{"id":"da47465f-c81f-49db-b627-46e1b4a3d9b9","arxiv_id":"1908.02830","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"The authors introduce Variable Input Length Map (VILMAP), an SOM variant that handles inputs of different lengths and can recognize words in phoneme sequences without forgetting earlier dimensions.","lead":"VILMAP is a new self-organizing map that learns patterns of different lengths in one model, sliding and stretching its internal prototypes to match variable-size inputs. The paper tests it on finding repeated motifs in time series and on recognizing words in phoneme data, aiming at biologically plausible speech segmentation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Experiment 3 does not measure word segmentation, so the claimed competitive segmentation result is unsupported; no variable-length motif discovery experiment is reported either.","rationale":"The reader's verdict (REJECT, moderate confidence) is justified. My main concern is not the technical inheritance assumption in Section III-B, which the reader flagged; it is the more direct empirical mismatch between the tasks the paper claims and the tasks it measures. Experiment 3's protocol converts word segmentation into word/non-word classification, so Table II is not an apples-to-apples comparison with the cited segmentation algorithms. Experiment 1, the only motif-discovery experiment, uses fixed-length GunPoint and therefore does not exercise the variable-length mechanism that is the paper's novelty. The combination of lack of direct evidence for the central claims and optimistic parameter selection supports rejecting the abstract's assertions as currently stated. I would not move the verdict because the reader already reached REJECT; I agree with that outcome, though my path to it emphasizes the task mismatch rather than the truncated-update geometry.","tokens_in":10632,"tokens_out":7171,"duration_ms":78646,"concrete_test":"Run VILMAP as an actual segmenter on the Brent-Siskind corpus: train with the same phoneme-feature preprocessing, then for each utterance present its phoneme stream left to right; let the winning node's prototype length define a word-hypothesis ending at that position, and record the resulting boundary sequence. Score these boundaries against the gold-standard word boundaries using the same boundary token F-score used by [19] (e.g., the standard word-segmentation evaluation script). If the boundary F-score is close to Table II's 0.750 or above, the segmentation claim gains support; if it is far below, the claim that VILMAP is competitive on word segmentation is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that VILMAP can identify motifs of different lengths and achieves results similar or superior to other methods for word segmentation. The reported experiments do not support either half. Experiment 1 (Section IV-A) uses GunPoint, a fixed-length 150-point two-class dataset; finding class averages in a fixed-length setting does not test variable-length motif discovery, and no baseline comparison is reported. Experiment 3 (Section IV-C) is explicitly not a segmentation experiment: the text states the model 'does not identify precisely the word boundaries,' so the authors construct a binary word/non-word recognition task and compute F-measure on those pre-separated inputs. Table II then compares that recognition score with segmentation scores from [19] for DiBS, TPs, PUDDLE, and AGu; the metrics are not commensurable. A model that labels isolated word-length sequences cannot segment fluent speech, where boundaries are unknown. In addition, the reported 0.750 F-measure was selected as the best of 100 LHS parameter sets without a separate validation split or error bars, so it is an optimistic upper bound. The Section V admission that the evaluation is 'preliminary' is accurate but does not repair the mismatch between the measured task and the asserted task.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes VILMAP, an extension of the LARFDSSOM self-organizing map in which each node's center, relevance, and distance vectors can grow to match the length of the input pattern. Activation is computed by regular, sliding-window, or truncated comparisons, and node vectors are extended with hand-initialized entries when a smaller node wins over a larger input. The authors claim that VILMAP can discover time-series motifs of different lengths, avoids catastrophic forgetting when trained with inputs of increasing dimension, and achieves word-segmentation results similar to or better than published methods. Three experiments are reported: a fixed-length GunPoint motif-identification task, a TIMIT-derived phoneme experiment with input dimensions increasing from 24 to 72, and a Brent-Siskind word/non-word recognition experiment compared against DiBS, TPs, PUDDLE, and AGu.","tokens_in":10857,"tokens_out":4044,"duration_ms":48010,"significance":"If the variable-length extension worked as claimed, VILMAP would be a useful and apparently novel SOM variant, particularly for online, biologically motivated learning. Positive features of the submission include a sufficiently detailed algorithmic description, a public code repository, and an Experiment 2 design that targets the real and interesting failure mode of catastrophic forgetting under growing input dimensionality. However, the evidence as presented does not support the two headline claims: no experiment actually varies motif length during discovery, and the word-segmentation evaluation is a binary word/non-word recognition task, not a boundary-prediction task. The comparison against boundary-based segmentation