Machine unlearning: programmed Alzheimer's disease?

April 11, 2016 | 15:00
On the crossroads of memory
On the crossroads of memory
Internet is great for finding stuff. This is possible because the users (meaning you and me) stock enormous amounts of data online that is exploited by clever search algorithms capable of identifying relationships. Post a piece of information, a photo, a video, an article, your credit card details, anything on the internet and within minutes it has been absorbed by all kinds of information collecting systems to be remembered forever. So what to do if you want to be forgotten?

Removing a piece of information is not enough to make "The System" forget because the data has not only been stored, it has also been consumed. In order to be completely forgotten, the data itself must of course be deleted but also its so-called lineage. Doing this requires advanced machine unlearning algorithms. Researchers Yinzhi Cao and Junfeng Yang of Columbia University propose a new kind of unlearning algorithm based on a technique that makes learning systems only stock incremental data instead of the individual pieces. Forgetting something would be much simpler and faster because it only requires recalculation of incremental sums instead of doing the whole calculation all over again.

We look forward to a future where finding information will again be difficult; a future where we no longer know that once we knew; a future where we can reinvent a past that we do not want to remember. (Do I feel a headache coming up?)
Loading comments...
related items