Why prevent the loss of convergence in the convolution neural network?

Question

I use to train shadoworflow konvetov with a set of training 15,000 images from 22 classes. I have 2 conv-layer and single layer fully connected. I trained a network with 15,000 images and experienced convergence and high accuracy on a training kit.

However, my test suite has much lower accuracy, so I assume the network is more suitable. To combat this, I added an exception to the full level is connected to my network.

However, adding an exception led to the network not converging after many iterations. I was wondering why this could be. I even used the high probability of dropping out (still possible 0.9) and have experienced the same results.

Answer 1

Well, if you leave the dropout probability as a percentage of 0.9, it means that there is a 10% chance that the neuron connection will come out at each iteration. Thus, for dropout must also be an optimum value.

As in the above, you can understand that we are also scalable, our neurons on a roll. The above case is 0.5. If this o.9, it will again be another scaling.

So, basically, if he has fallen 0.9, we have to scale it by 0.9. This means that we have more than 0.1 in the test.

This is what you can understand how an attack can affect. Thus, for some it is likely to saturate your nodes, etc., which causes the problem of convergence.

Answer 2

You can add a drop to your dense layers after the convolutional layers and remove the drop from the convolutional layers. If you want to have many more examples, you can put some white noise (5% random pixels) on each image and have a P, P 'option for each image. It can improve your results.