The paper introduces a label‑free validation method called geometric consistency for selecting checkpoints in audio active learning. It uses pretrained embeddings to generate pseudo‑labels for unlabeled data, scoring checkpoints by agreement with these pseudo‑labels and cross‑entropy on the labeled set. The method works with any query strategy, requires only one additional forward pass per round, and achieves top performance in area under the labeling curve without using extra labeled data.
arXiv:2607. 13571v1 Announce Type: cross Abstract: Sound event detection relies on frame-level strong labels whose annotation is expensive.
arXiv:2609.39934v1 Announce Type: cross Abstract: Checkpoint selection in domain generalization often relies on source-validation accuracy, yet the selected checkpoint need not provide reliable proba...
The paper investigates how pseudo‑labeled examples should be incorporated into mini‑batches during continual active learning (CAL) to balance annotation savings and training cost. It formalizes the problem as two decisions: selecting which unlabeled examples receive pseudo‑labels and determining their share of each mini‑batch. The proposed semi‑supervised continual active learning (SSCAL) method, compatible with any acquisition function, uses distance‑based pseudo‑labeling, caps their batch share, and reweights labeled losses to maintain gradient balance, achieving top performance across multiple audio datasets and settings.
arXiv:2607. 01297v1 Announce Type: cross Abstract: Most existing audio classification methods suppose that each query (testing) sample belongs to a class of support (training) samples, and misrecognize samples of unseen classes as seen classes (cannot reject samples of unseen classes).
arXiv:2609.36569v1 Announce Type: cross Abstract: Checkpoint selection is a routine decision in supervised fine-tuning (SFT): training produces multiple checkpoints, but only one is retained. Yet fix...
The paper introduces Cumulative Sample Loss (CSL), a metric derived from checkpoint loss dynamics, to detect annotation errors in temporally labeled videos. CSL captures persistent disagreement between a frame’s annotation and the learned visual‑temporal structure, flagging high‑CSL frames as likely mislabeling or temporal disordering. Experiments on EgoPER and Cholec80 show CSL outperforming final‑checkpoint loss, achieving up to a 4.2‑point AUC improvement on EgoPER and 92.0/78.5 AUC on Cholec80.
arXiv:2307.12226v3 Announce Type: replace-cross Abstract: Machine learning models -- including prominent zero-shot models -- are often trained on datasets whose labels are only a small proportion of...
arXiv:2607. 06179v1 Announce Type: cross Abstract: There are some datasets of varying scales for audio classification (AC) applied to different tasks.
arXiv:2608. 14824v1 Announce Type: cross Abstract: We present a parameter-free episodic evaluation of nearest-centroid classification for elephant vocalisations on fixed pretrained acoustic embeddings, across the Elephant Voices (EV) and Linguistic Data Consortium (LDC) datasets.
arXiv:2607. 13555v1 Announce Type: cross Abstract: Bioacoustic call-type classification relies on costly expert annotation.
arXiv:2609.37169v1 Announce Type: cross Abstract: Mid-training equips pretrained large language models with specialized and reasoning capabilities, but the returns of this stage are bounded since add...
The paper investigates why direct pairwise AUC surrogate training, though aligned with the ranking metric, can fail to provide corrective updates during stochastic training. It introduces the concept of Corrective Power, showing that in certain regions severe ranking errors persist without generating strong gradient signals, a phenomenon linked to degeneracies in the U‑statistics of pairwise gradients. The authors explain empirical observations—such as the effectiveness of squared‑hinge loss, the benefit of cross‑entropy warm‑up, and the advantage of parameter‑efficient fine‑tuning—through this theoretical lens, and validate their claims with experiments on eight computer‑vision and eight NLP tasks using various models.