you could inspect a machine learning model, with no coding required?
Building effective machine learning systems means asking a lot of questions. It's not enough to train a model and walk away. Instead, good practitioners act as detectives, probing to understand their model better.
But answering these kinds of questions isn't easy. Probing "what if" scenarios often means writing custom, one-off code to analyze a specific model. Not only is this process inefficient, it makes it hard for non-programmers to participate in the process of shaping and improving machine learning models. For us, making it easier for a broad set of people to examine, evaluate, and debug machine learning systems is a key concern.
That's why we built the What-If Tool. Built into the open-source TensorBoard web application - a standard part of the TensorFlow platform - the tool allows users to analyze an machine learning model without the need for writing any further code. Given pointers to a TensorFlow model and a dataset, the What-If Tool offers an interactive visual interface for exploring model results.
What can you do with the What-If Tool?
Visualize inference results
Examples are colored by inference results until changed. Organize results into confusion matrices and other custom layouts to show inference results faceted by numerous features.
Edit a datapoint and see how your model performs
Edit, add or remove features or feature values for any selected datapoint and then run inference to test model performance. Alternatively, you can duplicate or upload a whole new example to see where it stands vis-à-vis loaded examples.
Explore the effects of a single feature
Explore auto-generated plots for individual features of a selected datapoint that show the change in inference results for different valid values of the feature.
Explore counterfactual examples
Seek out the most similar example of a different classification for any datapoint for classification models along L1 and L2 distances calculated using the distribution of feature values across all loaded examples. WIT highlights the delta between the initial example and its counterfactualWachter, S., Mittelstadt, B., & Russell, C. (2017). Counterfactual explanations without opening the black box: Automated decisions and the GDPR..
Arrange examples by similarity
Create distance features from a selected datapoint using either L1 or L2 distances and apply it to your visualizations for further analysis.
View confusion matrices and ROC curves
For binary classification models and examples that include a feature describing the true label, explore model performance interactively using thresholds, ROC curves, numeric confusion matrices and cost ratios.
Test algorithmic fairness constraints
For binary classification models, slice your dataset into subgroups and explore the effect of different algorithmic fairness constraints with the push of a button, such as "equality of opportunity"Hardt, M., Price, E., & Srebro, N. (2016). Equality of opportunity in supervised learning. In Advances in neural information processing systems (pp. 3315-3323)., for your model on those subgroups.
This binary classification model predicts whether a person earns more than $50k a year based on their census informationFisher, R. A. UCI Machine Learning Repository. Irvine, CA: University of California, School of Information and Computer Science.. Discover how different features affect the model's predictions.
This regression model predicts a person's age from their census informationFisher, R. A. UCI Machine Learning Repository. Irvine, CA: University of California, School of Information and Computer Science.. Slice data and explore inference results, such as aggregated inference error measures across different subgroups.
This binary classification model predicts whether an image contains a smiling faceLiu, Z, Luo, P, Wang, X, Tang, X. Deep Learning Face Attributes in the Wild. Proceedings of International Conference on Computer Vision (ICCV) 2015.. Can you figure out what group was missing from the training data, resulting in a biased model?
This multi-class classification model predicts flower type from plant measurementsLichman, M. UCI Machine Learning Repository. Irvine, CA: University of California, School of Information and Computer Science.. Look at the correlations between different features and flower type.
The People + AI Research initiative (PAIR) brings together researchers across Google to study and redesign the ways people interact with AI systems. We focus on the "human side" of AI: the relationship between users and technology, the new applications it enables, and how to make it broadly inclusive. Our goal isn't just to publish research; we're also releasing open source tools for researchers and other experts to use.
The What-If Tool was a collaborative effort. We would like to thank the Google teams that piloted the tool and provided valuable feedback and the TensorBoard team for all their help.
 Wachter, S., Mittelstadt, B., & Russell, C. (2017). Counterfactual explanations without opening the black box: Automated decisions and the GDPR.https://arxiv.org/abs/1711.00399
 Hardt, M., Price, E., & Srebro, N. (2016). Equality of opportunity in supervised learning. In Advances in neural information processing systems (pp. 3315-3323). https://arxiv.org/abs/1610.02413