[1904.12848] Unsupervised Data Augmentation for Consistency Training
6
公开标注数
6
参与人数
2026-07-28 09:43:22
首次 Whisper
本页的公开 Whisper
划选高亮2026-07-28 16:07:22
原文高亮摘录
“arXivLabs: experimental projects with community collaborators”
Whisper 随想笔记
Is this just for the staff or can anyone pitch in?
划选高亮2026-07-28 13:04:22
原文高亮摘录
“use of consistency training on a large amount of unlabe”
Whisper 随想笔记
So basically, they found that how you add noise matters more than the amount of unlabeled data?
划选高亮2026-07-28 12:55:22
原文高亮摘录
“use of consistency training on a large amount of unlabe”
Whisper 随想笔记
Consistency training with better noise is the real winner here, not just the semi-supervised part.
划选高亮2026-07-28 10:01:22
原文高亮摘录
“Semi-supervised learning lately has shown much promise in improving deep learning models when labeled data is scarce”
Whisper 随想笔记
We tried RandAugment in our pipeline and yeah, it made a bigger dent than I expected.
划选高亮2026-07-28 09:52:22
原文高亮摘录
“Semi-supervised learning lately has shown much promise in improving deep learning models when labeled data is scarce”
Whisper 随想笔记
Still skeptical about those IMDb numbers with just 20 labels, but the idea is solid.
划选高亮2026-07-28 09:43:22
原文高亮摘录
“Semi-supervised learning lately has shown much promise in improving deep learning models when labeled data is scarce”
Whisper 随想笔记
Finally someone pointing out that the noise itself matters, not just the consistency trick.
分享本页 Whisper
短链接
https://domwhisper.com/s/88fe27608f89嵌入代码
<iframe src="https://domwhisper.com/embed/88fe27608f89" width="100%" height="480" style="border:0;border-radius:16px" loading="lazy"></iframe>