Audio-L3A - efficient audio class-incremental learning
Audio CIL · PyTorch · CNN14 / PANNs · scikit-learn · SLURM · 2026.02 - 2026.05
An efficient multi-label audio class-incremental learner combining a frozen backbone, a Ridge closed-form Weighted Analytic Classifier, frequency weighting, and per-class threshold calibration.
Results. Reached 45.16% mAP on five-phase AudioSet-50, outperforming the strongest CIL baseline, LwF, by 2.65 points and finishing only 0.71 points below the offline joint-training upper bound. After phase 0, the model requires no backpropagation; total training time is approximately 1/9 of the offline upper-bound run.
Calibration. Diagnosed why raw Ridge scores produced weak F1@0.5. Per-class Platt scaling raised F1 from 27.51% to 38.93%; sigmoid plus learned per-class thresholds aligned calibration with the F1 objective and reached 45.33% F1 across five random seeds.
Ablations and baselines. Implemented FT, EWC, SI, LwF, and CIL-ML-KD across CNN14 and TCResNet-8. Isolated the effects of frequency weighting, embedding dimension, and threshold calibration; mAP standard deviation across five class orders was only 0.10.
HPC engineering. Reduced a TCResNet-8 baseline from 24 hours to 20 minutes with MFCC feature caching and automated more than 80 experiments across GPU and CPU SLURM partitions.
Paper and code are currently private while the thesis is in progress.
面向多标签音频类增量学习，结合冻结 backbone、Ridge 闭式解加权解析分类器、频率加权和 per-class threshold 校准。
核心结果。 在五阶段 AudioSet-50 上达到 45.16% mAP，比最强 CIL 基线 LwF 高 2.65 个百分点，距离离线联合训练上界仅 0.71 个百分点。Phase 0 后无需反向传播，总训练时间约为离线上界的 1/9。
校准。 定位 Ridge raw score 导致 F1@0.5 偏低的问题。Per-class Platt Scaling 将 F1 从 27.51% 提升至 38.93%；进一步采用 sigmoid + 每类独立学习阈值，使五个随机种子下的 F1 达到 45.33%。
消融与基线。 在 CNN14 与 TCResNet-8 上实现 FT、EWC、SI、LwF、CIL-ML-KD 五种基线，量化频率加权、embedding 维度和阈值校准的贡献；五种类别顺序下 mAP 标准差仅 0.10。
HPC 工程。 使用 MFCC 磁盘缓存将 TCResNet-8 baseline 从 24 小时降到 20 分钟，并通过 SLURM GPU / CPU 多分区自动编排 80 余组实验。
论文与代码在毕业论文完成前暂不公开。