What Are Invertible Neural Networks and Why Do They Matter?
While looking at the fancy generative network concept, I wondered whether there were any neural networks that could be inverted. How may such neural networks be beneficial? Throughout my research, I came across an intriguing study. It shows how to design an invertible neural network (INN). INNs can be used for data compression, encoding, and denoising, as I discovered. They are also faster to compute than typical neural networks. In this post, I will define invertibility and describe how to formalize an invertible neural network.
What is an INN?
An invertible neural network is a network that can reconstruct the input data from the output data with high accuracy. Before we get into details, let's first understand invertibility. with an example:
Consider the function f(x), whose domain and range are A and B (domain is the set of possible inputs and range is the set of possible outputs).
Let f be a function from a set A to a set B. The inverse function of f, denoted as f^-1, is a function from B to A such that for every y in the range of f, and for every x in the domain of f, the following two conditions hold:
- If f(x) = y, then f^(-1)(y) = x.
- If f^(-1)(y) = x, then f(x) = y.
In other words, the inverse function of f “reverses” the action of f by mapping every output value back to its corresponding input value.
