BUPT / clubber.ml

A.M.C. - Artificial Intelligence & Machine Learning CLUB: Friends for Code, Paper, and Beers!🍻
https://ai-ml.club
Apache License 2.0
120 stars 37 forks source link

第三季6场 AI ML Club 活动纪要 - AI ML Club #345

Open utterances-bot opened 4 years ago

utterances-bot commented 4 years ago

第三季6场 AI ML Club 活动纪要 - AI ML Club

« 本次沙龙通知: #325 «

https://ai-ml.club/events/seminar-meeting-minutes-3-6/

huan commented 4 years ago

补充一个心得

Euclidean 距离和 Cosine 距离从数学角度来讲基本等价,除了 cosine 距离中,限制了向量的模值必须等于1。“模值等于1”这个约束,对训练有力,所以实际应用中,arc face 使用了 cosine 距离,比 facenet 产生了一定优势。(感谢范弘炜的分享)

AMC Logo Update

AMC Logo 中,已经将 "AI ML Club" 字样,改为了我们的域名:"AI-ML.Club"

huan commented 4 years ago

主席 Manual 需要修订

hwfan commented 4 years ago

补充一个心得

Euclidean 距离和 Cosine 距离从数学角度来讲基本等价,除了 cosine 距离中,限制了向量的模值必须等于1。“模值等于1”这个约束,对训练有力,所以实际应用中,arc face 使用了 cosine 距离,比 facenet 产生了一定优势。(感谢范宏炜的分享)

AMC Logo Update

AMC Logo 中,已经将 "AI ML Club" 字样,改为了我们的域名:"AI-ML.Club"

弘 not 宏 :-) (为防之后的人也写错,还是决定回复更正一下了哈哈)

另补充:欧氏度量和余弦度量的区别在于模值,而这个区别的影响主要是来自训练中的一个经验规律,即在人脸数据训练中,特征的模值往往代表样本的质量。这是因为最后得到的特征经过了BN模块的归一化处理,使得更一般化的特征会往中间集聚,更特殊化的特征会往周边弥散(笔者在MegaFace训练数据库上,也得到了能够佐证中这个规律的实验结果)。因此需要进行模值归一化,让网络对于不同质量的样本作等同的重视。

huan commented 4 years ago
- 范宏炜
+ 范弘炜

收到,fixed!

huan commented 4 years ago

@hwfan Please join our BUPT organization!

You've invited hwfan ! They'll be receiving an email shortly. They can also visit https://github.com/BUPT to accept the invitation.

huan commented 4 years ago

Please join our BUPT organization!

You've invited @AlaskaaWong ! They'll be receiving an email shortly. They can also visit https://github.com/BUPT to accept the invitation.

newip commented 4 years ago

Registered for the S3E7 event. Refer to @huan BTW: Why can't I reply from ai-ml.club directly?

huan commented 4 years ago

@newip agree to refer to me.