systems is therefore not commensurable. The reported numbers are also optimistic upper bounds because they are selected as the best of 100 parameter samples on the test data without a validation split or error bars. The contribution is better framed as a preliminary proof-of-concept for a variable-length SOM extension, with the central claims substantially narrowed.","major_comments":[{"comment":"The only motif-discovery experiment uses the GunPoint dataset, whose inputs all have fixed length 150. This cannot support the abstract's claim that VILMAP can 'identify Motifs with different lengths in time series,' because no variable-length motif is presented and no quantitative error measure or baseline comparison is reported; the result is a visual match of prototypes to class averages. An experiment with known motifs of several lengths, or a subsequence-level evaluation, is needed for the central claim.","section":"Section IV-A"},{"comment":"The word-segmentation experiment explicitly concedes that VILMAP 'does not identify precisely the word boundaries'; instead, words and non-words are separated beforehand and the network is asked to recognize word-length inputs. This measures word/non-word classification, not segmentation. Consequently, Table II's comparison with DiBS, TPs, PUDDLE, and AGu, whose scores from [19] are boundary-based segmentation results, compares different tasks, and the reported F-measure of 0.750 cannot be read as a competitive segmentation result.","section":"Section IV-C"},{"comment":"The reported F-measures are selected as the best of 100 Latin Hypercube parameter sets evaluated on the test data, with no validation split, no error bars, and no significance tests. This makes the numbers optimistic upper bounds and prevents any claim of similarity or superiority to the cited methods. The authors should report the full distribution of results or select parameters on a separate validation set and then evaluate on held-out test data.","section":"Sections IV-B and IV-C"},{"comment":"The activation threshold a_t directly controls node creation, and the truncated comparison and truncated update rules for unequal node/input lengths are heuristic: when a node grows, new center entries are set to the input values and new relevance entries are set to 0.5. The manuscript gives no analysis or sensitivity study of how this initialization changes the existing distance geometry or how a_t interacts with the sliding-window and truncated activations. Because these rules determine node counts and downstream clustering, a threshold sweep or ablation would be needed before the model's behavior can be considered robust.","section":"Section III-A and III-B"}],"minor_comments":[{"comment":"The winner rule in Eq. (3) does not specify tie-breaking; with the shifting used for unequal lengths, an explicit tie rule would improve reproducibility.","section":"Eq. (3)"},{"comment":"The captions of Figures 4(a) and 4(b) do not define Procedure A and Procedure B; the text describes them partially, but the captions should be self-contained.","section":"Figure 4"},{"comment":"The construction of the False dataset is underspecified: the text does not state how many random sequences are generated, how labels are assigned, or how accidental matches to true words are handled, all of which are needed for replication.","section":"Section IV-B"},{"comment":"The paper refers to [19] for the comparison scores but does not state whether the same corpus split, tokenization, or scoring implementation is used; the evaluation protocol should be clarified.","section":"Section IV-C"},{"comment":"The authors state that parameters for Experiment 1 were adjusted by trial and error while Experiment 3 uses Latin Hypercube Sampling; the difference in tuning methodology should be explained and justified.","section":"Section IV-A"}],"recommendation":"reject","confidential_remarks":"For the editor: the manuscript's advertised contributions are not supported by the reported measurements. The decisive issue is that Experiment 3 measures word/non-word recognition rather than segmentation, so the central comparative claim fails, and the variable-length motif-discovery claim has no corresponding experiment. The released code, the algorithmic clarity, and the catastrophic-forgetting observation are salvageable in a substantially revised and more modest paper, but as submitted the gap between the claims and the evidence is too large for journal publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nVILMAP is a genuine extension of LARFDSSOM: it lets prototypes grow and compare inputs via a sliding window, which is a real gap in the SOM family. The code is up, the mechanism is clearly described, and the experiment on dimension growth does suggest the model can avoid catastrophic forgetting in a controlled setting. That is the paper's actual strength.\n\nThe soft spots are where the claims exceed the evidence. Experiment 3 is not a word segmentation experiment. The authors say the model does not identify word boundaries, then they feed it pre-separated words and non-words and compute F-measure. Comparing that number to segmentation algorithms is comparing recognition to segmentation; the metrics are not commensurable. The paper's abstract claim about superior word segmentation is therefore unsupported. On top of that, the reported 0.750 F-measure is the best of 100 LHS parameter samples selected on test data with no validation split and no error bars, so it is an optimistic upper bound. The GunPoint experiment is fixed-length and qualitative, so it does not demonstrate variable-length motif discovery either.