{"id":"799216fb-e20d-4220-8bfe-5a19ca9bf66e","arxiv_id":"2505.11817","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An exemplar-free keyword spotter updates a linear classifier with recursive least squares on frozen acoustic features, matching joint-training accuracy without replay buffers.","lead":"AnalyticKWS lets a keyword-spotting model learn new voice commands without storing any old recordings, by updating the classifier with a single calculation instead of backpropagation. The method reports near-joint accuracy on up to 100 keywords while using far less memory and training time than replay-based baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (7), the update claimed to replicate the joint analytic least-squares solution, is malformed as printed and the appendix proof contains a sign error; the equivalence with joint learning is therefore unverified.","rationale":"My stress test agrees with the reader's overall CONDITIONAL verdict, but the most load-bearing concern is not the frozen-feature generalization limit. The reader's weakest_assumption is that task-0 features may not linearly separate future keywords; that is a real boundary condition, and the paper honestly acknowledges it. However, the method's strongest claim is a mathematical one: Eq. (7) is supposed to be exactly the recursive form of Eq. (5), giving 'absolute memorization' and equivalence with joint analytic learning. If Eq. (7) is algebraically invalid, that guarantee is absent regardless of feature quality. The appendix is not just incomplete; it contains a sign error and a reduction that drops the new-task contribution. The empirical results are suggestive and the RLS approach is credible, so rejection is not warranted; the correct response is a conditional acceptance requiring a corrected derivation, a precise statement of the update (including how new-class columns are added), released code, and repeated-run statistics. Since the reader already reached CONDITIONAL, my verdict recommendation is UNCHANGED, but the reasoning should foreground the Eq. (7) derivation failure rather than only the feature-transfer boundary.","tokens_in":15041,"tokens_out":8734,"duration_ms":87984,"concrete_test":"Implement Eq. (5), the literal Eq. (7) as printed, and the standard recursive least-squares form θ_t = θ_{t−1} + A_t S_t'^T (Y_t − S_t' θ_{t−1}) on randomly generated feature and label blocks for 3–5 small tasks. Compute the Frobenius norms ‖θ_literal − θ_joint‖ and ‖θ_RLS − θ_joint‖. If the literal equation fails while the standard RLS form matches the closed-form joint solution, the paper's stated formula is a typo but the intended method is recoverable; if the standard form also fails under the block-diagonal label construction used in Eq. (4), then the claimed equivalence is false for the class-incremental formulation as defined.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the recursive update Eq. (7) is identical to the joint solution Eq. (5). As printed, Eq. (7) is not a well-formed matrix expression: it reads θ_t = [θ_{t−1} − A_t S_t^T S'_t θ_{t−1} A_t S'_t^T y_t], with no operator between the last two terms. The appendix proof then states in Eq. (b) that θ_t = A_t [Q_{t−1} − S_t'^T Y_t], but the joint solution Eq. (5) requires a plus sign: A_t [Q_{t−1} + S_t'^T Y_t]. The subsequent reduction in Eq. (d), θ_t = θ_{t−1} − A_t S_t'^T θ_{t−1}, drops the new-task term and is therefore not equivalent to Eq. (5). This is not merely a missing derivation step: the stated algebra does not establish the claimed identity. The high-level recursive least-squares idea is plausible and similar identities appear in the analytic learning literature, but the paper must supply a correct derivation of the exact update used in the experiments. Without that, the strongest claim, that incremental learning gives the same answer as joint learning, is unsupported. The frozen-feature generalization concern is a boundary condition on usefulness; the Eq. (7) defect is a direct challenge to the method's core theoretical guarantee.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AnalyticKWS, an exemplar-free class-incremental learning method for small-footprint keyword spotting. A CNN backbone is first trained on task 0 and then frozen. A random linear expansion layer (AFE) is inserted, and the classifier is recalibrated with ridge regression. For each subsequent keyword task, the method updates a stored feature autocorrelation matrix (AFAM) and the analytic classifier through a recursive least-squares rule, claiming that this incremental update is identical to the joint analytic least-squares solution over all tasks. Experiments on GSC-v1, GSC-v2, and SC-100 with up to 51 tasks compare against EWC, iCaRL, BiC, Rwalk, RK, and DE-KWS, reporting accuracy, backward transfer, per-epoch training time, and extra memory. The reported results show strong accuracy and much lower training time than the baselines, without storing any exemplars.","tokens_in":15303,"tokens_out":5151,"duration_ms":54636,"significance":"If the recursive equivalence is correctly established and the experimental results are reproducible, the method is genuinely attractive for edge-device KWS: it is exemplar-free, privacy-preserving, requires only one analytic update per task, and reports near-joint accuracy with very low per-task training time. The paper has clear strengths: it tackles a practical problem, evaluates on three datasets including a 100-class setting, reports computational efficiency, and builds on a standard recursive-least-squares formulation rather than ad hoc fitting. However, the central