Skip to content

Patterns for neural symbolic systems

Jump to bottom
Till Mossakowski edited this page Sep 29, 2022 · 7 revisions

Hets features a logic NeSyPatterns that provides a pattern language for the specification of neural-symbolic systems. Support for this logic is still under development. A parser, a printer, static analysis and comuptation of colimits as well as an initial visulation have been developed; some debugging is currently taking place.

How to use it. Consider the following DOL text from the paper below:

%prefix( : <https://ontohub.org/meta/> )%

logic NeSyPatterns    
pattern Model = data NeSyPatterns.omn
  Model;
end
pattern Train = data NeSyPatterns.omn
  Symbol -> Training -> Model;
end
pattern SemanticDeduction = data NeSyPatterns.omn
  Symbol -> d : Deduction -> Symbol;
  Semantic_Model -> d : Deduction; 
end
view R1 : Model to Train =
  Model |-> Model
end
view R2 : Model to SemanticDeduction =
  Model |-> Semantic_Model
end
network N =
  Train, SemanticDeduction, R1, R2
end
pattern SemanticGenerateAndTrain =
  combine N
end

Either

  • store the into a file, say pattern.dol, install Hets, and call hets -g pattern.dol. This will show a graph of different patterns. Right-click on the node at the bottom, and select Taxonomy graphs, then Concept graph. This will display the combined pattern as specified by SemanticGenerateAndTrain.
  • Alternatively, paste the above text into the Hets online page. Then, the graph of patterns will be displayed. You can click on the node at the bottom and get a textual representation of the combined pattern. (However, there is no display of the individual patterns yet.)

References

Clone this wiki locally