On Sat, May 19, 2018 at 1:00 PM, Alexey Potapov <pota...@aideus.com> wrote:
> > > Well... traditional probabilistic programming is a logical probabilistic > programming. It's definitely not about lambda-calculus. > I don't know what to do with this statement. There is a famous theorem, the church-turing theorem, dating to the 1930's, that states that anything turing-computable is equivalent to lambda calculus. There have been many extensions, refinements, generalizations and clarifications of that theorem, since then. If you have a probabilistic programming language working on a modern-day digital computer, then its lambda-calculus. If you have a theoretical algebra working on infinite-precision topological spaces, that's something else. The quantum-computing machines are often understood as infinite-precision topological vector-space machines (where the space is complex-projective, and the operators are unitary). Topological computing is .. interesting, but I never got the sense, from quick skims of the literature, that this is what was being explored. > I think much of what neural nets and deep learning do also fits into this > general framework; I want to write a paper on > this, but have not had the > time yet. > > You can also map a functional programming (with algebraic types, pattern matching, etc.) > to neural networks. One my student has written a nice diploma work on this topic. > So, it's cool, but this doen't give us much per se... Well, one of the problems in the unsupervised natural-language-learning project is to factor large tensor products into approximately diagonal components. The factorization can be done slowly, by walking over all elements, comparing them sorting them. I claim that the factorization can also be done quickly, using NN algorithms, but discussions about this have always gotten stuck in various misunderstandings. Thus, having this explicitly written down is important. In a very abstract, hand-wavey fashion: there is this general concept of "integrated information". The unsupervised natural-language-learning project is all about finding the those parts which are least-integrated, and performing explicit cuts there. What remains are the highly-integrated parts, grouped up into classes: nouns, verbs, morphemes, syntactic relations, semantic similarity, etc. I guess you could say that its "discrimination", but the field is not some 2D pixel field, but instead this certain abstract graph. -- Linas -- cassette tapes - analog TV - film cameras - you -- You received this message because you are subscribed to the Google Groups "opencog" group. To unsubscribe from this group and stop receiving emails from it, send an email to opencog+unsubscr...@googlegroups.com. To post to this group, send email to opencog@googlegroups.com. Visit this group at https://groups.google.com/group/opencog. To view this discussion on the web visit https://groups.google.com/d/msgid/opencog/CAHrUA37Vq5-wSTaE5XmBG35qBDYHzAMKPqSjvLL8mvyNJ5NRVg%40mail.gmail.com. For more options, visit https://groups.google.com/d/optout.