theoretical derivation as printed contains a malformed update equation and an appendix proof with sign and algebraic errors, so the paper's strongest claim — that incremental learning gives the same answer as joint learning — is not currently supported. The frozen-feature premise is also load-bearing and is only partially acknowledged in the limitations paragraph.","major_comments":[{"comment":"The recursive update printed as Eq. (7) is not a well-formed matrix expression: it reads θ_t = [θ_{t−1} − A_t S_t^T S'_t θ_{t−1} A_t S'_t^T y_t], with no operator between the last two terms, and it mixes S_t and S'_t notation. If the intended formula is the standard recursive least-squares update, it should take the form θ_t = θ_{t−1} + A_t S_t'^T (Y_t − S_t' θ_{t−1}) (or an algebraically equivalent expression). As printed, Eq. (7) cannot be evaluated and is not shown to equal Eq. (5). This is the central theoretical guarantee of the paper and must be corrected with a full, valid derivation of the exact update implemented in the experiments.","section":"§3.2.3, Eq. (7)"},{"comment":"The appendix proof contains a sign error and a dropped term. Step (b) writes θ_t = A_t(Q_{t−1} − S_t'^T Y_t), but the joint solution in Eq. (5) requires a plus sign, Q_{t−1} + S_t'^T Y_t. Step (d) then reduces the expression to θ_t = θ_{t−1} − A_t S_t'^T θ_{t−1}, which discards the new-task term Y_t. The final line of the proof is therefore not equivalent to Eq. (5). The recursive-least-squares identity is standard and likely repairable with a correct Woodbury-based derivation, but the proof as written does not establish the claimed equality, and the manuscript cannot be accepted with this defect in the core argument.","section":"Appendix A, steps (b) and (d)"},{"comment":"The method's practical accuracy depends on the premise that features from the task-0-trained, frozen CNN extractor, after a random linear expansion, are linearly separable for all future unseen keywords. The analytic update and AFAM can only rearrange the fixed feature space; they cannot recover information lost because the backbone is never adapted. The limitations paragraph concedes that the extractor 'might not be optimal for every domain.' Because this assumption is load-bearing for the claim of near-joint accuracy, the paper should provide evidence: for example, vary the size and composition of task 0, measure linear separability of the AFE features for future tasks, or show what happens when task 0 is small or unrepresentative.","section":"§3.2.1 and Limitations"},{"comment":"The Joint upper-bound row reports only three accuracy values (94.93, 94.76, 95.32) for the nine experimental settings; no joint accuracy is reported for GSC-v2 or SC-100. Since the paper repeatedly claims that AnalyticKWS approaches joint-training performance, missing the upper bound for most settings weakens this claim. Please report joint accuracy for every dataset and task configuration, or explicitly state that a single joint model is used for all settings and justify why that is appropriate.","section":"Table 1"}],"minor_comments":[{"comment":"The notation is inconsistent: the left factor uses S_t and the right factor uses S'_t; the expanded feature matrix should be denoted uniformly, e.g., S'_t throughout the update.","section":"§3.2.3, Eq. (7)"},{"comment":"The label matrix y_0 is used without definition; please state explicitly that it is a one-hot encoded matrix whose rows correspond to the samples in S'_0.","section":"§3.2.2, Eq. (2)"},{"comment":"There is a typo in the text: 'AFAuM' should be 'AFAM.' The sentence 'Now we proof calculate' is also ungrammatical.","section":"Appendix A"},{"comment":"The memory comparison is hard to follow because the dimensions behind '16000×1×1×500' are not explained. Please give the tensor shape or number of stored scalars for each method.","section":"§5.3"},{"comment":"The ablation finds the best accuracy with a 512-dimensional expansion, but the main results use AnalyticKWS-128 and AnalyticKWS-256. Please clarify why the best configuration is not used in the headline comparisons, or add AnalyticKWS-512 to Table 1.","section":"Table 3"},{"comment":"The experimental setup does not specify the regularization coefficient γ, the pretraining epochs and optimizer for task 0, or the hyperparameters for baselines such as EWC and BiC. These details should be provided to make the comparisons reproducible.","section":"§4.2 and Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is essentially an application of the ACIL/BRMP recursive-least-squares idea to keyword spotting, with the AFE expansion as an additional component. The central proof defect is fixable and is not a reason to reject on novelty grounds, but it must be repaired before the paper can be considered for publication. The authors should also carefully distinguish their contribution from the existing analytic class-incremental learning literature in the final version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read of AnalyticKWS. The genuinely useful part is the recipe: freeze a small CNN trained on the first task, expand its features with a fixed random projection, and update a linear classifier via recursive least squares, with no exemplars and no gradient steps for new keywords. The empirical work is solid and reasonably broad—three datasets, multiple task splits, accuracy/BWT/time/memory comparisons, and an ablation of the expansion size and regularization. The efficiency numbers are the real story: orders of magnitude less training time than rehearsal baselines, with accuracy close to joint training. For edge KWS, that's worth paying attention to.