We introduce the Information and Community Introspection toolkit DNIKit, an open supply Python framework for analyzing machine studying fashions and datasets. DNIKit accommodates a group of algorithms that every one function on intermediate community responses, offering a singular understanding of how the community perceives knowledge all through the totally different levels of computation.
With DNIKit, you’ll be able to:
create a complete dataset evaluation reportfind dataset samples which are close to duplicates of every otherdiscover uncommon knowledge samples, annotation errors, or mannequin biasescompress networks by eradicating extremely correlated neurons detect inactive models in a mannequin
To visualise sure analyses, DNIKit additionally works with Symphony, a analysis platform for creating interactive knowledge science elements we initially revealed at ACM CHI 2022. Now open-sourced, Symphony elements allow a number of stakeholders in cross-functional AI/ML groups to discover, visualize, and share analyses for AI/ML. Symphony helps quite a lot of knowledge varieties and fashions, and can be utilized throughout platforms reminiscent of Jupyter Notebooks to standalone web-based dashboards. Symphony additionally has particular elements to visualise the outcomes from DNIKit analyses, reminiscent of computing dataset familiarity and duplicates.
We use Symphony along with DNIKit for interactive, visible dataset evaluation – most notably, the Dataset Report.