Re: unboxed structure fields (or class slots)
"David McClain (as dbm at refined-audiometrics dot com)" <[email protected]>
| Newsgroups | gmane.lisp.lispworks.general |
|---|---|
| Message-ID | <[email protected]> |
That last point bears some emphasis… up until the recent convolutional tricks used in LLMs, you had to just guess at how to form feature vectors. There was no directed search for proper feature vectors. We barely understood how the NNs actually learn the interior boundaries. Frankly, back propagation training seems like the weakest approach you could devise. But we don’t know how else to deterministically train a recognition network to recognize data regions that we can’t humanly visualize. > On Jan 14, 2026, at 14:58, David McClain (as dbm at refined-audiometrics dot com) <[email protected]> wrote: > >> No, this is not how it works. You do not propagate answers back to the input: > > > That’s cute… my abbreviated description prompted your response. > > >>> I don’t see a path to feedback training of Perceptron networks, with any layer depth. >> >> Well, that has been happening a lot in the last 15 or so years, and as you can judge by the current LLM boom, not necessarily unsuccessfully. > > Example from my past experience… 250K training iterations on the “Rubber Mirror”, taking all weekend long on a DG Eclipse. That doesn’t sound very amenable to immediate training feedback. > > >> They are not as good as humans, but there have been giant leaps of progress in the recent years. China has built surveillance systems that > > Actually, they have always been much better than humans for higher dimensional data. The very best I have seen from humans are the projection plots over data in an R Program. That is to say… not very good. You have to be lucky to guess a good set of feature vectors to apply to your data, shake it up nonlinearly if you can, and hope to see some separation in the higher dimensional space, as typified by a NN classifier. > > > > > _______________________________________________ > Lisp Hug - the mailing list for LispWorks users > [email protected] > http://www.lispworks.com/support/lisp-hug.html _______________________________________________ Lisp Hug - the mailing list for LispWorks users [email protected] http://www.lispworks.com/support/lisp-hug.html