Last year Google open-sourced Fairness Indicators for sliced evaluation of TensorFlow Models, now Google announced TensorFlow MinDiff which is towards addressing the unfair bias in the Machine Learning models. The TensorFlow MinDiff will help the Data Scientist in fairness evaluation and fixing the Machine Learning model to work well.
Google announced the release of TensorFlow MiniDiff on its official blogpost saying “Today, we are announcing the release of MinDiff, a new regularization technique available in the TF Model Remediation library for effectively and efficiently mitigating unfair biases when training ML models.”
The TensorFlow MinDiff is a technique for addressing unfair bias issue with the machine learning model (ML Models). This library can be used to run the test against different data sets and then penalizing the models based on the differences in the predictions.
During model training phase Data Scientists always try to minimize the by optimizing the data distribution. But once the model is trained it becomes difficult to find bias in large unseen data. At this point TensorFlow MinDiff will help in evaluating the model. This tool is another step towards making TensorFlow a good library for Model Remediation, which will be a suitable technique for many different types of use cases.
The newly released TensorFlow MinDiff library will help Data Scientists in fixing the bias in the Machine Learning model and to find out the working of models for various users. This library can be used to check the performance of the model on a different slice of data and to evaluate the unfair bias of the Machine Learning model. This library will help Data Scientists in experimenting with the different data sets to avoid unfair bias in the model.
TensorFlow MiniDiff is model remediation techniques which are developed to balance the error rate across the different slices of data. This API works by penalizing the distributional differences between different data samples. Thus helping Data Scientists in developing better Machine Learning Models.
The TensorFlow MiniDiff library can be used to minimize either false-negative or false-positive between the slice data of different classes or groups.
MiniDiff can be used to find out the unfair bias in the ML Models and help the Data Scientists in developing models for solving a wide variety of societal challenges in a much better way. If there are two given datasets for evaluating the model, then this technique penalizes the models based on the dissimilarity in the distributing scores between the data sets. Then the penalty applied is in the direct proportion to the differences between the two datasets. The penalty is applied based on prediction scores between the datasets.
Finally the penalty is applied to the model by adding components to the loss and in this way model tries to minimize the penalty by bringing the distribution together.
The good thing here is that the performance of the model does not deteriorate much with the use of MinDiff API and its performance remains in the acceptable range. So, this is a very good API for mitigating the unfair bias issues in Machine Learning Models.TensorFlow MinDif
10,522 total views, 12 views today
