Abstract
Machine Learning methods are a much-discussed topic today in storage industry. Everyone wants to have a data driven insightful decision-making capabilities into their products. Moreover, when dealing with risk-sensitive systems – where the cost of a bad decision can be very high, and prediction accuracy is not the only objective; a multidimensional perspective about the quality of prediction needs to be considered. The reality is that reliable estimation of prediction confidence remains a significant challenge in machine learning. In this talk, we’ll discuss the issues with conventional machine learning, and introduce Conformal Prediction which can be used with any underlying learning method by creating a nonconformity measure for the algorithm used, allowing predictions complemented with valid confidence measures, i.e. confidence of each prediction and credibility indicated by informativeness of data. We discuss how to build a conformal prediction framework, it's validity and implementation challenges, and list some applications in storage.