Work log 2016-07-13

We could learn a compact representation of a generative model; already dont with e.g. graphs.

For each symbol S(i) assign probabilities P(i, 0), P(i, 1), etc.

Initial model: P(S(i)) = P(i, 0)

Expand:

λv(i) -> <term>
<term> <term>
case <term> of
  <pat> -> <term>
  ...

Not the brillest; how about Church Encoding?

~~ λs0 s1 s2 ... -> <term> ~~

For n symbols:

term n = λ s1 s2 ... sn -> <expr n>

expr n = vi | 0 <= 1 < n
       | λ -> <expr (S n)>
       | <expr n> <expr n>

What can we compare recurrent clustering to?

• Non-recurrent clustering
• Simpler feature extraction/distance, e.g. n-grams, bag-of-words, tree kernel

Ranking would be good too; select first N closest matches.

We want to learn a rule from 𝒫(Def) -> 𝒫(Def) (in general). Maybe more practical for 𝒫(Def) -> Def -> [0,1]