Intervene in the operations of a language model using an interactive debugger

LM-Debugger builds upon our findings from this work https://t.co/fnEBgFe4b8
to provide three core capabilities for single-prediction debugging and model analysis 2/8 pic.twitter.com/Boer9BVEJD

— Mor Geva (@megamor2) April 27, 2022

(2) It also lets the user intervene in the prediction process by changing the weights of FFN updates of her choice, e.g. increasing (decreasing) an update that promotes music-related (teaching-related) tokens 4/8

— Mor Geva (@megamor2) April 27, 2022

Check out a demonstration of LM-Debugger here:https://t.co/nm6G4jrZCN
And try it out for yourself with our two demos:
GPT2 Medium: https://t.co/supnwK26CU
GPT2 Large: https://t.co/mYyZxKsm1n
6/8

— Mor Geva (@megamor2) April 27, 2022

Be sure to check out the video. It's very cool. Pay close attention, though. If you've never worked with such systems – I haven't – you may be puzzled on first viewing. It made more sense to me the second time around.

What I'm  wondering is if something like this could be used to "bootstrap" a symbolic component into a neural net. I'm thinking of some posts where I discuss Vygotsky's account of language acquisition: Vygotsky Tutorial (for Connected Courses) and this one, Dual-system mentation in humans and machines [updated]. The second one involves a hybrid AI system with symbolic and neural net components.