What is machine ‘unlearning’ and why is Google pursuing it?

Yet the announcement very much goes hand in hand with AI and addresses some of the concerns people have with it, mainly privacy. You’ve heard of machine learning, but what about machine unlearning?

In late June, Google announced a machine "unlearning" competition. The contest, which is open to anyone, begun in mid-July and will run to mid-September, and is geared toward removing sensitive data from AI systems.

Machine learning is used to creating new content, predicting outcomes, or answering complicated queries based on what it's trained on. These machinations are what an AI system then operates off of. The problem is, once that data has been used to train the system, it can remain in the arteries of the system for years later, or in some cases, indefinitely.

Google is therefore planning to introduce selective ‘amnesia’ to its AI algorithms to remove all traces of a particular data set from its machine learning systems, without affecting their performance. So this means the data fed to the system to help it learn and reach an optimal level of performance is then wiped once its purpose of training the system is finished. Although perhaps not all the information can be removed, those deemed superfluous can be.

Google believes training algorithms to forget the data they've already been trained on would give people more control over sensitive information. This would make it easier, for instance, for the company to serve users who apply for the right to be forgotten.

Equally, AI systems that retain data can be subject to misuse of data by cybercriminals to bully and blackmail users after having hacked the system and obtained their data. Data poisoning, denial of access to online activities; tricking face recognition, and creating deepfakes, all are more of a possibility with traditional deep learning systems.

As AI’s growth, machine unlearning would make it possible for someone to remove their data from an algorithm and ensure that no one else profits from it, while also protecting themselves from the dangers of AI.

This is increasingly important as new data-handling laws are either being implemented or catching up to meet the growth in AI. In fact, Google’s bid to induce machine unlearning is partly responding to regulation such as the EU’s GDPR guidelines, where people can demand data deletion from a business if they have concerns about the information they revealed or gave to it.

A growing sentiment to regulate AI is gaining momentum across the world due to fears over privacy, misreporting or even intellectual property infringement. Dr Geoffrey Hinton, dubbed the ‘Godfather’ of AI quit his role at Google with warnings against AI and AI chatbots. Equally, even the creator of chatbot ChatGPT Sam Altman called for AI regulation as he testified in front of US congress for "development and release of AI models above a threshold of capabilities".