{"id":"05fe16ea-4ea2-48a0-a5f6-6470866504a6","arxiv_id":"2508.20732","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using random projections and ridge regression on frozen PANNs audio embeddings, the proposed method learns new classes and domains online without base-task training, outperforming linear probes and nearest-class-mean baselines on ESC-50 and TAU Urban Acoustic Scenes 2019.","lead":"This paper adapts a known continual-learning recipe, random projections plus ridge regression, to audio classification. It shows that a frozen pretrained audio model can learn new sound classes and new acoustic domains in a single pass, with low forgetting.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline numbers may be inflated by choosing Q on the same test tasks, and the method is never compared to RanPAC, the approach it directly extends.","rationale":"The reader's weakest assumption already flagged the projection dimension being selected on the same evaluation tasks and the lack of comparison to actual SOTA. I agree with that and make it the central concern: the reported numbers are likely optimistic because Q is chosen from the test-task ablation, and the method is an audio application of RanPAC without a direct RanPAC baseline. The math itself is coherent, the method is reproducible in principle, and the claim 'outperforms the investigated baselines' is supported by Table 1. Thus the appropriate verdict remains CONDITIONAL, not stronger or weaker. The concern is about external validity and evaluation protocol, not internal inconsistency. My agreement is partial because I place additional weight on the missing RanPAC comparison, which is not in the reader's weakest_assumption but is in the reader's rationale.","tokens_in":8456,"tokens_out":7746,"duration_ms":88471,"concrete_test":"Fix Q by nested validation: on ESC-50, choose Q from {2048,4096,8192,16384} using only the validation fold, then retrain with that Q and evaluate on the held-out test fold; on TAU, choose Q on a subset of cities/domains and test on held-out cities. Also run RanPAC [24] (both with and without base-task fine-tuning) on the same frozen PANNs embeddings and task orders. If the held-out-chosen Q is not 8192, or if RanPAC matches/exceeds the reported AAT, the central superiority claim is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the proposed method is a strong, base-task-free online learner for audio CIL/DIL rests on the reported final accuracies (93.4% CIL, 61.4% DIL). These numbers use Q=8192, selected from the ablation in Fig. 4 on exactly the same ESC-50 and TAU evaluation tasks that are later reported. Since the test split is used to pick Q, the final accuracies are selected estimates, not unbiased evaluations of a pre-specified method. The per-task lambda is also tuned on the current task's validation split, which is less concerning but adds another selection step. In addition, the paper explicitly says the method relies on RanPAC, yet RanPAC [24] itself is never run on these audio tasks; the only comparisons are linear-probe variants and NCM. Therefore the paper supports 'outperforms the methods we implemented,' but not the broader conclusion that the approach is suitable and superior for arbitrary audio CIL/DIL tasks. A held-out-task evaluation and a direct RanPAC comparison are needed before the reported superiority should be taken at face value.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an online incremental learning method for audio classification using a frozen PANNs CNN14 model. A random ReLU projection expands the 2048-dimensional embeddings to Q=8192, and ridge regression is applied to accumulated Gram and class-sum matrices (Eqs. 1-3). The method is evaluated on class-incremental learning with ESC-50 (93.4% final average accuracy) and domain-incremental learning with TAU Urban Acoustic Scenes 2019 (61.4%), and is compared to linear-probe, joint-linear-probe, and NCM baselines. The authors conclude the approach is suitable for both CIL and DIL and outperforms all investigated methods.","tokens_in":8722,"tokens_out":4562,"duration_ms":49219,"significance":"If the reported results hold, the paper offers a simple, base-task-free continual audio learner that leverages a strong pretrained representation and requires no iterative training. The ablation study in Figs. 3-4 provides useful evidence that the nonlinear random projection is the key component. The method is reproducible from the public PANNs model and standard datasets. However, the evaluation protocol and the absence of a direct comparison to RanPAC, the method on which it explicitly relies, currently prevent the broad superiority claim from being accepted at face value.","major_comments":[{"comment":"The projection dimension Q=8192 is selected via ablation on exactly the same ESC-50 and TAU tasks that are later