On Wed, Aug 21, 2024 at 12:39 PM Matt Mahoney <mattmahone...@gmail.com> wrote:
> ... > > But what I think you are asking is how to convert a neural network to a > set of logical rules that you can understand... > Yan Kiing Yin can correct me if I am wrong about the relationship between abductive logic programming and his approach, but the ultimate grounding is in what logic programming calls "facts" aka literal relationships aka phenomena aka observations aka data. These are logical entities absent any generalization and are the basis for phenomenology: The logic of phenomena. Abductive logic programming starts with such atomic "facts". The way it proceeds from there is to *abduce* rules that generate said "facts" without doing violence to their statistics. The way it does _that_ is similar to the way Prolog (with its admittedly nasty implementation of first order logic) goes about satisfying a relation (a predicate with unbound variables) in a depth-first traversal of the rules to provide answers that satisfy that relation. The biggest problem I see with first order logic is not its inability to be "fuzzy" -- since it clearly can by generating redundant answers to represent more than one case satisfying a relation (redundancy meaning case count for that answer) -- but rather its inability to enter into second order logic where one can treat relationships in "quotes" or to phenomenologically bracket a fact by making the "fact" part of a relationship with provenance: Attributed assertions. This is absolutely essential to deal with scientific forensics: Identifying sources of "bias" -- not in the sense meant by AI alignment researchers (which always presumes a set of *values* that are assumed to be "universal" against which the behavior of AI's is to be judged) -- but rather in the sense meant by a community of interest regarding their selection of data that they deem worthy of taking under consideration for discovering the nature of reality. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T04cca5b54df55d05-M433a36cec4aaa66da8fc0a8f Delivery options: https://agi.topicbox.com/groups/agi/subscription
