Machine learning of the well known things

Abstract

Machine learning (ML) in its current form implies that an answer to any problem can be well approximated by a function of a very peculiar form: a specially adjusted iteration of Heavyside theta-functions. It is natural to ask if the answers to the questions, which we already know, can be naturally represented in this form. We provide elementary, still non-evident examples that this is indeed possible, and suggest to look for a systematic reformulation of existing knowledge in a ML-consistent way. Success or a failure of these attempts can shed light on a variety of problems, both scientific and epistemological.

The Splendors and Miseries of Heavisidisation

“…Together with the results of [2], by now we have the following set of network building blocks (which may be used/combined as sub-networks):…”

“…i.e. points in the segment [2,3] belong to the class 1, and the rest to 0. Lets take a level-1 network.…”

“…In [2] we formulated an applicability condition for ML to scientific problems, i.e. to those which have a well defined ("objective") answer: this answer should have a representation G : X −→ Z with G( x) expressed as a combination of iterated Heaviside functions.…”

“…In this paper we further elaborate on the elementary examples of [2] and attempt to extend them further. One of the goals (not fully achieved yet) is Heavisidisation of algebraic numbers.…”

Dual Geometry of Entanglement Entropy via Deep Learning

“…In Ref. [18][19][20][21][22], the authors utilized deep layers satisfying a specific recurrence relation. In this case, experimental or observational data allow us to find the recurrence relation determining the bulk geometry.…”