{"id":3269,"date":"2016-05-23T17:44:40","date_gmt":"2016-05-23T15:44:40","guid":{"rendered":"https:\/\/geekosas.com\/?p=3269"},"modified":"2026-05-23T17:45:06","modified_gmt":"2026-05-23T15:45:06","slug":"playing-with-neural-networks","status":"publish","type":"post","link":"https:\/\/geekosas.com\/index.php\/2016\/05\/23\/playing-with-neural-networks\/","title":{"rendered":"Playing with Neural Networks"},"content":{"rendered":"

A few days ago I wrote about neural networks<\/a>, trying to explain in a simple way how they work and why this type of technology is called "artificial intelligence."<\/p>\n

The important thing about this technique and what makes it so revolutionary is the concept of training, but it is often difficult to imagine how this happens and how the parameters affect the result.<\/p>\n

To help us out, our friends at TensorFlow<\/a> created a graphical tool that shows us the training process of a neural network, available at http:\/\/playground.tensorflow.org\/<\/a><\/a><\/p>\n

For those who don’t really know how to interpret the application, I’ll explain the basic components. To do that, I’ll use the following screenshot as an example:
\n <\/p>\n


\n\"\"<\/p>\n

For the impatient who already want to see the app: If you press the "play" button, it will start training the neural network to predict the colors of the dots on the right.<\/p>\n

For the more geeky ones, the explanation begins here (I was impatient too). The application is divided into 5 sections: the top menu (in white), DATA, INPUT, HIDDEN LAYERS, and OUTPUT. The orange colors represent negative values, while the blue ones represent positive values. A white color means something close to 0.<\/p>\n

DATA<\/strong> contains parameters about the sample to train the neural network. First, you can choose the general shape of the data, then the percentage of data left to test the model (see OUTPUT section), a noise factor to make the work more "real" for the neural network, and finally the number of cases that will enter each training cycle.<\/p>\n

INPUT<\/strong> is the second section, and it is for choosing the variables that will enter the model. x1 and x2 are predefined, representing the vertical and horizontal axes of the OUTPUT section, but you can also choose transformations on these, such as x1*x2, sin(x1), or sin(x2). The lines coming out of each box correspond to the value of the constant associated with the parameter input in the activation or response function.<\/p>\n

HIDDEN LAYERS<\/strong> is for configuring the connections the neural network will have. In the boxes, you can see the shape the function takes as x1 and x2 move. The lines connecting them represent the value of the constant associated with the data coming from the previous function\/neuron. In other words, if it is thick, it means it has high impact, and if it is blue, it has a positive constant, while if it is orange, it is negative.<\/p>\n

OUTPUT<\/strong> shows the contrast between the value obtained by the neural network and the data it should fit (read the same way as the boxes in HIDDEN LAYER). When you start training, above the large graph, you will see a smaller graph that shows how the error percentage of the neural network evolves for both the control group and the training group. Below the large graph, you can choose to discretize (orange or blue) the network output and also show the control group. The control group is used to measure the model’s predictiveness; the network is never trained on these.<\/p>\n

Top Menu<\/strong> is where the magic happens. To start training, you must press the play button and everything will begin. In "Interactions," you see how many times it has trained with the entire sample of dots. Then come the parameters:<\/p>\n