- Sep 2018
-
192.168.199.102:5000 192.168.199.102:5000
-
生成模型 vs. 判别模型
总体来看,如果样本足够多,判别模型的正确率高于生成模型的正确率。
生成模型和判别模型最大的区别在于,生成模型预先假设了很多东西,比如预先假设数据来自高斯,伯努利,符合朴素贝叶斯等等,相当于预先假设了 Hypothesis 函数集,只有在此基础上才有可能求出这个概率分布的参数。
生成模型,进行了大量脑补。脑补听起来并不是一件好事,但是当你的数据量太小的时候,则必须要求你的模型具备一定的脑补能力。
判别模型非常依赖样本,他就是很传统,死板,而生成模型比较有想象力,可以“想象”出不存在于当前样本集中的样本,所以他不那么依赖样本。
关于 想象出不能存在于当前样本集的样本 ,见本课程 40:00 老师举例。
生成模型在如下情形比判别模型好:
- 数据量较小时。
- 数据是noisy,标签存在noisy。
- 先验概率和类别相关的概率可以统计自不同的来源。
释疑第三条优点:老师举例,在语音辨识问题中,语音辨识部分虽然是 DNN --- 一个判别模型,但其整体确实一个生成模型,DNN 只是其中一块而已。为什么会这样呢?因为你还是要去算一个先验概率 --- 某一句话被说出来的概率,而获得这个概率并不需要样本一定是声音,只要去网络上爬很多文字对话,就可以估算出这个概率。只有 类别相关的概率 才需要声音和文字pair,才需要判别模型 --- DNN 出马。
Tags
Annotators
URL
-
- Nov 2017
-
genomebiology.biomedcentral.com genomebiology.biomedcentral.com
-
MCC
Matthews correlation coefficient
-
- Oct 2017
- Jun 2017
-
offconvex.github.io offconvex.github.io
-
One explanation of Non-convex optimization
Tags
Annotators
URL
-
- May 2017
-
www.analyticsvidhya.com www.analyticsvidhya.com
-
Precision: It is a measure of correctness achieved in positive prediction i.e. of observations labeled as positive, how many are actually labeled positive. Precision = TP / (TP + FP) Recall: It is a measure of actual observations which are labeled (predicted) correctly i.e. how many observations of positive class are labeled correctly. It is also known as ‘Sensitivity’. Recall = TP / (TP + FN)
Example: In cancer research you may want higher recall, Since you want all actual positive observations to classified as True Positive. A lower Precision maybe alright because some healthy people classified as cancerous can be rectified later.
-
- Apr 2017
-
demo.clab.cs.cmu.edu demo.clab.cs.cmu.edu
-
if your goal is word representation learning,you should consider both NCE and negative sampling
Wonder if anyone has compared these two approaches
-
- Feb 2017
-
sebastianraschka.com sebastianraschka.com
-
Feature scaling in depth tutorial
-
- Dec 2015
-
cs231n.stanford.edu cs231n.stanford.edu
-
some of the deep learning libraries we may look at later in the class
what c++ libraries are used?
-
- Apr 2015
-
arxiv.org arxiv.org
-
Same as above but the pre- trained vectors are fine-tuned for each task.
How?
Backpropagating to the input layer, changing the vector representation with the training examples?
-
- Sep 2014
-
sites.psu.edu sites.psu.edu
-
you
us
-
you
we
-
your
our
-
your
our
-
your
our
-
ethical challenges.
the ethical challenges you will come up against along the way.
-