Researchers at the Tokyo University of Science developed a means of making AI forgetful, reasoning current large language models are energy-intensive, and raise privacy and security issues due to being all-encompassing.

In a research paper, the university’s team detail their development of Black-Box Forgetting, a method of optimising the text prompts used in AI models to make it overlook certain classes of data.

The research programme was headed by associate professor Go Irie, who argued it is often unnecessary for all object classes to be classified in real-world uses.

“For example, in an autonomous driving system, it would be sufficient to recognise limited classes of objects such as cars, pedestrians and traffic signs.”

The academic explained putting effort into training AI to cover “classes that do not need to be recognised” could be detrimental to the overall accuracy of classifications, while also raising the prospect of wasting “computational resources” and increasing the “risk of information leakage”.

In many practical scenarios, people want AI to “fulfil specific roles rather than be jacks-of-all-trades”, the team argues.

They noted there already some methods to make pre-trained models forget, but explained these “assume a white box setting, where the user has access to the internal parameters and architecture”.

In the case of black-box models, where users do not have this access, they worked to develop a “so-called derivative-free optimisation strategy” requiring no access to the underlying set-up.

Method
The researchers expanded the covariance matrix adaptation evolution strategy (CMA-ES) method, using the contrastive language-image pre-training (CLIP) models AI companies use as their target.

What they call an “evolutionary algorithm” samples “various candidate prompts to feed to the model” and then evaluate results using defined “objective functions, updating a multivariate distribution based on the calculated values”.

To address degradation in the performance of the approach when dealing with larger problems, the team developed “a new parameterisation technique called latent context sharing”, which handles smaller chunks “considered to be unique to a prompt token”, in turn making the whole process more manageable.

Benchmark testing proved the approach was effective in forgetting 40 per cent of the classes in a given dataset.

Irie argued the outcome could help service providers to comply with privacy requirements without retraining a model from scratch.