used to report the final accuracies. Because Fig. 4 shows average accuracy after each task (i.e., on the test splits), the headline numbers 93.4% and 61.4% are selected estimates, not unbiased evaluations of a pre-specified configuration. A held-out-task protocol or nested selection is needed; at minimum, the numbers should be labeled as having Q selected on the test tasks.","section":"§3.2, Fig. 4"},{"comment":"The paper claims adaptation 'in a single forward pass' (Abstract and §2.2). However, the λ-selection procedure requires updating G and K for 17 candidate λ values on a random 80% subset of the current task, each involving the Q×Q inverse in Eq. (3), followed by a final update over the full task. This is not a single forward pass. If λ tuning is intended as separate hyperparameter selection, that should be stated explicitly and its computational cost quantified.","section":"§2.2, Eq. (3)"},{"comment":"The method is described as 'relies on RanPAC [24]', yet RanPAC is never run on the audio tasks, and no comparison to prior audio continual-learning methods (e.g., [4], [7], [10]) is provided. The conclusion that the method 'outperforms all other investigated methods' therefore only covers the simple baselines implemented in this paper. A direct RanPAC comparison under the same CIL/DIL protocols is necessary to support the claimed advantage.","section":"§2, §4, Table 1"},{"comment":"The advantage over the offline joint linear probe is 1.9 percentage points for CIL and 1.1 for DIL, with only n=5 runs. Given the small number of runs and the overlapping standard deviations (DIL: 61.4±0.1 vs 60.3±0.7), statistical significance testing or confidence intervals should be provided before making the 'outperforms' claim.","section":"Table 1"}],"minor_comments":[{"comment":"The wording 'outperforms all other investigated methods' should be qualified as 'baselines implemented in this paper' to avoid overclaiming.","section":"Abstract and §4"},{"comment":"K is defined as unnormalized class sums ('CP for class y without averaging'). In balanced datasets this is harmless, but for imbalanced tasks the lack of normalization can bias Wo. This design choice should be justified or at least noted.","section":"Eq. (2)"},{"comment":"For the NCM baseline it is not stated whether cosine similarity is computed on the original 2048-dimensional embeddings or on the projected features. Please clarify, since Figs. 3-4 suggest the projection is central to the proposed method.","section":"§3.1"},{"comment":"The term 'online' is used to mean a single pass over the task data rather than a streaming/sample-by-sample setting; consider making this explicit in the definitions.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a straightforward adaptation of RanPAC to frozen audio embeddings; its main value is the empirical demonstration on ESC-50 and TAU. The blocking issues are the test-set-based selection of Q and the missing RanPAC baseline. If the authors add a proper held-out evaluation and a direct RanPAC comparison, the paper could be acceptable as an application-oriented contribution. The novelty relative to [24] is modest, so the framing should emphasize the audio-specific analysis rather than a new algorithmic claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a straightforward, well-documented adaptation of RanPAC to audio classification, with one meaningful twist: there is no base-task fine-tuning, and updates are fully online (single forward pass) on frozen PANNs CNN14 embeddings. The method is exactly RanPAC's random projection + ReLU + regularized inverse on accumulated Gram and label sums, so the novelty is modest, but the paper explicitly acknowledges this and doesn't oversell the algorithm itself.\n\nWhat's actually new is the application to audio with a frozen generic encoder, and the demonstration that it works for both class-incremental and domain-incremental setups without touching the backbone. The ablations on projection dimension and nonlinearity are useful, and the evaluation using average accuracy and forgetting is standard. The results are internally consistent and the improvement over the implemented baselines is plausible.\n\nThe soft spots are real but not fatal. First, Q=8192 is chosen from an ablation on the same ESC-50 and TAU tasks that are later reported, so the final accuracies are selected estimates. Second, the comparison set is weak: no RanPAC itself, no prior audio continual learning method, only linear probes and NCM. The DIL gain over a joint linear probe is only 1.1 percentage points, which undercuts any strong superiority claim. Third, per-task lambda tuning on a validation split is okay, but it adds another selection step.\n\nNone of these invalidate the paper. The central claim is defensible but should be read as 'outperforms the methods we implemented' rather than a general state-of-the-art result.