Are AI systems limited in their ability to learn/incorporate new information?

This past week Nature magazine published an article at https://doi.org/10.1038/s41586-024-07711-7  titled "Loss of Plasticity in Deep Continual Learning" by Shibhansh Dohare and others.

It basically claims that AI systems can not go on learning new things forever.

In contrast, the Human brain can continually learn and incorporate new information, over the whole of it's "defect free" lifetime.

The paper suggests a possible way of correcting this basic AI learning problem by resetting certain weights used in its neural network. This effectively introduces noise into the system.

Peter Brooks

Palm Bay Florida 

  • Hello Andrew:

    No need to apologize for anything. Life threw you a curve ball (as they say here in the US) and that is that.

    I just gave you an example of what I (and my wife) did under our conditions.

    For example volunteering in a hospital may not be a viable option in the UK. However it got me interested in Medical science. 

    Regarding adding noise in the system - that was a quote by the Guy who wrote the paper - it was made in a podcast the author also made.

    I believe he was saying, for AI systems one has to give the stable system in kick in the butt by changing some the lower level neuron weights, to get it started. In an analog system it would be like adding noise into the system.

    Peter Brooks

    Palm Bay 

      

  • Hi Peter,

    Perhaps Kirsten McCormick and Dr. Elise Tapping could provide a more detailed explanation regarding the introduction of noise into the AI system.


    - Andrew

  • Hello Andrew:

    I am a little surprised that they haven't already contributed to this discussion.

    I did find a web site that deals with this subject:-

    larksuite.com/en_us/topics/ai-glossary/noise 

    but it doesn't seem to suggest changing the weights.

    Peter Brooks

    Palm Bay