A brief analysis of the influence of parameters on the performance of deep learning model in zhihu Answer

Question: Can parameter tuning greatly improve the performance of the deep learning model? The answer of

The details are as follows:

ZFNet can learn about the structure between Alex Net and VGG Net, the Image Net champion in 2013. Modified some parameters compared to Alex Net. In fact, VGG is also calculated by Alex

 

 

As can be seen from the above chart, the simpler the network parameters, the greater the improvement. This is basically caused by two reasons: 1, simple network can optimize more space, 2, simple network performance is general, relatively easy to improve.

In fact, ResNet V2, also calculated by V1 tuning optimization, the improvement is not so obvious.

 

In fact, individual ResNet V2 contributes more. Although the lifting effect is general, a general lifting residual block method is proposed, which is more widely used.