你可能会需要的术语介绍
众所周知,一个领域的黑话对新人来说是比较不友好的,为此我从知乎上找了一篇黑话大赏(bushi)做了点改良放在这里。如果遇到看不懂的词了可以来这找找。在系统学习之前可以先无视这篇文章,遇到问题再来找找。
作者:Young
链接:https://www.zhihu.com/question/469612040/answer/2008770105
来源:知乎
- feature:一个向量
- representation:还是一个向量
- embedding:把输入映射成向量,有时作为名词 = feature
- 提高泛化性:在各种东西上预测更准了
- 过拟合:训练过头了
- attention:加权提取特征,越重要的 feature 权重越高
- adaptive:还是加权
- few-shot learning:看了几个样本就学会了
- zero-shot learning:一个没看就开始用自带的知识瞎蒙
- self-supervised:自学(自监督)
- semi-supervised:教一点自学一点(半监督)
- unsupervised:没人教了,跟谁学?(无监督)
- end-to-end:一套操作,行云流水搞到底(输入是图像,输出也是图像就算)
- multi-stage:发现不行,还得一步一步来
- domain:我圈起来一堆样本,就管他叫一个domain
- transfer:我非得在这一堆样本上训练,用在另一堆样本上,就是不直接训练,就是玩 ~
- adversarial:我加了一部分就是让 loss 增大
- robust:很稳我不会让 loss 变大的(但也不容易变小了)
- state of the art(sota):我(吹 nb)第一
- outperform:我虽然没第一,但是我比 baseline 强
- baseline:(故意) 选出来的方法,让我能够看起来比它强
- empirically:我做实验了,但不知道为啥 work
- theoretically:我以为我知道为啥 work,但没做实验,或者只做了个 toy model
- multi 开头词组
- multi-task:把几个不同任务的 loss 加一起,完事
- multi-domain:把几堆儿样本混一块训练,完事
- multi-modality:把视频语音文字图像 graph 点云 xxx 混一块训练,完事
- multi-domain multi-modal multi-media model:mua~mua~mua~mua……
- 消融实验:删掉某模块做对比实验
- 长尾数据:出现频率低的类别很多
- ... Is all you need:骗你的,就是把你骗进来。除了... 你还要一堆 trick
- 体素:我把世界变成 MC 了,世界是一堆方块,他们在不同视角下有各自的颜色和透明度
- 点云:我每采样一次得到一个点,由这些点去表示我要的物体,不太直观,来张图
这是我用照片重建的独角兽稀疏点云,红色的不用管,是照相机视角(图不够多,巨糊)
先这些,后续想起来了可能会补充。