Two thoughts, which aren’t really coherent or informed enough to be called questions.
If naive Bayes == neural network, and you need 3 layers of neuron to make a general classifier(hyperplanes → convex hulls → arbitrary hypershapes), do you need 3 layers of naive Bayes?
Would an algorithm for inducing names be: poke around looking for things that have mutual information but are not screened off, and induce a name that screens them off. When you find names that have mutual information, decide whether they ought to be merged (a clump of all the pieces has mutual information) or clustered under a hierarchically higher name (otherwise).
Two thoughts, which aren’t really coherent or informed enough to be called questions.
If naive Bayes == neural network, and you need 3 layers of neuron to make a general classifier(hyperplanes → convex hulls → arbitrary hypershapes), do you need 3 layers of naive Bayes?
Would an algorithm for inducing names be: poke around looking for things that have mutual information but are not screened off, and induce a name that screens them off. When you find names that have mutual information, decide whether they ought to be merged (a clump of all the pieces has mutual information) or clustered under a hierarchically higher name (otherwise).