Google wants its AI systems to be able to forget things they’ve learned

Unlearning technology is really hard, so Google has launched a “challenge”.

Artificial intelligences have reached a moment of maximum development that a few years ago was unthinkable. This has brought us great benefits as a society, but it has also generated a series of inevitable problems and serious discussions about certain professions that are going to disappear as a result of its gradual replacement by the new machines that sooner or later seem to occupy all the spaces of our lives. In this frame, Google has a new target for its artificial intelligence: that they learn to unlearn, which they have dubbed machine-unlearning and have launched a challenge on this issue so that users learn how to do it.

It is true that until now this type of issue has always been a problem since artificial intelligences are having a lot of trouble unlearning things that they do not need or that they’ve just been outclassed.

Machine-Unlearning, Google’s commitment to learning to unlearn

Machine-Unlearning basically consists of the ability of machines to forget information that they no longer need. According to Google itself:

Deep learning has led us to tremendous progress in many applications, from generating realistic images to language models that can have conversations like humans. While this progress is very excitingWidespread use of neural network models requires caution: as Google’s AI principles explain, we also seek to develop responsible AI technologies by understanding and mitigating potential riskssuch as propagation and amplification of biases and protecting user privacy.

This is how Google itself explained how machine-unlearning works | Image: Google

Thus, it seems that the main importance of machine-unlearning is the privacy protection. Thus, information that is inaccurate or out of date can be deleted without problems. The thing does not stop there, but You can also remove dangerous or tampered content and, therefore, it can harm the people of whom it is spoken.

The problem is in the unlearn algorithm. It is a process in which some pre-trained models are necessary for it. Ideally, these models should be no different than those that have not been trained to forget, but at the moment there is the problem to make it work. Although it may be surprising, it is a very difficult model This has caused serious problems for researchers. In fact, there is not much bibliography about it since it is a really complicated issue and it is quite scattered.

It is for this reason that Google has launched a challenge so that AI experts can test their knowledge and we can all benefit from it. The challenge will begin in the middle of this month of July and the test consists of teaching the AI ​​to forget a series of faces and the ages that we find in them.