I’m sure you’ve heard/read about machine learning on many occasions, but what about machine unlearning? The truth is like this this concept is fundamental to the future of artificial intelligence, more specifically generative models, taking into account the problems that have been appearing in them for some time. Problems that, as it becomes more and more evident every day, seriously threaten its reliability and thus its future possibilities.

We’ve talked about AI bugs in the past (and most likely will continue to do so in the future). For example, at the end of last year (when ChatGPT started to become popular), we defined three types of errors: ignorance, faulty sources and unreal content, which we can directly relate to hallucinations. This kind of behavior of generative models has already led to problems like this.

Currently, the main method of correcting these problems is to retrain the model., either by adding some data, removing others, and of course in each case reviewing the algorithm to find and catch possible improvements as well as potential bugs. However, this process is extremely expensive (it is estimated that currently training a model with GPT-3 is about four million dollars, and that due to its increasing complexity, this cost will rise to $500 million in 2030) and requires resources of time, personnel and data processing capacity. Therefore, the need to constantly retrain is extremely inefficient.

That is, in the face of this problem, when the draft machine learning first introduced in 2007 by researchers Alexander Strehl and Joydeep Ghosh. Researchers have proposed using machine learning to improve the accuracy of machine learning classifiers. In their work, they showed that machine learning could reduce the bias of classifiers and improve their accuracy, even when the training data was limited.

So we are talking about a relatively new technique, and that’s why still has a lot of room for development, which is something that is increasing due to the emergence of generative models that we have experienced in recent years. Several techniques are currently used, such as deleting training data known to be erroneous or biased, or adjusting their weights in the model to reduce their role in the generated responses.

Recently, Google has launched the first edition of its machine learning challenge, a competition that started in mid-July and will run until mid-September. Its goal is to verify the effectiveness of different learning methods for the same task, as well as to establish metrics that will help identify their reliability in a standardized way.