自定义博客皮肤VIP专享

*博客头图：

点击选择上传的图片

格式为PNG、JPG，宽度*高度大于1920*100像素，不超过2MB，主视觉建议放在右侧，请参照线上博客头图

请上传大于1920*100像素的图片！

博客底图：

点击选择上传的图片

图片格式为PNG、JPG，不超过1MB，可上下左右平铺至整个背景

栏目图：

点击选择上传的图片

图片格式为PNG、JPG，图片宽度*高度为300*38像素，不超过0.5MB

主标题颜色：

RGB颜色，例如：#AFAFAF

Hover：

RGB颜色，例如：#AFAFAF

副标题颜色：

RGB颜色，例如：#AFAFAF

预览取消提交

自定义博客皮肤

-+

上一步保存

ID246783的博客

博客等级

码龄4年

162
原创

225
点赞

353
收藏

200
粉丝

关注

私信

热门文章

分类专栏

最新评论

机器学习--神经网络
2401_84079994: I guess the math part is used in certain functions in the whole process flow. I did not try to memorize all the related functions. At least not doing so before I can catch the supporting logic and/or math prove behind. In fact I am trying to run the simple network on a very simple demo example provided in a book, line by line, for about 2 months on and off. Kind of frustrated by the line by line explanations on the book. Only understand what it is doing but lack why it is doing so. Well, it is college experience. Professor can not tell everything and student can not learn everything.
机器学习--神经网络
aWty_: In a particular training, those parameters are randomly given in the initialization part, and the SGD algorithm would help you adjust those parameters automatically to reach a lower loss, which means reaching higher accuracy for prediction. So the initialization for parameters is not a part of network design, but a part of training process. That is my understanding for your question.
机器学习--神经网络
aWty_: In a particular training, those parameters are randomly given in the initialization part, and the SGD algorithm would help you adjust those parameters automatically to reach a lower loss, which means reaching higher accuracy for prediction. So the initialization for parameters is a part of network design, but a part of training process. That is my understanding for your question.
机器学习--神经网络
2401_84079994: thanks a lot for reply. any input fills my curious mind. for deep network, my thought is that it may run much less times in loop while simple network may run much more time in loop, given a timeframe. but, since the designs are different with unknown effect on back and forth process between any two layers, the deeper network may still get better result than simpler network. by the way let me ask a dummy question about the math part. is it true that, given a designed network, you do have a fixed set of parameters (or variables, like x1,x2, etc) and therefore those partial derivatives comes out in play? if true, those pre-fixed parameters are part of the neural network design for a problem? if it is still true, I would guess that the deep learning process may have certain capacity to increase/decrease parameters and change parameter values at the same time in order to generate better result. to this point, it suddenly tastes like partial random evolution process. LOL. thanks again. If I have more energy I may get into the math part a little bit and ask for more understanding.
机器学习--神经网络
aWty_: Firstly, sorry for being unable to answer your questions, I'm also a green hand learner in the machine learning field. However, I would still like to share my viewpoint on your questions. 1. The network cannot guarantee 100% accuracy for predictions, yet you can increase it by adjusting parameters while training your network with SGD. 2. If you are curious about how to design a network, I recommend you learn something about LeNet AlexNet VGG GoogLeNet ResNet NiN and so on. These are both the nets that are proven to be efficient in solving problems like computer vision. 3. As for the third question, the answer is absolutely no. As we all know, the more middle layers a net has, the more time it requires to train. So in a fixed timeframe, an over-complex net may even unable to finish training, let alone provide a better prediction. Additionally, if a network has too many layers, it may encounter a problem called overfitting. So the answer is NO.

sortest path

关注

a

关注数：文章数：4 文章阅读量：471 文章收藏量：0

作者: aWty_

在 AFO 的边缘徘徊的 OIer

展开

专栏收录文章