Can a Machine Learning System Unlearn?

I was inspired by this article as well as my experience before starting my MBA. Several years ago, I encountered the theme of “learn, unlearn, relearn” and made an effort to incorporate it into my work. Now, as a graduate student, I try and incorporate it into my education as I encounter new ideas that require me to unlearn old ways of thinking. 

The first part “learn” is straightforward – we absorb information or learn processes. The third part “relearn” is also straightforward – sometimes we need to relearn a skill or knowledge. Take riding a bicycle after many years or rediscovering a language; it may take some time, but you can relearn and even strengthen the skill or knowledge. 

The last part “unlearn” is complicated. Unlearning can refer to unlearning a fact, a way of thinking, or a way of acting. A far too simple example is we learned that Pluto was a planet growing up, but now it is not classified as a planet. We had to unlearn that fact then replace it with new knowledge. Or unlearning shortcuts/tools for an old OS system when an update drops. A more complex example is unlearning some work habits when you transitioned to work from home. Unlearning can free up space and allow us to gain new insights to approach a challenge differently. 

With digital disruption all around us, organizations need to unlearn old systems or approaches to make way for new processes or tactics. This HBR article goes into detail about how unlearning can help organizations adapt and grow in a rapidly changing environment. Our brains cannot keep adding information, so we need to unlearn certain processes to make room for new and improved processes. 

What makes unlearning complex is not fully knowing how deep that way of acting or thinking affects other habits or perspectives. As a result, how do we know what to target to unlearn something? Moreover, how can we fully know if we have unlearned a way of thinking or acting?

https://www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/its-time-for-businesses-to-chart-a-course-for-reinforcement-learning

A more interesting application of “learn, unlearn, relearn” comes to light when referring to machine learning systems. An integral part of the machine learning process is learning based on a set of evaluations. As the system examines more and more data and goes through evaluation, the system can improve. And yet, data scientists may not fully comprehend how it works. As algorithms become more complex, it will become more difficult to accurately explain how a machine learning system accomplishes its task. After the system engages in learning and relearning, it may be impossible to identify a specific subset of data and how that directly impacts the system’s algorithm.

This leads me to the first question: Can a machine learning system truly unlearn?

If data scientists cannot understand how these complex machine learning systems function, then how can they know whether a system has truly unlearned something? They would have to clearly define what they want it to unlearn then be able to assess if it has unlearned that. It would be like asking students to unlearn a strategy for solving a math equation, but then not being able to observe them on any future attempts. Take this article that discusses AI and machine learning in healthcare. A machine learning system was looking at images from radiology to help with x-rays and mammograms. In the process, the system was able to accurately identify the race of patients, even though that was not its purpose. The data scientists could not determine which factors the program used to do this. Some data scientists are working on ways to allow machine learning programs to unlearn. Currently, it may require building a new system from scratch.

Which brings me to the second question: Why would you want a machine learning system to unlearn?

First, new regulations and laws may give more power to an individual when it comes to data privacy. If these new regulations permit me to force organizations to stop using specific personal data, they could delete it. But has the machine learning system retained some aspect of my data? Is my data now an integral part of the system? Am I ok with that? 

Second, making machine learning systems unlearn can allow for more adaptable systems that align with ethical and moral standards. As these systems become more complex and further expand into our everyday lives, data scientists may discover inherent biases within them. Instead of starting from scratch, which takes time and money, the ability to unlearn certain patterns could allow the machine learning system to change and still function. 

While the application or necessity of a machine learning system unlearning may not be fully realized, it may help engineers better understand systems and organizations better adapt systems to meet changing regulations and needs. I believe organizations need more tools to refine machine learning systems and the process of unlearning can be a critical tool in the toolbox. 

The complexity and challenges of unlearning for humans also applies to machine learning. Maybe data scientists need to unlearn certain ways of thinking about machine learning to address this challenge?

(I also enjoyed watching this video explaining how machines learn – https://youtu.be/R9OHn5ZF4Uo)

10 comments

  1. I think this is a really thought provoking post. I was reading it and immediately was brought to the current environment where there is divide over facts. I think you addressed the complexity in unlearning, but compound that where what-to-learn is disputed and I think you are left with a total mess. Ultimately though, you are correct– humans probably need to address this first before machines.

  2. This is a fascinating post and something I have never really considered. I think our brains are automatically “programmed” to do certain things, such as unlearning old information and replacing it with a new one. I assumed that machines would be rebuilt from scratch or fully reprogrammed when the old data is obsolete. The concept of machines working like human brains is something that fascinates me and something I want to explore in more detail.

  3. Great read, really cool topic. As I was thinking through your question, I wonder if “unlearning” for machines wouldn’t really quite be “unlearning” information so to speak, but rather learning so much contrary information that the information you want it to forget becomes minuscule and obsolete. AI and predictive models determine outcomes or answers based on many different historical factors that coalesce into a prediction or action (e.g. determining a patient’s race), couldn’t we in theory contribute enough information to the algorithm to weigh the old information less heavily. Ethics and manipulation would become an additional concern here though, just food for thought….

    1. Interesting point @barrinja1! I think the increase in the velocity and quality of data may shift this idea from “unlearning” to the learning point you made.

  4. I’m probably butchering the mechanics of operating and maintaining an AI, but can’t you just re-image the AI’s memory drive to a point in time where re-learning ought to begin? In essence, you are pushing the delete button on the information you no longer want the AI to consider in its processes without functionally altering the AI. Nevertheless, thank you for introducing this topic and I hope to provide some use cases for this in subsequent publications.

    1. @collarcs I think one of the problems with “deleting” versus “unlearning” is that there are times when the algorithm shouldn’t be deleted, but rather needs to untrain certain data that were input into the algorithm. For instance, if a user all of a sudden no longer wants to share their data with a company that is using it to run marketing algorithms, the AI/ML algo would have to unlearn that user’s data. There has been legislature lately that requires organizations to give users the ability to delete their data, it’s called the Right to be Forgotten. I’ve been wondering about this topic for a while, but what happens when someone passes away? Does their data live on? Should it? Could someone put in their will that all of their electronic data be deleted?

  5. I do think it would be non-trivial to have ML models unlearn something, since we don’t really know what it learned to begin with. That said, it is important for models to relearn when conditions change. A great example is the housing crisis, where mortgage risk models didn’t fully account for the conditions on which they were built fundamentally changing, leading them to massively miscalculate the risk profile of certain types of loans.

  6. Great post, Parker! I think it’s a very interesting topic and question. When designing machine learning models, there probably isn’t a lot of thought on the possibility of needing to unlearn something. It would probably be easier to somehow relearn or learn that something from before should be forgotten rather than just forgetting. But now it just seems like a logical problem and I’m already confusing myself.
    Also love the idea of learn, unlearn, relearn. I’ll try to implement it into my own studies as well!

  7. While reading this, I’ve realized I still don’t fully grasp what machine learning is, meanwhile people are already thinking of ways to change some of the underlying understanding of the machine learning platforms that it has developed over the past few years. Just another example of how technology moves so quickly! I imagine another massive reason this would be required is when we identify a new process or find efficiencies in the process that we may want to feed into the machine learning platform. If the platform is inundated with the less efficient processes, then it may not be able to prioritize the newer, more efficient processes.

  8. This blog post was really thought-provoking. I never consciously thought about “learn, unlearn and relearn” until reading this. It’s something that subconsciously happens as I’m introduced to new ideas but perhaps might be more effective if I am more conscious of it! I do think the ML models out there are not as sophisticated as we think

Leave a Reply

Please log in using one of these methods to post your comment:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: