AI Machine Learning & Data Science Research

Facebook HiPlot ‘Makes Understanding High Dimensional Data Easy’

Facebook’s new HiPlot is a lightweight interactive visualization tool that takes this further, using parallel plots to discover correlations and patterns in such high-dimensional data.

In this age of global informatisation, the amount of data has exploded — as has the complexity of that data. How to effectively deal with high-dimensional data and find the correlations and patterns hidden within is a serious challenge. In recent years visualization and visual analysis have been applied to the task with some positive results. Facebook’s new HiPlot is a lightweight interactive visualization tool that takes this further, using parallel plots to discover correlations and patterns in such high-dimensional data.

HiPlot outperforms other existing visualization tools in terms of interactivity, simplicity, and extendability. There are a number of reasons for this. Firstly, HiPlot uses interactive parallel plots (a useful way to visualize and filter high-dimensional data) and other graphical methods to more clearly present information. The parallel plots are interactive, and it is easy to change data visualization style at users’ discretion by plotting along one or several axes in various value ranges or using different colours. Secondly, HiPlot can be used directly through an IPython notebook, where the simple syntax enables users to view multiple experiments simultaneously. Thirdly, the incompatibility of different systems’ log formats can make data analysis difficult, but HiPlot is compatible with logs from open-source Facebook AI libraries to help researchers better perform hyperparametric searches. HiPlot’s web server can read the CSV or JSON files by default, and users can also provide their custom Python parsers to convert their experiments into HiPlot experiments.

For example, in the case of a population-based training visualization, since the training task for existing hyperparameter tuning methods might be forked several times with different hyperparameters, it is challenging to analyze such experiments, and they may contain hard-to-find errors. HiPlot however can display the boundaries between related data points, enabling such experiments to be much more easily and accurately visualized.

image.png

Facebook HiPlot can help mitigate issues related to increasing model complexity by effectively analyzing the hyperparameter tuning of deep neural networks. Facebook AI hopes that other researchers will use HiPlot to explore their experiment data more thoroughly and to provide a foundation for more productive future training techniques.

For more information on HiPlot, please check the project page. The HiPlot tool can be downloaded from GitHub.


Author: Herin Zhao | Editor: Michael Sarazen

1 comment on “Facebook HiPlot ‘Makes Understanding High Dimensional Data Easy’

  1. Topic is really excellent !

Leave a Reply to Ziane.ziane Cancel reply

Your email address will not be published.

%d bloggers like this: