Insights into what is happening behind neural network

Course Q&A Deep Learning Specialization Improving Deep Neural Networks: Hyperparameter tun
1

Good day,

in this course we have seen some lab activities that illustrate how each layer separates the dataset. For planar data points I can clearly see that the relu activations draw half spaces to separate the data in different categories. My questions is: How can we gain insights in the case of image recognition?

In the lecture Andrew cited the article to showcase what each layer is doing. How are those images in the article generated? Specifically the images are in section 4.1 of the article I linked.

I am playing with the rice image data set on Kaggle. I currently achieve scores around 91- 94 percent, which I think is reasonable as a beginner. Nevertheless, I would love to see similar images in the cited article in my case. How would I go about producing those images?

The model I attempted is here.

chi-yu

2

Hello chi-yu @u5470152,

Let me quote from that paper’s section 4.1:

The first layer consisted of 24 groups (or “bases”) of 10x10 pixel filters, while the second layer consisted of 100 bases, each one 10x10 as well.

Given that Figure 2 showed exactly 24 and 100 images for the first and the second layer, I guess they were just plotting each of the filters as if it is a 10 x 10 grey scale image.

Raymond