Github Copilot is a new GitHub service providing an AI-based code assistant. Functionally, it works as an autocomplete plugin to some editors (as of the time this post is being written, VSCode, JetBrains IDEs and NeoVim are supported). It doesn’t feel like any usual code completion tool though: while traditional tools are able to suggest simple completions such as variable names and method arguments (based on the source code semantics), Copilot generates suggestions that vary from a few characters to entire files.

In my last post I described a method for computing feature importance that can be applied to Deep Neural Networks trained to solve NLP tasks. Today I will show how the magnitude and direction information recovered by this method can be used to guide editing movie reviews in the desired sentiment direction (positive or negative).

In this post I will describe a technique I have been applying at Kunumi to explain the output of Deep Neural Networks trained to solve Natural Language Processing (NLP) problems.