Reproduce “SpecAugment++: A Hidden Space Data Augmentation Method for Acoustic Scene Classification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Helin Wang, Yuexian Zou, Wenwu Wang
SpecAugment++, a data augmentation method for Acoustic Scene Classification, extends masking techniques to both input spectrograms and intermediate hidden states of deep neural networks. The approach, particularly with its mini-batch based mixture masking scheme, achieved accuracy gains of up to 4.7% on standard DCASE datasets, outperforming established augmentation strategies.
Key insights
- Augmenting intermediate hidden representations, in addition to input data, significantly improves model generalization and robustness in Acoustic Scene Classification.
- Masking schemes that introduce 'interference' or content from other samples (MM and CM) are more effective than simple zero-value masking, compelling the network to learn more discriminative features.
- Applying augmentation across multiple layers, from input spectrograms to deeper hidden states, yields superior performance, indicating a synergistic effect from diverse feature perturbations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.