◎ Tech Daily Reporter Zhang Mengran
Can machines learn, but can they forget?
A friend said, “If you unplug the power without saving, won’t it forget!”
The issue is not that simple.
The purpose of machine learning is to use computers as tools to simulate human learning in a real-time and authentic manner. It can categorize existing content into knowledge structures and is widely applied to solve complex problems in engineering and science. Today, machine learning is considered one of the most intelligent research fields, but scientists have raised a new question: Can machines learn, but can they forget?
Screenshot from Wired magazine
Recently, an article on the Wired website stated that as an emerging field in computer science, machine learning researchers have begun exploring methods to induce “selective amnesia” in AI, aiming to delete sensitive data related to specific individuals or points from machine learning without affecting model performance. If this can be achieved in the future, the concept will help people better control data.
Video screenshot. Researchers are testing whether they can delete sensitive data without needing to retrain the system from “scratch”.
In reality, while the way machines learn mimics humans, their method of “forgetting” is quite different from ours.
The “forgetting” in machine learning is quite intuitive for users who have needs—such as those who regret sharing content online. However, from a technical perspective,the traditional method to eliminate the influence of specific data points is to “rebuild the system from scratch,” which can be a prohibitively expensive task for companies.
Specifically, users in certain regions have the right to request companies to delete all their data if they change their minds about the content they disclosed. However, achieving “complete erasure” is difficult because once trained, machine learning systems do not easily change, and even the trainers themselves may not fully understand how the system acquires these capabilities—they cannot completely comprehend the algorithms they debug or train.
Image Source: Vision China
In 2019, some scientists suggested that machine learning projects could segment source data into multiple parts to achieve “forgetting” of individual data points, but this idea has recently been proven flawed—if deletion requests are submitted in a specific order, whether accidentally or maliciously, the machine learning system could crash.
Concerns about AI systems infringing on personal privacy are growing. How to enable AI to learn “selective amnesia” so that sensitive data can be deleted without requiring “from scratch” retraining of the system has become a hot research topic. It relates to whether data can be better controlled and the value derived from it. To realize the concept of “selective amnesia,” scientists may need to make new explorations in computer science.
“When they (users) request data deletion, can we eliminate all impacts of their data while avoiding the full costs of retraining from scratch?” said Aaron Roth, a machine learning professor at the University of Pennsylvania. His current research aims to find some “middle ground.” Perhaps in the near future, a path will be found that can both control data and protect the value derived from it.
Source: Tech Daily
Editor: Zhang Shuang
Reviewer: Yue Jiang
Final Review: Liu Haiying