\n\nThere is also a small gap in the mechanics: truncated update initializes new dimensions with relevance 0.5 and zero distance, and the paper doesn't analyze how that affects the activation threshold or the existing distance geometry. That's a minor concern relative to the experiment mismatch, but worth putting on the table.\n\nI agree with your stress-test. This is a plausible model with an honest preliminary evaluation, but the central empirical claims are not supported as stated. The authors themselves call it preliminary in Section V, and that is the right word.\n\nWho is this for? People working on growing SOMs or online learning with changing input dimensionality will find the mechanism useful. The psycholinguistics application would only be credible after a proper segmentation evaluation and a statistically grounded parameter selection.\n\nIt deserves a serious referee, not a desk reject, because the mechanism is new and the catastrophic forgetting result is worth checking. But as it stands it should not be accepted. I'd send it back for major revision with a demand for a real boundary-evaluation setup, a validation procedure, and error bars.","headline":"A real extension to variable-length SOMs undermined by an evaluation that measures word recognition, not segmentation.","tokens_in":11356,"tokens_out":2068,"would_cite":false,"duration_ms":21665,"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":"VILMAP extends self-organizing maps to variable input lengths, enabling motif discovery and word segmentation while avoiding catastrophic forgetting.","keywords":["self-organizing maps","variable input length","time series motif discovery","word segmentation","subspace clustering","catastrophic forgetting","online learning"],"falsifier":"Train VILMAP on a set of short patterns, then on the same patterns extended by a constant tail, and then present the unextended short patterns again. If the map no longer clusters them as before (for example, if they activate a different node or the F-measure on the short set drops), the truncated update has changed the stored geometry and catastrophic forgetting is not avoided. Similarly, on a labeled time series with motifs of two different lengths, one could compare VILMAP's clusters against a brute-force sliding-window motif search to see whether motifs of both lengths are recovered.","tokens_in":1563,"feed_emoji":"🧠","tokens_out":2122,"duration_ms":80342,"temperature":0.7,"pith_summary":"The paper proposes VILMAP, a self-organizing map whose prototype nodes can have different lengths, so one map can cluster time-series patterns of varying size without a fixed input window. The motivating applications are time series motif discovery and word segmentation in speech transcribed as phoneme sequences. The authors claim this is the first SOM-based method that supports inputs with different lengths, that it finds motifs in a standard dataset, and that it avoids catastrophic forgetting when the input dimension grows during training. On word segmentation it reports F-measure 0.750, precision 0.856, and recall 0.667, outperforming three of four baselines and losing only to AGu, a batch model with infinite memory. The significance would be a biologically plausible, online neural model for a language-acquisition problem usually handled by statistical or Bayesian segmenters.","feed_headline":"A growing self-organizing map learns motifs of many lengths","feed_subtitle":"It beats three of four word-segmentation baselines and avoids forgetting old patterns as inputs grow.","key_machinery":"The central mechanism is the weighted activation function of LARFDSSOM adapted to variable lengths. Each node stores a center, a relevance vector, and a distance vector; activation is $\\sum_i \\omega_{ji} / (\\sum_i \\omega_{ji} + D_\\omega(x, c_j) + \\epsilon)$, where $D_\\omega$ is the weighted Euclidean distance. Three comparison modes (regular, sliding window, and truncated) decide how a node of one length matches an input of another, and the truncated update grows all three vectors when a smaller winning node meets a longer input. This growth rule is what lets the map handle increasing input sizes and what the authors say prevents catastrophic forgetting.","core_discovery":"On its own terms, the central discovery is that the dimension-selective update rules of LARFDSSOM can be extended to variable-size inputs by adding three length-matching modes. When node and input match, the original distance and update apply. When the input is shorter than the node, a sliding-window comparison takes the best shift and updates only the overlapping part. When the input is longer, the node's center, relevance, and distance vectors grow to the input size, with the new center entries copied from the input, new relevance entries set to $0.5$, and new distance entries set to zero, and only then is the standard update applied. The paper argues these rules let the map discover motifs of different lengths and, because growing a node's dimensions does not disturb its old dimensions, let it keep performing on short inputs after training on long ones.","pith_inferences":["A natural test the paper does not run is whether VILMAP finds a motif that appears at several different lengths within the same time series, since the sliding-window mode only handles an input shorter than a node.","The neutrality of the $0.5$ initialization for new relevance dimensions is an assumption that could be tested by ablation: starting new dimensions at relevance $1$ might let long patterns distort old clusters, while starting at $0$ might prevent learning them.","If the word-segmentation result transfers, it