Counterfactual Logit Pairing

Posted by Bhaktipriya Radharapu, Software Engineer

is an open source toolkit that showcases solutions to help mitigate unfair bias in Machine Learning models. The toolkit offers resources to build fairer models for everyone – in line with (CLP) to address unintended bias in ML models.

ML models are prone to making incorrect predictions when a sensitive attribute in an input is removed or replaced, leading to unintended bias. For instance, the .

Counterfactual Logit Pairing (CLP) is a technique that addresses such issues to ensure that a model’s prediction doesn’t change when a sensitive attribute referenced in an example is either removed or replaced. It improves a model’s robustness to such perturbations, and can positively influence a model’s stability, fairness, and safety.

CLP mitigates such counterfactual fairness issues at training time. It does so by adding an additional loss to the model’s training loss, which penalizes the difference in the model’s outputs between training examples and their counterfactuals.

Another advantage of using CLP is that you can use this even on unlabelled data. As long as the model treats the counterfactual examples similarly you can validate that your model is adhering to counterfactual fairness.

For an in-depth discussion on this topic, see research on , and codelab provides an end-to-end example. In this overview, we'll emphasize key points from the notebook, while providing additional context.

The notebook trains a text classifier to identify toxic content. This type of model attempts to identify content that is rude, disrespectful or otherwise likely to make someone leave a discussion, and assigns the content a toxicity score. For this task, our baseline model will be a simple Keras sequential model pre-trained on the of the classifier on original and counterfactual examples.

Build a counterfactual dataset using CounterfactualPackedInputs by performing a naive ablation based on term matching.

Improve performance on flip rate and flip count by training with CLP.

Evaluate the new model’s performance on flip rate and flip count.

Be aware that this is a minimal workflow to demonstrate usage of the CLP technique, and not a complete approach to fairness in machine learning. CLP addresses one specific challenge that may impact fairness in machine learning. See the are you trying to achieve?

Consider when counterfactual pairs should have the same prediction. Many syntactic counterfactuals generated by token substitution may not require identical output. Consider the application space and the potential societal impact of your model and understand . You can try experimenting with other metrics in the suite to know which options offer the best results.

counterfactual_weight = 1.0 counterfactual_model = counterfactual.keras.CounterfactualModel( baseline_model, loss=counterfactual.losses.PairwiseMSELoss(), loss_weight=counterfactual_weight) # Compile the model normally after wrapping the original model. # Note that this means we use the baseline's model's loss here. optimizer = tf.keras.optimizers.Adam(learning_rate=0.001) loss = tf.keras.losses.BinaryCrossentropy() counterfactual_model.compile(optimizer=optimizer, loss=loss, metrics=['accuracy']) counterfactual_model.fit(counterfactual_packed_input, epochs=1)

Once again, we evaluate the results by looking at the flip count and flip rate. Select “flip_rate/overall” within Fairness Indicators and compare the results for female and male between the two models. You should notice that the flip rate for overall, female, and male have all decreased by about 90%, which leaves the final flip rate for female at approximately 1.3% and male at approximately 1.4%.

You can get started with Counterfactual by visiting .

Acknowledgements

The Counterfactual framework was developed in collaboration with

• Amy Wang, Ben Packer, Bhaktipriya Radharapu, Christina Greer, Nick Blumm, Parker Barnes, Piyush Kumar, Sean O’Keefe, Shivam Jindal, Shivani Poddar, Summer Misherghi, Thomas Greenspan.

This research effort was jointly led by

• Alex Beutel, Jilin Chen, Tulsee Doshi in collaboration with Sahaj Garg, Vincent Perot, Nicole Limtiaco, Ankur Taly, Ed H. Chi.

Further, this work was pursued in collaboration with

• Andrew Smart, Francois Chollet, Molly FitzMorris, Tomer Kaftan, Mark Daoust, Daniel 'Wolff' Dobson, Soo Sung.

Vollständiger Original-Bericht

Ausführliche Details, Code-Beispiele & Hersteller-Stellungnahme auf blog.tensorflow.org.

↗ Original-Artikel auf blog.tensorflow.org lesen