search

nuric / pix2rule

From pixels to symbolic rule learning
https://arxiv.org/abs/2106.07487
MIT License
12 stars 3 forks source link

Enquire of different semi-symbolic layer #4

Closed less-and-less-bugs closed 1 year ago

less-and-less-bugs commented 1 year ago

Thanks for your great work!

What's the difference between Base DNF, weight layer, and real-value layer? I have not seen this point in your paper yet.

Lookforward to your reply!

nuric commented 1 year ago

These are the different formulations we tried for the DNF layer but ended up using only the WeightedDNF in the paper. The BaseDNF sets up the permutation and computation and other classes override how the semi-symbolic layer is computed.

RealDNF uses soft_minimum to compute conjunction and DNF (plain version) uses your regular multiplication to compute conjunction. You can trace the differences in the compute_conjunction , reduce_existential and compute_disjunction functions.

With RealDNF the assumption was truth values have no bounds (they are real numbers) with plain DNF they are constrained to fuzzy logic [0,1] ranges and Weighted DNF constrains them to [-1,1] and uses tanh.

Hope that helps.

less-and-less-bugs commented 1 year ago

Thanks for your reply!