Neural networks and stuff (it is not about unboxed structures anymore)
"Yuri Davidovsky (as work at disclosure dot ie)" <[email protected]>
| Newsgroups | gmane.lisp.lispworks.general |
|---|---|
| Message-ID | <[email protected]> |
> On 14 Jan 2026, at 11:15, Tim Bradshaw <[email protected]> wrote: > > Neural networks can simulate a universal Turing machine, with the same caveats that you need an unbounded amount of storage that apply to any such simulation. I do not get this obsession with neural networks having to be a Turing machine, as in if they are not, then they are not useful (referring back to the famous argument that a single layer network cannot do XOR). My firm opinion is that a statistical pattern matcher does not need a mind of its own. Its job is to extract features, normalise them into a form that computer algorithms can understand, and hand the normalised data over to an expert system to make decisions. Think of a UAV being airborne that is scanning the landscape looking up a potential target of interest and deciding whether to hand it over to the operator. The neural network continually performs a visual scan, detecting unexpected statistical outliers in the signal, and propagates them to an actual decision making algorithm after data normalisation. The algorithm then can decide whether to: 1. Ignore the interference. 2. Focus on it to gain more data. 3. Hand it over to the operator. 4. Escape the scene. 5. … whatever else. There are two wings to AI research, remember? It does not have be one or the other. _______________________________________________ Lisp Hug - the mailing list for LispWorks users [email protected] http://www.lispworks.com/support/lisp-hug.html