","tokens_in":20517,"tokens_out":3582,"would_cite":false,"duration_ms":35687,"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":"Treating local field potentials and spikes as two views of the same medial frontal signal and fusing them through a multi-view broad learning system decodes monkey oculomotor decisions more accurately than single-modality and…","keywords":["broad learning system","multi-view learning","local field potentials","spikes","oculomotor decision decoding","medial frontal cortex","brain-machine interface","multi-class classification"],"falsifier":"Run the same 4-class comparison on epochs that end before the earliest saccade, e.g. 0-90 ms after target onset, or on saccade-aligned windows ending at saccade onset. If MvBLS no longer beats the best single-view baseline there, the claim that it decodes the oculomotor decision and that LFPs and spikes are complementary for that decision would fail.","tokens_in":19500,"feed_emoji":"🧠","tokens_out":7559,"duration_ms":69346,"temperature":0.7,"pith_summary":"This paper extends the broad learning system, a wide single-layer neural network, to handle two views of the same neural signal at once, and tests it on recordings from the medial frontal cortex of two macaques performing an oculomotor gambling task. The two views are local field potentials (LFPs) and action potentials (spikes), recorded simultaneously from the same electrodes. The paper's central assertion is that these two signal types carry complementary information about the monkey's upcoming choice, and that a multi-view BLS can exploit that complementarity better than single-view classifiers or multi-view subspace methods. The reported 45-session mean four-class accuracy is 47.94%, above the best baseline of 46.51%, with the pairwise differences reported as statistically significant.","feed_headline":"Two neural signals beat one in decoding monkey choices","feed_subtitle":"A multi-view Broad Learning System reaches 47.94% on four-choice monkey decisions, besting every baseline.","key_machinery":"The central object is the multi-view broad learning system (MvBLS), a wide neural network that replaces deep stacking with broad feature construction. Each view is passed through its own set of sparse, de-noised linear feature nodes built by random projection followed by L1-regularized reconstruction; the two views' feature nodes are concatenated, mapped through randomly weighted orthonormal bases and a sigmoid to form enhancement nodes, and the whole is connected to a one-hot label matrix by ridge regression. The paper's argument is that keeping the two views separate in the first layer, rather than concatenating raw features before feature extraction, preserves each signal's specific structure and lets the output layer fuse them effectively, and that this yields better accuracy than subspace-alignment multi-view methods such as MvDA and MvMDA.","core_discovery":"On the paper's own terms, the discovery is a fusion architecture plus an empirical result: building separate sparse feature-node banks for the LFP view and the spike view, then feeding their concatenation into shared nonlinear enhancement nodes and a ridge-regression output layer, decodes the direction of a monkey's saccadic choice better than any of the six compared approaches. The gain appears in both four-way decoding and binary choice decoding, and it holds across most of the 45 sessions. The paper also reports that combined LFP+spike features outperform either modality alone under every classifier tested, which it reads as evidence that the two signals encode complementary aspects of the oculomotor decision.","pith_inferences":["A testable next step the authors do not run is to restrict the decoding epoch to the pre-saccade interval; if the accuracy gap persists there, the 'decision' reading is strengthened, and if it collapses, the decoder may be reading the executed saccade instead.","Because the LFP and spike feature banks are built independently, the same MvBLS could be applied to other paired neural or physiological streams, such as ECoG with spikes or EEG with eye tracking, without changing the architecture.","The reported robustness to parameter choices suggests that cross-validation cost could be reduced in future applications by fixing the structural parameters and tuning only the ridge regularization, though the paper does not claim this."],"forward_implications":["Fusing LFP and spike features improves oculomotor decision decoding over either modality alone in most of the 45 sessions, for every classifier tested.","MvBLS outperforms six classical