Visualizing The Data Transformation of a Neural Network

If you struggle to comprehend how a neural network learns complex non-linear data, I have created an animation that will surely help.

Please find the video here: Neural Network Animation.

For linearly inseparable data, the task boils down to projecting the data to a space where it becomes linearly separable.

Now, either you could do this manually by adding relevant features that will transform your data to a linear separable form. Consider concentric circles for instance. Passing a square of (x,y) coordinates as a feature will do this job.

But in most cases, the transformation is unknown or complex to figure out. Thus, non-linear activation functions are considered the best bet, and a neural network is allowed to figure out this "non-linear to linear transformation" on its own.

As shown in the animation, if we tweak the neural network by adding a 2D layer right before the output, and visualize this transformation, we see that the neural network has learned to linearly separate the data. We add a layer 2D because it is easy to visualize.

This linearly separable data can be easily classified by the last layer. To put it another way, the last layer is analogous to a logistic regression model which is given a linear separable input.

The code for this visualization experiment is available here: GitHub.

Find the code for my tips here: GitHub.

I like to explore, experiment and write about data science concepts and tools. You can read my articles on Medium. Also, you can connect with me on LinkedIn.