How to you get AI generated models to forget information that it was trained on, when that information was later retracted by it's authors?

During a human lifetime we learn facts that are later determined to be incomplete or wrong.

We effectively mark in our memory that retraction or modification has taken place. 

Over the past year numerous research papers has be retracted due to errors or fraud by the authors.

How do we know that these retracted research papers have not been used during the AI training cycle?

Then how do we get the AI developed model to forget this defective information? 

Peter Brooks

Palm Bay Florida 

  • Wow! And I thought I overthought! Joy

  • Throughout  a human lifetime, we all accumulate knowledge that may later be revised or refuted. Our cognitive flexibility allows us to update our understanding and correct misconceptions as new information becomes available. 

  • We effectively mark in our memory that retraction or modification has taken place.

    Only to some extent - even in traditional systems, incorrect or superseded information tends to hang around for quite a while - not just in collective memory (which can be surprisingly persistent) but in secondary texts (e.g. textbooks and other papers written when the original information what still thought to be current) - and we cope with that adequately. I suspect the situation with AI won't be much different. Some of the current publicly accessible AI models are based on a snapshot if the internet (from Sept 2021 IIRC) - so at some point they're be refreshed by taking a new shapshot - so removed information should then be lost from their knowledge. If they start to add to models continually rather than doing snapshots, then outdated information should gradually be filtered out.

       - Andy.

  • Training or retraining of AI systems is a very expensive proposition.

    As you correctly observed the incorrect information can live on in texts for a very long time. (The Bible is a good example)

    I personally remember attending a lecture by Fred Hoyle in the 1950's where he discussed his continuous creation theory of the universe, which was later retracted because of the Big Bang Theory. Now the Big Bang Theory is being question, as it doesn't consider quantum mechanics and virtual particles.

    What caused me to create this question was the recent retraction of a major medical research (often quoted) paper from 2005. How can we make sure that the AI systems are revised to eliminate this defective information?

    Peter Brooks

    Palm Bay Florida 

  • It is time to drop the other shoe concerning AI learning!

    How can we develop unbiased AI systems when important information is not formally published?

    Nobody wants to publish a research paper when the experiment/test did not support their pre-assumptions!

    This is very obvious in drug development where a large quality of trials on humans or animals are never published because they didn't work or had serious side effects. 

    We only reach the truth by highlighting our failures and determining the root causes.

    It also seems to me that the IET organization is very hesitant in publishing articles involving UK failures.

    Peter Brooks

    Palm Bay 

  • Hi Peter. Acknowledging the presence of public bias, it is imperative to conduct thorough analyses of failures alongside successes to mitigate the development of biased AI systems. The revelation regarding the IET’s publication practices is indeed surprising (Competency E5 - UK spec Joy) It underscores the necessity for open access to all research data, encompassing both successful and unsuccessful outcomes, to foster a more transparent and equitable field of artificial intelligence.

  • As you are aware ( I assume) the IET has out sourced it's bi-monthly and it's week day web site publications and it's focus has narrowed with the reduction in staff size.

    It should be noted that I typically forward two or more world wide validated articles each day to them, for their consideration.

    I do this not to get them published, but to try and train the team to expand their viewpoints. 

    Peter Brooks

    Palm Bay 

  • Your  commitment to enriching the IET’s perspectives is truly commendable. Thumbsup

    I personally wasn't aware of the IET’s outsourcing strategy for its bi-monthly and weekday website publications. Maybe initiatives as yours will become instrumental in catalysing a shift and potentially prompting the team to embrace a more extensive array of topics and perspectives.

  • The outsourcing was to Redactive.co.uk.

    Not too sure if I have mentioned it before, but I am also a member of the IEEE which publishes a low tech "Spectrum" magazine, like the E&T magazine. The IEEE also publishes "Proceedings" which contains "high tech" articles.

    Sometimes the IEEE Spectrum will publish very thoughtful articles about  ChatGPT and other AI software.

    If they are significant, I will also forward them to Ed Almond.

    Peter Brooks

    Palm Bay

  • Hi Peter 

    I wasn’t aware that you were a member of the IEEE, interesting! I just looked them up. At first, I thought IEEE was what the IET used to be called (IEE). Speaking of which, who is Ed Almond?

    The Spectrum magazine, definitely something I’d be keen to read.