4 ms·

Hybrid neurosymbolic AI is promising I'd say

by The_rationalist 6y ago

bsaul 6y ago
any recent link on existing work in that field ? i'm super interested to know if any progress is being done there.
- The_rationalist 6y ago
  There's a lot of things to say but I'm just going to do a non exhaustive lazy comment: Symbolic AI is a big family: For datasets babelnet has been a big improvement but is not that much recent. For symbolic formats the classic OWL/RDF way has been enhanced with SHACL which allow I believe to dynamically create schema that validate the semantic web graph HyperGraphs and metagraphs are trending (check it out it's a datastructure that is actually more expressive than graphs) Truth is: Hybrid AI is not original, it is obvious to anyone knowledgeable that you must use sota neural networks for preprocessing tasks such as POS tagging, universal dependency parsing, NER, etc Those gathered metadata will feed the symbolic engine/DB The truth is that almost zero hybrid AI systems exists or they have a very limited scope. Some tasks such as coreference resolution are currently neural only but a symbolic semantic understanding would improve the state of the art, same for word sense disambiguation. Even the task of semantic parsing is almost only using neural networks, symbolic just isn't funded nowadays or researchers are ashamed of giving a new life to this paradigm which is ridiculous. On the other hand some researchers are trying too hard to implement logic in neural networks which is to me a by design mismatch. The only projects worth following that I know are: Cyc Wolfram language And the only "promising" one opencog. I'm currently researching on how to implement true semantic parsing but I'm not gonna expand on my work until it is ready. To sum it up: NLU is not ready for primetime and almost nobody is truly working on it (but a lot of people pretend to work on it by just pushing a bit further the illusion of understanding that comes from amalgaming correlation as causation)

Hybrid neurosymbolic AI is promising I'd say

bsaul 6y ago

any recent link on existing work in that field ? i'm super interested to know if any progress is being done there.

The_rationalist 6y ago
There's a lot of things to say but I'm just going to do a non exhaustive lazy comment: Symbolic AI is a big family: For datasets babelnet has been a big improvement but is not that much recent. For symbolic formats the classic OWL/RDF way has been enhanced with SHACL which allow I believe to dynamically create schema that validate the semantic web graph HyperGraphs and metagraphs are trending (check it out it's a datastructure that is actually more expressive than graphs) Truth is: Hybrid AI is not original, it is obvious to anyone knowledgeable that you must use sota neural networks for preprocessing tasks such as POS tagging, universal dependency parsing, NER, etc Those gathered metadata will feed the symbolic engine/DB The truth is that almost zero hybrid AI systems exists or they have a very limited scope. Some tasks such as coreference resolution are currently neural only but a symbolic semantic understanding would improve the state of the art, same for word sense disambiguation. Even the task of semantic parsing is almost only using neural networks, symbolic just isn't funded nowadays or researchers are ashamed of giving a new life to this paradigm which is ridiculous. On the other hand some researchers are trying too hard to implement logic in neural networks which is to me a by design mismatch. The only projects worth following that I know are: Cyc Wolfram language And the only "promising" one opencog. I'm currently researching on how to implement true semantic parsing but I'm not gonna expand on my work until it is ready. To sum it up: NLU is not ready for primetime and almost nobody is truly working on it (but a lot of people pretend to work on it by just pushing a bit further the illusion of understanding that comes from amalgaming correlation as causation)