\n\nThe soft spot is the theoretical core. Eq. (7), the update claimed to reproduce the joint analytic solution, is not a well-formed matrix expression as printed—an operator is missing between terms. The appendix proof then gets the sign wrong relative to Eq. (5) and, in step (d), drops the new-task contribution entirely. So the paper's headline guarantee, that incremental learning equals joint learning, is not established by the text. The high-level identity is standard recursive least squares and was already in ACIL, which the paper cites, so I suspect the method works if implemented correctly. But a reader cannot verify that from this manuscript.\n\nAlso: no code, no repeated-run statistics, and the AFE expansion size appears to be tuned on the same benchmarks used for the final comparison. Those are fixable but should be addressed. The frozen-extractor concern (task-0 features may not generalize to future keywords) is acknowledged in the Limitations paragraph; I read it as a boundary condition, not a fatal flaw, and it doesn't affect the theoretical claim.\n\nCitation pattern looks honest: ACIL is credited, and the new contributions—the feature expansion, the KWS formulation, and the SC-100 evaluation—are real even if the core update is inherited.\n\nWho should read this: researchers working on on-device continual learning for speech. It's a well-targeted engineering contribution with a genuine privacy/efficiency angle. My recommendation: send it to peer review. A competent referee can check whether the implemented update matches ACIL, ask for code and variance estimates, and require a corrected derivation. I would not reject it outright; I'd ask for the theory to be fixed and the experimental reporting tightened.\n\nBest.","headline":"A useful exemplar-free KWS recipe with strong numbers, but the paper's central proof is botched as printed and needs correction and code before the joint-learning claim can be trusted.","tokens_in":15883,"tokens_out":3700,"would_cite":false,"duration_ms":37502,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a keyword-spotting model can keep learning new spoken commands from a closed-form recursive matrix update, storing no old audio, and reach accuracy matching joint training on all tasks.","keywords":["analytic learning","class incremental learning","exemplar-free continual learning","keyword spotting","catastrophic forgetting","recursive least squares","acoustic feature expansion","small-footprint KWS"],"falsifier":"A decisive check is to compare, on the same held-out feature matrices, the weights produced by the recursive update of Eq. (7) with those of the direct joint least-squares solution of Eq. (5): any difference beyond floating-point error would refute the claimed identity. Separately, if task 0 is drawn from one acoustic domain and later tasks from another, and final accuracy falls well short of the joint fit on the same frozen features, the frozen-extractor assumption is the failure point.","tokens_in":14788,"feed_emoji":"🎙️","tokens_out":6031,"duration_ms":61786,"temperature":0.7,"pith_summary":"The paper tries to show that keyword spotting can continually add new words without ever replaying old audio, by replacing gradient-based fine-tuning with a closed-form recursive least-squares update. If correct, continual learning on small devices no longer requires an exemplar buffer, so user data stays on-device and private. Each new task is adapted in a single epoch, without back-propagation, and the only stored artifact is a small square correlation matrix. The experiments report near-joint accuracy and near-zero backward transfer across up to 100 keywords, with far lower training time and memory than rehearsal-based baselines.","feed_headline":"One matrix update adds keywords with no stored audio","feed_subtitle":"AnalyticKWS learns new commands in a single epoch per task, storing only a small correlation matrix instead of past audio.","key_machinery":"The load-bearing object is the acoustic feature auto-correlation matrix (AFAM), $A_{t-1} = \\left(\\sum_{i=0}^{t-1} S_i'^\\top S_i' + \\gamma I\\right)^{-1}$, together with the recursive classifier update asserted in Eq. (7). The AFAM encodes the inverse regularized covariance of all expanded features seen so far; updating it and the classifier via the Woodbury matrix identity reproduces the joint least-squares solution without revisiting old data. A fixed random linear expansion layer, the acoustic feature expansion (AFE), raises the feature dimension before the fit, and the paper's ablation shows accuracy rising with expansion size.","core_discovery":"The central claim is that, once a CNN feature extractor is frozen and the acoustic features are passed through a fixed random linear expansion, the incremental classifier update computed only from current-task features and a stored correlation matrix is algebraically identical to the joint analytic least-squares solution over all tasks. The paper states this in Section 3.2.3: the recursive formulation gives the same answer as joint learning. Consequently, each new keyword task can be learned in one epoch without back-propagation and without storing any past utterances, because the acoustic feature auto-correlation matrix (AFAM) summarizes all past feature correlations and is itself updated recursively. The reported evidence is that AnalyticKWS matches or exceeds rehearsal-based baselines in accuracy and backward transfer on GSC-v1, GSC-v2, and SC-100, while using only a few seconds per task and a tiny constant memory footprint.","pith_inferences":["The