Research map / Label-efficient and robust learning / Self-supervised and contrastive
Unmute the Patch Tokens: Rethinking Probing in Multi-Label Audio Classification
Lukas Rauch, René Heinrich, Houtan Ghaffari, Lukas Miklautz, Ilyass Moummad, Bernhard Sick, Christoph Scholz
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the linear cls-token probe with binarized prototypical pooling | EAT frozen probe · AudioSet-20k | mAP: 17.29 -> 31.67 | Improved |
| replace the linear cls-token probe with binarized prototypical pooling | A-MAE frozen probe · AudioSet-20k | mAP: 8.36 -> 22.32 | Improved |
| use protobin instead of linear across general audio datasets | all six backbones · general multi-label audio | mAP: +14.41 | Improved |
| use protobin instead of linear on few-shot bioacoustics | all six backbones · birdset few-shot | mAP: +12.16 | Improved |
| use prototypical pooling instead of the best attentive pooling (mhca) | frozen backbones · general audio | mAP: -4.59 | Improved |
| replace mhca with protobin on polyphonic AudioSet-20k | BEATs · AudioSet-20k | mAP: 21.86 -> 31.54 | Improved |
| replace mhca with protobin on single-label ESC-50 | EAT · ESC-50 (accuracy) | mAP: 89.8 -> 86.8 | Hurt |
| replace mhca with protobin on single-label SC-2 | EAT · SC-2 (accuracy) | mAP: 93.2 -> 90.4 | Hurt |
| use protobin with ASiT backbone vs float class-dependent proto | ASiT · AudioSet-20k | mAP: 21.89 -> 20.96 | Hurt |
| add supervised AudioSet fine-tuning of the backbone before linear probing | BEATs linear probe · AudioSet-20k | mAP: 24.71 -> 40.30 | Improved |
| add supervised fine-tuning to the backbone before protobin probing | SSLAM protobin · AudioSet-20k | mAP: 30.94 -> 43.70 | Improved |
| use protobin on top of linear probe so as to close the gap to fine-tuning | EAT · AudioSet-20k | mAP: closes 63% of the gap | Improved |
Closest papers by meaning
- ALMTokenizer: A Low-bitrate and Semantic-rich Audio Codec Tokenizer for Audio Language Modeling (ICML 2025 · similarity 0.86)
- WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling (ICLR 2025 · similarity 0.86)
- Weakly-supervised Audio Separation via Bi-modal Semantic Similarity (ICLR 2024 · similarity 0.85)
- FLAM: Frame-Wise Language-Audio Modeling (ICML 2025 · similarity 0.85)
- Synthio: Augmenting Small-Scale Audio Classification Datasets with Synthetic Data (ICLR 2025 · similarity 0.85)
- Generative Audio Language Modeling with Continuous-valued Tokens and Masked Next-Token Prediction (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
