Scientists explore new problems of AI "selective amnesia"

    "Wired" magazine website, as an emerging field in computer science, machine learning researchers have begun to explore methods to induce "selective amnesia" in AI, with the goal of not affecting model performance. , Remove sensitive data of specific people or points from machine learning. If it can be realized in the future, this concept will help people better control data.

  The purpose of machine learning is to use computers as tools and to be committed to real and real-time simulation of human learning methods. It can divide existing content into knowledge structure, and then be widely used to solve complex problems in engineering applications and scientific fields. Nowadays, machine learning is regarded as the research field with the most intelligent features, but scientists have asked a new question: the machine can learn, but will it forget? In fact, although their learning methods are imitating human beings, their "forgetting" methods are quite different from ours.

       The "forgetting" of machine learning is actually very intuitive for users in need, that is, those who regret the content they share online. But from a technical point of view, the traditional method to eliminate the impact of specific data points is to rebuild the system "from scratch". This is a potentially very costly task that is almost unbearable for enterprises. Specifically, users in certain regions have the right to request the company to delete their data if they change their minds about what is disclosed. But it is difficult to completely erase this thing, because once trained, the machine learning system will not change easily. Even the trainers themselves do not know how the system masters these abilities, because they cannot fully understand them. Algorithms debugged or trained by yourself.

       In 2019, some scientists proposed that the source data of machine learning projects could be divided into multiple parts to achieve the "forgetting" of a single data point, but it has recently been proven to be flawed. If the submitted deletion requests appear in a specific order, whether accidental or malicious, the machine learning system will crash. Therefore, to realize the concept of "selective amnesia", scientists may need to make new explorations in computer science.

       "When they (users) ask to delete data, can we eliminate all the impact of their data while avoiding the full cost of retraining from scratch?" Aaron Ross, a professor of machine learning at the University of Pennsylvania, said that their current research is hope Some "middle grounds" can be found. Perhaps in the near future, it is expected to find a development path that can not only control data but also protect the value generated by data.

       Editor-in-chief

       In fact, it is not just machines facing the problem of "selective amnesia", humans have also not mastered this skill. Forgetting, which often happens inadvertently, is passive. People can't choose exactly what to remember and what to forget, otherwise, there are so many "toast to dissipate sorrow and worry more." The neural network training process of the machine learning model is like "alchemy". It's hard for you to know exactly what the pill is made from, so you don't dare to easily change the heat and the elements that enter the furnace. Possibly, it is necessary to understand the path of machine learning very well in order to achieve accurate data extraction. In short, this is indeed a problem that needs to be solved but is very difficult with conventional thinking.

(Note: transferred from Xinhuanet. If there is any infringement, please contact to delete)