Training

The ML process

For any complex problem that requires the computer to be able to identify patterns, there is an ML process to solve it.

Machine learning

Let’s recap the machine learning solution as we have understood it so far. Given the numerical dataset, we take it to the input space and partition the space through a decision boundary such that the “Good” class lies on one side of the decision boundary and the “Bad” class lies on the other side. This way, a general concept of good movies is learned.

Model

We have our first machine learning model, a perceptron, which can map the input space into a decision space and linearly separate the “Good” class from the “Bad.”

However, where do the right weight values come from? The movie dataset only has two features ($acting$,$direction$), so we were lucky to plot the entire dataset and see the complete 2D input space. We cannot do that for most of the real world problems that machine learning engineers would have to solve. For example, the galaxy data has many features as compared to the movie data. Each galaxy image may have thousands of pixels, and each pixel value is a feature. Thus, even a small 100 x 100 image needs to be represented in a 10,000-dimensional input space, which we can’t visualize by plotting in the 2D input space. We have no way of visualizing beyond 3D, thus far!

Learning to the rescue

Mathematically speaking, a function is defined as a mapping between input ...