and state-of-the-art single- and multi-view baselines with statistically significant pairwise differences in four-class decoding.","MvBLS is faster than deep multi-view alternatives, making online or high-throughput neural decoding feasible.","The architecture is insensitive to its structural parameters, so practitioners can choose small networks for speed without losing accuracy.","The paper expects the approach to transfer to other primate brain-state decoding tasks and beyond, since the two-view construction is not specific to oculomotor signals."],"supporting_citations":[{"why":"Supplies the single-view broad learning system that MvBLS extends to multiple views.","marker":"[5]"},{"why":"Provides the oculomotor gambling task, the 45-session medial frontal LFP and spike recordings, and the random-guess and permutation-test baselines.","marker":"[9]"},{"why":"Supplies the eight LFP frequency bands used to build the spectral features in the LFP view.","marker":"[14]"},{"why":"Provides the multi-view discriminant analysis baseline that MvBLS must beat.","marker":"[17]"},{"why":"Provides the multi-view modular discriminant analysis baseline that MvBLS must beat.","marker":"[18]"},{"why":"Provides the deep multi-view representation-learning baseline (DCCAE) used in the speed-accuracy comparison.","marker":"[22]"}],"fun_headline_variants":["Multi-view BLS fuses LFP and spikes to boost monkey decision decoding","Fusing spike and LFP views improves oculomotor decision decoding","Spike+LFP fusion beats single signals for monkey decision decoding","Multi-view BLS: two neural views beat one in decoding choices","Multi-view BLS outperforms in monkey oculomotor decision decoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim assumes that the 400-millisecond decoding window after target onset measures the monkey's choice intention rather than the executed eye movement itself, because the monkeys' reaction times are 100-300 milliseconds and the window therefore includes the saccade.","fun_headline_variants_meta":{"raw":{"variants":["Multi-view BLS fuses LFP and spikes to boost monkey decision decoding","Fusing spike and LFP views improves oculomotor decision decoding","Spike+LFP fusion beats single signals for monkey decision decoding","Multi-view BLS: two neural views beat one in decoding choices","Multi-view BLS outperforms in monkey oculomotor decision decoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000897,"raw_usage":{"total_tokens":3809,"prompt_tokens":834,"completion_tokens":2975,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":2883}},"tokens_in":450,"tokens_out":2975,"duration_ms":22816,"temperature":1.0,"reasoning_tokens":2883,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:53:33.009788+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 4-class comparison on epochs that end before the earliest saccade, e.g. 0-90 ms after target onset, or on saccade-aligned windows ending at saccade onset. If MvBLS no longer beats the best single-view baseline there, the claim that it decodes the oculomotor decision and that LFPs and spikes are complementary for that decision would fail.","supporting_citations":[{"cited_title":"Broad learning system: An effectiv e and efficient incremental learning system without the need for deep archi tecture,","cited_arxiv_id":null,"evidence_quote":"Supplies the single-view broad learning system that MvBLS extends to multiple views."},{"cited_title":"Sequential selection of econom ic good and action in medial frontal cortex of macaques during value -based decisions,","cited_arxiv_id":null,"evidence_quote":"Provides the oculomotor gambling task, the 45-session medial frontal LFP and spike recordings, and the random-guess and permutation-test baselines."},{"cited_title":"Multiscale modeling and decoding algorithms for spike-field activity,","cited_arxiv_id":null,"evidence_quote":"Supplies the eight LFP frequency bands used to build the spectral features in the LFP view."},{"cited_title":"Multi-vie w discriminant analysis,","cited_arxiv_id":null,"evidence_quote":"Provides the multi-view discriminant analysis baseline that MvBLS must beat."},{"cited_title":"Generaliz ed multi- view embedding for visual recognition and cross-modal retr ieval,","cited_arxiv_id":null,"evidence_quote":"Provides the multi-view modular discriminant analysis baseline that MvBLS must beat."},{"cited_title":"On deep m ulti-view representation learning,","cited_arxiv_id":null,"evidence_quote":"Provides the deep multi-view representation-learning baseline (DCCAE) used in the speed-accuracy comparison."}],"review_version":1}