In this post, we reverse the direction of traditional machine learning. Usually the direction is from data to a property: One builds a model that predicts for a data point a certain property. We are going to explore the other direction: from property to data point. Our high-level goal is to start with a desired property, and to modify a data point until that property is satisfied.

In this post, we reverse the direction of traditional machine learning. Usually the direction is from data to a property: One builds a model that predicts for a data point a certain property. We are going to explore the other direction: from property to data point. Our high-level goal is to start with a desired property, and to modify a data point until that property is satisfied.
Consider the toy dataset MNIST, where the usual goal is to build a classifier that takes as input an image of a handwritten digit and predicts the digit that is being depicted. This prediction can be understood as a likelihood vector that assigns a certain probability to each possible digit. In the figure below, we see that the classifier assigns the highest probability to the digit 7

Read full article on Medium.com