Google Announces "What-If" Tool for Analyzing Machine Learning Models

Posted on by Jamie McGibbon in Cloud Computing, and Google Cloud Platform

Machine learning is something that can be incredibly useful for a variety of applications across many different fields. When it comes to conceptualizing, implementing, and training these models, though, it can often be quite challenging. This is especially true when it comes to determining how different data and “what-if” scenarios will affect a model.

However, in a recent post on the Google AI blog, Google announced a new “What-If” tool that can be used to observe how a machine learning model works when adding or working with new data or datasets. This tool, which is a feature of the open-source TensorBoard web application, can take some of the stress out of analyzing, debugging, and fine-tuning machine learning models.

Traditionally, training machine learning models and analyzing different “what-if” scenarios required programmers who would write code to see how a model would be affected. And while programmers are still an important part of developing and working with such models, Google’s new tool makes it easy for those without a programming background to participate in the training and analysis of machine learning models through the use of a graphical user interface.

With the new “What-If” tool, programmers and non-programmers alike can explore and work with TensorFlow models to see what results may be produced after applying changes. Users can easily add, edit, and remove data points and can view the results visually, with the tool offering several different ways to display the results. This makes it easy to see, at a glance, how changing various factors or inputs might affect a model.

What’s more is that the “What-If” tool allows users to find and view counterfactual examples from within the data, with users also being able to arrange examples by similarity. There are also various performance indicators that are available to view, including thresholds, ROC curves, confusion matrices, and cost ratios.

The tool also allows datasets to be broken up into subgroups, which can be useful for testing changes to algorithms and what not on a smaller subset of data.

For those that are interested in trying the new “What-If” tool, there are several working demos available on the tool’s Github page, where users can test the tool on binary and multi-class classification, and regression models.

Machine learning is a field that can be greatly helpful in a variety of fields including technology, medicine, and finance, among others. And while building and training models can take a lot of time and effort, Google’s new “What-If” tool can help to make testing new ideas or scenarios a bit less of a lift.

BECOME A PETRI MEMBER:

Don't have a login but want to join the conversation? Sign up for a Petri Account

Register