Uber Open Sources Manifold: A Visual Tool for Debugging AI Models

2020-01-08

In order to reduce workload, Uber has developed Manifold internally, which is a model-agnostic visual tool that displays the differences in feature distributions (i.e., measurable attributes of observed phenomena). It has helped various product teams analyze countless AI models. As of today, it has been open-sourced on GitHub.


Uber machine learning software engineer Li Lezhi wrote in a blog post, "Since the creation of Manifold earlier this year, we have received a lot of feedback on its potential in general ML model debugging scenarios. When open-sourcing the standalone version of Manifold, we believe that the tool will benefit the ML community by providing interpretability and debuggability to ML workflows."


Manifold utilizes a clustering algorithm (k-Means) to divide prediction data into multiple segments based on their performance similarity. The algorithm ranks features based on their KL divergence, which is a measure of the difference between two contrasting distributions. Generally, higher divergence in the manifold indicates the factor that distinguishes two groups of segments for a given feature.


Manifold includes support for various algorithm types, including conventional binary classification and regression models. In terms of visualization, it can extract numerical and categorical as well as geospatial feature types. It integrates with Jupyter Notebook, one of the most widely used data science platforms for data scientists and ML engineers, and offers interactive data slicing and performance comparison based on per-instance prediction loss and other feature values.


Manifold's convenient performance comparison view allows for comparing the predictive performance of models and data subsets. The feature attribution view summarizes the feature distribution of data subsets with various performance levels defined by the user. Both provide an overview of model performance and map views of geospatial features, aiding in identifying underperforming data subsets for further investigation.


Manifold is packaged as a standalone installation or a bundled installation. After installation, there are two ways to input data into it: via comma-separated files or programmatically.


The features included in version 1 release are:



  • Model-agnostic support for general binary classification and regression model debugging. Users will be able to analyze and compare models of various algorithm types, enabling them to differentiate performance differences among various data slices.

  • Visualization support for tabular feature inputs, including numerical, categorical, and geospatial feature types. With feature value distribution information for each data slice, users can gain better insights into potential causes of certain performance issues, such as whether there is any correlation between the model's prediction loss and the geographic location and distribution of its data points.


The new upgrade of integrated blocks includes visualization support for geospatial features.



  • Integration with Jupyter Notebook. Through this integration, Manifold accepts data input as Pandas DataFrame objects and presents the visualization of this data in Jupyter. As Jupyter Notebook is one of the most widely used data science platforms for data scientists and ML engineers, this integration allows users to analyze their models without disrupting their normal workflow.


Manifold's integration with Jupyter Notebook accepts data input as Pandas DataFrame objects and presents the visualization in the Jupyter Notebook UI.



  • Interactive data slicing and performance comparison based on per-instance prediction loss and other feature values. Users will be able to slice and query data based on prediction loss, ground truth, or other features of interest. This feature will enable users to quickly validate or reject their hypotheses through a common data slicing logic.


Interactive data slicing based on per-instance prediction loss and feature values allows Manifold users to better understand ML model performance issues.

RECENT AI NEWS

RECENT AI TOOLS