w***g 发帖数: 5958 | 1 对raw audio/image提取特征算是一个killer app。而且非常依赖convolution。一旦上
了fully connected layer,参数数量立刻爆炸性增长,数据量不大的话直接就overfit
了。不知道还有哪些数据上NN能超越SVM。有经验的同学请过来说说。 | l*******m 发帖数: 1096 | 2 nlp还行,nn over sparse data 好像没人怎么搞过
overfit
【在 w***g 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 对raw audio/image提取特征算是一个killer app。而且非常依赖convolution。一旦上 : 了fully connected layer,参数数量立刻爆炸性增长,数据量不大的话直接就overfit : 了。不知道还有哪些数据上NN能超越SVM。有经验的同学请过来说说。
| b*****o 发帖数: 715 | 3 Deep learning is suitable for any content-based input, such as text, image,
video, or audio.
And as you said, the tricky thing is to choose the topology of the network,
which varies from one data type to another.
On the other hand, for structured data, deep learning does not have any
benefit over standard out-of-shelf methods, such as SVM or gradient boosting.
overfit
【在 w***g 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 对raw audio/image提取特征算是一个killer app。而且非常依赖convolution。一旦上 : 了fully connected layer,参数数量立刻爆炸性增长,数据量不大的话直接就overfit : 了。不知道还有哪些数据上NN能超越SVM。有经验的同学请过来说说。
|
|