\n\nThis paper is for researchers who want a quick, cheap baseline for audio continual learning with a pretrained encoder. It's not a breakthrough, but it's a clean, reproducible empirical study. I would send it to peer review, and I'd ask for (1) a direct RanPAC comparison or a clear explanation of why it's inapplicable, (2) a held-out task for selecting Q, and (3) at least one prior audio IL baseline. With those revisions it would be a solid contribution.","headline":"A clean, honest RanPAC adaptation to audio; the key weakness is Q chosen on the test tasks and the absence of a RanPAC baseline.","tokens_in":9158,"tokens_out":2499,"would_cite":false,"duration_ms":24450,"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 frozen pretrained audio model can keep learning new sound classes and new recording domains online, in a single forward pass per task, with minimal forgetting.","keywords":["online incremental learning","class-incremental learning","domain-incremental learning","audio classification","pretrained audio model","random projections","ridge regression","catastrophic forgetting"],"falsifier":"Run the identical online protocol on a held-out sequence of sound classes or acoustic domains that are semantically far from AudioSet, such as medical or marine sounds, keeping Q = 8192 and the per-task λ search unchanged. If single-pass accuracy drops to the level of a plain linear probe, the claim that the random expansion and decorrelation are what prevent forgetting would be refuted.","tokens_in":8390,"feed_emoji":"🎧","tokens_out":4922,"duration_ms":51516,"temperature":0.7,"pith_summary":"The paper claims that a frozen pretrained audio model can serve as a permanent feature extractor for online incremental learning: when new sound classes or new recording domains arrive, the model adapts in a single forward pass through the new data and forgets little of what it learned before. The key move is to insert a random, frozen projection layer with ReLU between the pretrained embeddings and the classifier, then update a Gram matrix and a class-prototype matrix and solve a ridge-regression readout. On ESC-50 class-incremental learning the method reaches 93.4% average accuracy, and on TAU Urban Acoustic Scenes 2019 domain-incremental learning it reaches 61.4%, beating every baseline the paper compares against, including joint linear probes that are allowed to see all previous data. If this holds, incremental audio classification no longer needs base-task training, replay buffers, task IDs, or multi-epoch fine-tuning when new classes or domains appear.","feed_headline":"Single-pass audio learner hits 93.4% on 50-class task","feed_subtitle":"A pretrained model plus random projection learns new classes and cities in one pass, no retraining.","key_machinery":"The central mechanism is the RanPAC recipe adapted to audio: a random nonlinear expansion followed by a ridge-regression readout. Frozen 2048-dimensional pretrained embeddings are projected by a fixed random matrix W into Q = 8192 dimensions, passed through ReLU, and accumulated into a Q×Q Gram matrix and a Q×C class-prototype matrix; inference uses Wo = (G + λI)^−1 K. The random expansion improves linear separability of classes, while the ridge inversion decorrelates the class prototypes so that updating the readout on new tasks does not overwrite the classification structure of old tasks.","core_discovery":"The paper demonstrates that a fixed pretrained audio encoder can be turned into an online continual learner without any base-task training. Features from the frozen PANNs CNN14 model are projected through a random matrix into 8192 dimensions, passed through ReLU, and accumulated into a Gram matrix G and an unnormalized class-prototype matrix K; the classifier weights are then computed as Wo = (G + λI)^−1 K, with λ chosen per task on a validation split. This single-pass update yields 93.4% final average accuracy with 2.5% forgetting on the 50-class ESC-50 class-incremental setup, and 61.4% accuracy with 2.0% forgetting on the 9-city TAU Urban Acoustic Scenes 2019 domain-incremental setup. The","pith_inferences":["The method's success depends on how well PANNs CNN14 embeddings separate sound classes that were not part of its training distribution; a natural test is to swap in other pretrained audio encoders and see whether accuracy tracks the quality of their embeddings.","The authors chose Q = 8192 and the per-task λ search using validation splits on the same ESC-50 and TAU tasks; a stricter evaluation would fix these hyperparameters on a held-out sequence of new tasks to confirm the numbers are not inflated by benchmark-specific tuning.","Because the readout is linear in random ReLU features, the method is essentially a kernel-style regressor; explicit kernel features such as random Fourier features