Google has announced the first Machine Unlearning Challenge in collaboration with various academic and industrial researchers. The challenge is aimed at advancing the state of machine unlearning, a subfield of machine learning that works to remove the influence of specific training data subsets, known as the “forget set”, from a trained model. This is particularly crucial in cases where the data requested to be deleted from the model can still influence its function and potentially pose privacy risks.
The challenge invites participants to develop unlearning algorithms that can effectively forget a subset of training images while maintaining model utility. The competition will be hosted on Kaggle between mid-July and mid-September 2023, and will involve a scenario where a certain subset of training images must be “forgotten” to protect individual privacy.
Machine unlearning has wider applications, including erasing outdated or inaccurate information from trained models or removing harmful data. This subfield of machine learning also relates to differential privacy, life-long learning, and fairness in model training.
The challenge aims to address the inconsistency of evaluation metrics and the lack of a standardized protocol, which are currently impeding progress in the field of machine unlearning. The competition hopes to provide direct comparisons between different unlearning methods, fostering novel solutions and shedding light on open challenges and opportunities.
If you are interested to read the full article, you can find it here.