suggests that boundary-free whole-word recognition could be a viable developmental pathway, complementing statistical segmentation models rather than competing with them.","A concrete next step would be feeding VILMAP progressively longer sequences of subword units and checking whether the map's stored prototypes grow compositionally, which would test whether the dimension-growth tolerance reaches beyond 72 features."],"forward_implications":["A single VILMAP can be trained on a corpus of utterances of varying phonetic length and then used to cluster new utterances without knowing the motif length in advance.","Because the map is online and passes through the data once, it offers a candidate neural account of how infants could track recurring phoneme sequences without batch statistics.","The lack of catastrophic forgetting when dimensions grow suggests the model can be a building block for a recurrent growing map that learns longer expressions over time, as the authors propose.","The competitive word-segmentation scores come without detecting exact word boundaries, implying that recognizing words as whole inputs is enough for much of the segmentation task."],"supporting_citations":[{"why":"Supplies the LARFDSSOM model whose relevance and receptive-field updates VILMAP inherits and extends.","marker":"[5]"},{"why":"Introduces the weighted distance and dimension-selective relevance mechanism that the activation function is built on.","marker":"[12]"},{"why":"Defines variable-length motif discovery and motivates the need for methods accepting different sample sizes.","marker":"[13]"},{"why":"Provides the GunPoint dataset used as the standard motif-discovery test.","marker":"[14]"},{"why":"Provides the word-segmentation protocol and baseline results VILMAP is compared against.","marker":"[19]"},{"why":"Describes the Adaptor Grammar baseline (AGu) that achieves the best F-measure in the comparison.","marker":"[23]"},{"why":"Supplies the child-directed speech corpus used for the word-segmentation experiment.","marker":"[28]"}],"fun_headline_variants":["Self-organizing map that grows to fit motifs of any length","Variable-length SOM finds motifs, avoids forgetting","Growing map learns motifs and words without forgetting","SOM with adaptable input length beats word-seg baselines","Map that grows with inputs discovers motifs, segments words"],"cache_read_input_tokens":13568,"weakest_assumption_plain":"The claim rests on the assumption that when a node is stretched to a longer input, initializing the newly added dimensions with neutral relevance and zero distance leaves everything the node already learned intact, so longer inputs do not corrupt shorter-pattern clusters.","fun_headline_variants_meta":{"raw":{"variants":["Self-organizing map that grows to fit motifs of any length","Variable-length SOM finds motifs, avoids forgetting","Growing map learns motifs and words without forgetting","SOM with adaptable input length beats word-seg baselines","Map that grows with inputs discovers motifs, segments words"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1315,"prompt_tokens":875,"completion_tokens":440,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":364}},"tokens_in":491,"tokens_out":440,"duration_ms":5224,"temperature":1.0,"reasoning_tokens":364,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:32:05.245365+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train VILMAP on a set of short patterns, then on the same patterns extended by a constant tail, and then present the unextended short patterns again. If the map no longer clusters them as before (for example, if they activate a different node or the F-measure on the short set drops), the truncated update has changed the stored geometry and catastrophic forgetting is not avoided. Similarly, on a labeled time series with motifs of two different lengths, one could compare VILMAP's clusters against a brute-force sliding-window motif search to see whether motifs of both lengths are recovered.","supporting_citations":[{"cited_title":"Dimension Selective Self- Organizing Maps With Time-Varying Structure for Subspace and Projected Clustering,","cited_arxiv_id":null,"evidence_quote":"Supplies the LARFDSSOM model whose relevance and receptive-field updates VILMAP inherits and extends."},{"cited_title":"Dimension selective self-organizing maps for clustering high dimensional data,","cited_arxiv_id":null,"evidence_quote":"Introduces the weighted distance and dimension-selective relevance mechanism that the activation function is built on."},{"cited_title":"Dis- covery of variable length time series motif,","cited_arxiv_id":null,"evidence_quote":"Defines variable-length motif discovery and motivates the need for methods accepting different sample sizes."},{"cited_title":"The ucr time series classification archive,","cited_arxiv_id":null,"evidence_quote":"Provides the GunPoint dataset used as the standard motif-discovery test."},{"cited_title":"Relating unsupervised word segmentation to reported vocabulary acquisition,","cited_arxiv_id":null,"evidence_quote":"Provides the word-segmentation protocol and baseline results VILMAP is compared against."},{"cited_title":"Adaptor grammars: A framework for specifying compositional nonparametric bayesian mod- els,","cited_arxiv_id":null,"evidence_quote":"Describes the Adaptor Grammar baseline (AGu) that achieves the best F-measure in the comparison."},{"cited_title":"The role of exposure to isolated words in early vocabulary development,","cited_arxiv_id":null,"evidence_quote":"Supplies the child-directed speech corpus used for the word-segmentation experiment."}],"review_version":1}