Romuald Texier wrote: > > Ken Seehof wrote: > > > Excellent idea, Dan. That's conveniently sidesteps the most difficult > > issue: getting the neural network to actually come up with linguistic > > rules. Once an intelligent human specifies the set of rules, the neural > > net should have no difficulty coming up with an optimal non-linear > > function of pre-processed features (i.e. the "rules") to identify spam. > > Analysis of the weights after training will help remove rules that turn > > out to be irrelevant. > > Wouldn't decision tree or other rule inference algorithms be more accurate > than neural networks for that kind of machine learning ? Moreover, neural > nets are "black box" : you do not get a logical rule set (that may be > edited by humans or exchanged) but an ugly matrix of floats... > > -- > Romuald Texier Not for all neural net algorithms. There are a variety of neural net algorithms that, when backpropagation is applied to a node, it's closest neighbour also recieves a little bit of this weighting adjustment as well. What you end up with is a grouping of similarly behaved nodes, which should be easier to extract rule sets from. However, if a simple weighting of rules are all that one is after, maybe a genetic algorithm is all that is required, which does give the benefit of simple rule extraction by humans. Give it an ~ 20 sized population of weightings, apply stociastic propagation with all the mutation, cross-over functions etc., and let the population climb to an optimal peak. Of course, neural nets have the advantage that they include the posibility of non-linear behaviour. To include non-linearity for the genetic algorithm, you'd need to do things like this: f_n is the n'th filtering rule w_n is the linear weighting associated with this rule w_nm is the bilinear weighting associated with rules n and m And run the genetic algoritm for this, with penalties for using w_nm's. Only hassle is, for a rule set of size N, a population that tries to solve a complete solution has to solve for N + N^2 (allowing for w_nn values) weightings. Maybe a genetic algorithm which allows incomplete but expandable solutions, that penalises for computational size? Everything I wrote beyond this point was too vague to include so I'm shutting up now. ____________snip___________________ Joal Heagney/AncientHart
RetroSearch is an open source project built by @garambo | Open a GitHub Issue
Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo
HTML:
3.2
| Encoding:
UTF-8
| Version:
0.7.4