Demystifying Machine Learning

We all have heard about machine learning and its path-breaking applications. So, what is machine unlearning and why the need for it? Well, let’s first start with the basic understanding of what it means. Machine learning is a type of artificial intelligence that enables machines to analyse, learn, and adapt to the surroundings, which was earlier done on an entirely different data set. Machine unlearning is just the opposite of it. The device can unlearn the learned stuff whenever needed.

Machine Unlearning: The Need For It

Vast chunks of big data explode into organisations and this data helps machine learning algorithms learn to act in new environments. However, we sometimes want our systems to unlearn these digital assets for several reasons. Firstly, continuous retention of data leads to data privacy issues. The recent news on Facebook says it all - the change in privacy policy to send user’s information across the Web have led to distrust among users and they ended up deleting their accounts. Similarly, the iCloud Data Breach incident had left users unhappy, which led them to search online teaching articles that could help them remove their data from iCloud. Secondly, along with privacy, meeting the security constraint with continuous machine learning is also a big challenge. For instance, an anomaly detection system, which is typically meant for detecting intrusions, fails to perform its job with machine learning techniques. Consider a situation where an attacker injects some data into the parent data to pollute a system. What will happen if the system will adopt this new data as its own and perform future tasks? Hence, here the security fails. In this case, we would apparently want the system to forget the polluted data after detection and get back the security. Thirdly, you might have come across a situation where Google sends you news or updates that are not important to you. This happens when you or your friend have searched for that particular topic and the system learns that you are interested in that specific issue. Here we want the system to unlearn and only send relevant news or recommendations to you. These are only a few reasons that have generated the need for machine unlearning techniques.

Machine Unlearning: Wipes Out Unwanted Data Quickly

Researchers Yinzhi Cao and Junfeng Yang proposed a system that can forget data efficiently. The model is based on two criteria, completeness and timeliness. Completeness meant the difference between the training data and the unlearned data, whereas timeliness has got to do with how soon the system had wiped the unnecessary data. Researchers tested and ran this model with different machine learning systems successfully. After this success, researchers aim to adapt the machine unlearning techniques to other systems and ultimately delete the unwanted data.

No doubt, machine learning techniques have led to incredible and groundbreaking innovations but they have also led to few issues that must be resolved at the earliest. How to fix them? The answer is machine unlearning.