recursive identity behind the AFAM is not specific to speech; the same update should transfer to any class-incremental problem with a frozen feature extractor, meaning the privacy and memory benefits are generic properties of analytic continual learning rather than of keyword spotting.","Since the frozen task-0 feature extractor is the single point of failure, a natural testable extension is to periodically adapt the expansion layer or maintain a small set of feature anchors from the current task, trading a little privacy for plasticity; the paper's own limitation section points at this gap.","The flat accuracy across T=11 to T=51 suggests that, for a fixed frozen backbone, the bottleneck is feature quality rather than forgetting; comparing different backbone capacities under the same analytic update would isolate how much of the gain comes from the representation versus the update rule.","The ablation trend of improving accuracy with expansion size up to 512 dimensions suggests a testable extension: whether even larger, or learned rather than random, expansions continue to help or saturate."],"forward_implications":["No user audio needs to be stored for continual KWS adaptation, so privacy requirements such as user deletion requests are satisfied by construction.","Per-task adaptation requires one epoch and no gradient back-propagation, cutting training time on the reported hardware from hundreds of seconds to single digits per task.","The AFAM is a square matrix whose size depends only on the expanded feature dimension (e.g., 128 by 128), so extra memory stays small and constant as the number of tasks grows.","Accuracy and backward transfer approach those of joint training, with near-zero BWT across up to 51 tasks even when rehearsal baselines use a 500- or 1000-sample buffer.","Small-footprint backbones like TC-ResNet-8 can keep incremental keyword spotting on-device, because the only saved artifact is the AFAM plus the current classifier weights."],"supporting_citations":[{"why":"Supplies the analytic class-incremental learning formulation and the recursive least-squares update that AnalyticKWS adapts to speech.","marker":"(Zhuang et al., 2022)"},{"why":"Provides the recursive least-squares (BRMP) update rule that lets the AFAM be updated without revisiting old samples.","marker":"(Zhuang et al., 2021)"},{"why":"Defines the TC-ResNet-8 backbone used as the frozen acoustic feature extractor.","marker":"(Choi et al., 2019)"},{"why":"A rehearsal-based incremental KWS baseline that AnalyticKWS must beat while using no buffer.","marker":"(Xiao et al., 2022a)"},{"why":"The most recent rehearsal-based KWS baseline, combining dark knowledge distillation with exemplar storage, compared against AnalyticKWS.","marker":"(Peng and Xiao, 2024)"},{"why":"Supplies the EWC regularization-based continual learning baseline used in the comparisons.","marker":"(Kirkpatrick et al., 2017)"},{"why":"Source of the SC-100 dataset used for the large-scale 100-keyword incremental evaluation.","marker":"(Song et al., 2024)"},{"why":"Supplies the iCaRL exemplar-based baseline whose fixed-size buffer degrades as the number of tasks grows.","marker":"(Rebuffi et al., 2017)"}],"fun_headline_variants":["One matrix update adds keywords with no stored audio","Closed-form solution learns new keywords without old data","Exemplar-free KWS: single-epoch updates, tiny memory","No backprop, no replay: analytic update for new keywords"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a CNN feature extractor, pretrained only on the first task and then frozen, yields features in which all later, unseen keywords are linearly separable after a fixed random expansion.","fun_headline_variants_meta":{"raw":{"variants":["One matrix update adds keywords with no stored audio","Closed-form solution learns new keywords without old data","Exemplar-free KWS: single-epoch updates, tiny memory","No backprop, no replay: analytic update for new keywords"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000402,"raw_usage":{"total_tokens":2101,"prompt_tokens":955,"completion_tokens":1146,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":1079}},"tokens_in":571,"tokens_out":1146,"duration_ms":12137,"temperature":1.0,"reasoning_tokens":1079,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:47:40.442121+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check is to compare, on the same held-out feature matrices, the weights produced by the recursive update of Eq. (7) with those of the direct joint least-squares solution of Eq. (5): any difference beyond floating-point error would refute the claimed identity. Separately, if task 0 is drawn from one acoustic domain and later tasks from another, and final accuracy falls well short of the joint fit on the same frozen features, the frozen-extractor assumption is the failure point.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the analytic class-incremental learning formulation and the recursive least-squares update that AnalyticKWS adapts to speech."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the recursive least-squares (BRMP) update rule that lets the AFAM be updated without revisiting old samples."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the TC-ResNet-8 backbone used as the frozen acoustic feature extractor."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the EWC regularization-based continual learning baseline used in the comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the iCaRL exemplar-based baseline whose fixed-size buffer degrades as the number of tasks grows."}],"review_version":1}