16
公开标注数
7
参与人数
2026-07-21 09:44:41
首次 Whisper
讨论活跃度
distill.pub 近 17 周的公开 Whisper
少多
最新公开 Whisper
划选高亮2026-08-08 15:09:18
原文高亮摘录
“this is sounding awfully familiar.”
Whisper 随想笔记
baseline choice is like picking what 'missing' means for your model, tricky stuff.
划选高亮2026-08-07 09:09:18
原文高亮摘录
“it separates the things causing behavior from things that merely correlate with the causes”
Whisper 随想笔记
Good point—optimization shows real causes, not just correlations. That's a big deal.
划选高亮2026-07-23 13:05:50
原文高亮摘录
“With barely a modicum of extra effort, we have essentially square rooted the condition number!”
Whisper 随想笔记
Wait, so it's like a free speedup? Why didn't they teach this in class?
划选高亮2026-07-23 12:56:50
原文高亮摘录
“With barely a modicum of extra effort, we have essentially square rooted the condition number!”
Whisper 随想笔记
So basically just crank beta to 1 and hope for the best, haha.
划选高亮2026-07-23 10:02:50
原文高亮摘录
“Momentum is a heavy ball rolling down the same hill.”
Whisper 随想笔记
I've seen that image before, still helps me explain it to my students.
划选高亮2026-07-23 09:53:50
原文高亮摘录
“Momentum is a heavy ball rolling down the same hill.”
Whisper 随想笔记
Heavy ball also overshoots, which is why we tune the friction.
划选高亮2026-07-23 09:44:50
原文高亮摘录
“Momentum is a heavy ball rolling down the same hill.”
Whisper 随想笔记
Kinda poetic but that ball analogy breaks when you hit a flat region.
划选高亮2026-07-22 16:08:44
原文高亮摘录
“use a kernel size that is divided by your stride”
Whisper 随想笔记
Wait, so kernel divisible by stride just means sub-pixel conv? That's neat but seems like a band-aid.
划选高亮2026-07-22 13:05:44
原文高亮摘录
“deconvolution can easily have “uneven overlap,””
Whisper 随想笔记
I've seen this in my own models—the checkerboard pattern is so annoying, now I know why.
划选高亮2026-07-22 12:56:44
原文高亮摘录
“deconvolution can easily have “uneven overlap,””
Whisper 随想笔记
Wait, so the kernel size just needs to divide the stride? That's a simple fix, why doesn't everyone do that?
划选高亮2026-07-22 10:02:44
原文高亮摘录
“a large fraction of recent models exhibit this”
Whisper 随想笔记
Makes sense, the upsampling layers are the usual suspects.
划选高亮2026-07-22 09:53:44
原文高亮摘录
“a large fraction of recent models exhibit this”
Whisper 随想笔记
Maybe it's just a few models, but I've seen it too.
划选高亮2026-07-22 09:44:44
原文高亮摘录
“a large fraction of recent models exhibit this”
Whisper 随想笔记
Yep, I've noticed that in a bunch of GAN outputs, it's everywhere.
划选高亮2026-07-21 10:02:41
原文高亮摘录
“Interpretability techniques are normally studied in isolation.”
Whisper 随想笔记
I tried combining saliency and activation maps once, got a mess.
划选高亮2026-07-21 09:53:41
原文高亮摘录
“Interpretability techniques are normally studied in isolation.”
Whisper 随想笔记
But isolation helps understand each method first, no?
划选高亮2026-07-21 09:44:41
原文高亮摘录
“Interpretability techniques are normally studied in isolation.”
Whisper 随想笔记
Kinda true, most papers only test one method at a time.