AI is used worldwide, which has profound implications for businesses. The general focus is on how well AI learns things and what can be done to help it learn better, but AI unlearning can be just as important.
AI models are often used to create new ones. This is efficient and saves resources, but the data transferred may be deleted from the source and become part of the latest model’s derivative. Unlearning an AI model before training a new one is essential to secure data for many businesses.
What Is AI Unlearning
In artificial intelligence, unlearning occurs when an AI model “forgets” or removes a particular data element from its knowledge base. If the unlearning is unsuccessful, the AI model will not leverage this data in any future predictions or results that it generates.
The Human Element
AI models are programmed and trained by humans. This system is how they learn how to craft curated answers based on its user’s preferences. Human unlearning is different than forgetting. Humans can ignore some aspects of demand, and what they forget is largely beyond their control. People do have a vague capacity to remove consideration of specific information when making decisions, but this can be cloudy. What humans learn affects their behavior in ways most may not be conscious of.
Unlearning Is Hard for AI
Like the humans who train it, AI finds unlearning difficult. This makes intuitive sense, as its learning capacity is strongly connected with who trains it. In an AI model, patterns from previously studied data are entered into model coefficients in ways that are not always obvious.
Extracting the impact of a particular data point on millions of numbers is a challenge that has yet to be solved. As AI models acquire more information, it is possible for them to “forget” or lose the impact of their data, but this process is not as specific as people may want.
The obvious way to train an AI model to unlearn is by retraining the model from scratch. However, this approach can be impractical given the high cost of AI training, especially in large batches. Another hurdle is AI’s ability to leverage one AI to create another.
Using technologies such as Transfer Learning, an AI model that has already been trained can be used as the base to make a second model. This helps speed up model creation and lower model costs. Still, it has the added complication that whatever the initial AI model has not forgotten will now be the new model’s derivative.
Why AI Unlearning Is Important for Business
Business applications often need to remove data, unlike humans. Legal requirements sometimes require the secure deletion of data records after a specific time frame. Computer technologies, from databases to storage devices, have mechanisms in place for secure deletion designed to meet regulatory and customer criteria.
AI presents a challenge because the information that was once part of AI training persists in newer AIs after secure deletion has been performed on the data. AI unlearning is critical for 100% information deletion, especially when handling sensitive data.
