Can AIs Forget?

Check out this answer from Consensus:

The question of whether AIs can forget is complex, involving legal, ethical, and technical dimensions. While the RTBF provides a legal framework for data deletion, its implementation in AI systems is challenging. Technical solutions such as targeted forgetting and intentional forgetting offer promising avenues for managing data in AI, but further research is needed to fully realize these capabilities. As AI continues to evolve, the ability to forget will become increasingly important in ensuring both compliance with legal standards and the efficient functioning of AI systems.

The concept of forgetting is intrinsic to human cognition, allowing individuals to manage and prioritize information effectively. However, in the realm of artificial intelligence (AI), the notion of forgetting is complex and multifaceted. This article explores whether AIs can forget, examining the technical, legal, and ethical dimensions of this capability.

The Right to Be Forgotten

The Right to Be Forgotten (RTBF) is a legal principle that allows individuals to request the deletion of their personal data from various platforms. This principle, enshrined in E.U. privacy law, aims to balance privacy and transparency. However, its application to AI systems presents significant challenges. AI models, which learn from vast datasets, may find it difficult to erase specific data points without affecting overall performance1 5.

Legal and Ethical Considerations

The RTBF raises critical questions about the feasibility and desirability of data deletion in AI contexts. While the law mandates data erasure, the technical implementation in AI systems is fraught with difficulties. The antagonism between privacy and transparency further complicates this issue, as strict adherence to data deletion could undermine the functionality of AI models1 5.

Technical Challenges of Forgetting in AI

Forgetting in AI is not merely about deleting data; it involves complex processes that ensure the AI system does not retain any residual information. This is particularly challenging in machine learning models, where data is deeply embedded in the learned parameters.

Catastrophic Forgetting

One of the primary technical challenges is catastrophic forgetting, where an AI model loses previously learned information upon acquiring new knowledge. This issue is prevalent in neural networks and poses a significant barrier to achieving continual learning4. Techniques such as spike-timing-dependent plasticity (STDP) and neural redundancy have been proposed to mitigate this problem, enabling AI systems to retain old information while learning new tasks4.

Intentional Forgetting

Intentional forgetting is a concept borrowed from human cognition, where irrelevant or outdated information is selectively discarded. In AI, this approach can help manage the ever-increasing data influx and improve system efficiency. Research in this area focuses on developing algorithms that can intentionally forget certain data points without compromising overall performance3 6.

Practical Approaches to AI Forgetting

Several practical approaches have been proposed to address the issue of forgetting in AI systems. These methods aim to balance the need for data retention with the legal and ethical requirements for data deletion.

Targeted Forgetting

Targeted forgetting involves selectively removing specific data points from an AI model. This approach is particularly relevant in sensitive domains such as healthcare, where patient data must be handled with utmost care. Experiments have shown that targeted forgetting can improve model performance while adhering to data deletion requirements2.

Algorithmic Memory

Algorithmic memory refers to the way AI systems process and store information. Unlike human memory, which involves understanding and context, algorithmic memory operates purely on data. This distinction allows for innovative approaches to forgetting, such as reinforcing forgetting by multiplying data points rather than erasing them7.

Can AIs forget?

Scott E. Fahlman has answered Likely

An expert from Carnegie Mellon University in Artificial Intelligence

A big problem for people doing AI via DL networks if that if you train a network on one task or set of data, and then move to a different task, performance on the first task can be degraded or ruined. So some DL researchers are working to combat “catastrophic forgetting”.

Can AIs forget?

David Tuffley has answered Unlikely

An expert from Griffith University in Artificial Intelligence, Software Science

Typically, no, since silicon based neural networks do not suffer the same degradation over time as organic neural tissue. The exception is if the AI is programmed to forget information that corresponds to certain defined criteria, though why this would be useful is not at all clear.

Can AIs forget?

Zdenka Kuncic has answered Likely

An expert from University of Sydney in Artificial Intelligence, Astrophysics

Yes. But more importantly, they can’t remember.

Can AIs forget?

Roman Yampolskiy has answered Likely

An expert from University of Louisville in Computer Science, Artificial Intelligence

Yes

Can AIs forget?

Kate Saenko has answered Likely

An expert from Boston University in Computer Science, Artificial Intelligence

AI is a computer making predictions that mimic human intelligence. For example, an AI program can analyze a recording of a person talking and predict what words they said, similar to how a human can understand speech. Theoretically, an AI program does not need to forget if there is infinite storage available. In practice, something like ‘forgetting’ can happen if an AI program is continuously adjusted on new data but has limited storage, so that the older program becomes overwritten with the new one. In our example, if we keep adjusting AI to new languages, then eventually it might ‘forget’ how to recognize the earlier language, meaning it would have worse recognition accuracy.

Can AIs forget?

Kay Kirkpatrick has answered Likely

An expert from University of Illinois at Urbana-Champaign in Mathematics, Artificial Intelligence

An AI can lose information from its memory or overwrite information, so in this sense an AI can forget.

Can AIs forget?

Mark Lee has answered Likely

An expert from Aberystwyth University in Computer Science

Yes. Although computer hardware chips have perfect memory, it is possible to design any computer program to forget, so, in principle, an AI system can forget too. The usual way a computer system loses information is because old or unused data is either deleted or removed to an archive. For example, banks usually keep customer account records for between five and ten years. But AI systems can learn and this gives many other ways that memories can be lost. Learning involves discovering new data, usually through a process of training. During training many examples of what is to be learned are shown to the system, which then generalises its experience and recognises new, unseen cases. Scientists found that AIs based on artificial neural networks that had previously been trained to recognise images of flowers could be retrained on, say, horses. They would then do a good job of recognising horses, but when flowers were presented again they were hopeless – they had effectively overwritten their previous flower expertise with horse expertise. This effect was called catastrophic forgetting and shows how unlike human memory such AI memories can be. Other kinds of AI programs forget gradually but none behave the same as our unique human memory.