might offer a principled way to tune the expansion dimension.","The framework could plausibly be combined with a small memory buffer for settings where class distributions shift more abruptly, though the paper does not test that combination."],"forward_implications":["One unified online framework handles both class-incremental and domain-incremental audio learning without task IDs, replay buffers, or base-task training.","Adapting to a new task requires only a single forward pass through its training samples, so deployment on streaming or resource-constrained devices becomes practical.","The reported accuracy exceeds even a joint linear probe that trains on all tasks together, suggesting that the nonlinear random expansion plus decorrelation recovers useful structure from frozen features.","Forgetting stays around 2 percentage points in both setups, indicating that stability-plasticity balance is achieved without explicit memory of past data.","Trainable parameters are limited to the Q×C readout plus the accumulated Gram matrix, far fewer than the 80.8M parameters of the frozen CNN14 backbone."],"supporting_citations":[{"why":"Supplies the frozen PANNs CNN14 pretrained audio model and the 2048-dimensional embeddings used for all experiments.","marker":"[12]"},{"why":"Provides the RanPAC method of random projections plus Gram and prototype matrices with ridge-regression readout that the paper adapts to audio.","marker":"[24]"},{"why":"Supplies the ESC-50 dataset of 50 environmental sound classes used for the class-incremental learning evaluation.","marker":"[25]"},{"why":"Provides the 24/8/8 train/validation/test split protocol used for ESC-50 in the experiments.","marker":"[26]"},{"why":"Supplies the TAU Urban Acoustic Scenes 2019 development dataset of city-specific acoustic scenes used for the domain-incremental learning evaluation.","marker":"[27]"}],"fun_headline_variants":["One-pass audio learner forgets almost nothing","Pretrained audio model learns new classes in a single pass","Audio AI gains new skills without forgetting old ones","Single-pass update for audio classification: 93.4% accuracy","Frozen audio encoder adapts in one forward pass"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole method bets that PANNs CNN14's frozen 2048-dimensional embeddings remain informative for any sound class and domain that arrives later, and that the projection size and ridge penalty chosen on the same benchmark tasks are not silently inflating the reported accuracy.","fun_headline_variants_meta":{"raw":{"variants":["One-pass audio learner forgets almost nothing","Pretrained audio model learns new classes in a single pass","Audio AI gains new skills without forgetting old ones","Single-pass update for audio classification: 93.4% accuracy","Frozen audio encoder adapts in one forward pass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000201,"raw_usage":{"total_tokens":1226,"prompt_tokens":766,"completion_tokens":460,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":383}},"tokens_in":510,"tokens_out":460,"duration_ms":4361,"temperature":1.0,"reasoning_tokens":383,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:51:33.869707+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical online protocol on a held-out sequence of sound classes or acoustic domains that are semantically far from AudioSet, such as medical or marine sounds, keeping Q = 8192 and the per-task λ search unchanged. If single-pass accuracy drops to the level of a plain linear probe, the claim that the random expansion and decorrelation are what prevent forgetting would be refuted.","supporting_citations":[{"cited_title":"PANNs: Large-scale pretrained audio neural networks for audio pattern recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen PANNs CNN14 pretrained audio model and the 2048-dimensional embeddings used for all experiments."},{"cited_title":"Ranpac: Random projections and pre-trained models for continual learning,","cited_arxiv_id":null,"evidence_quote":"Provides the RanPAC method of random projections plus Gram and prototype matrices with ridge-regression readout that the paper adapts to audio."},{"cited_title":"ESC: Dataset for Environmental Sound Classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the ESC-50 dataset of 50 environmental sound classes used for the class-incremental learning evaluation."},{"cited_title":"Environmental sound classification with convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the 24/8/8 train/validation/test split protocol used for ESC-50 in the experiments."},{"cited_title":"A multi-device dataset for urban acoustic scene classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the TAU Urban Acoustic Scenes 2019 development dataset of city-specific acoustic scenes used for the domain-incremental learning evaluation."}],"review_version":1}