Sunday, January 5, 2020

CogAlg News - January 2020

Several new developers are trying to join lately, but the code progress seems cosmetic. The project has finally accepted my mid 2019 notice that the low level Frame_blobs of CogAlg in fact is a form of breadh-first-search traversal vertical flood-filll. It seems I've been cited, but without being referenced:
https://github.com/boris-kz/CogAlg/commit/33db6b0cf171591e0da52f6a267908ccd58bc263#diff-9b63966e9c5ddf883920a1c522318f75

I challenge the final claims in Readme:

https://github.com/boris-kz/CogAlg/commit/b0a88d9c54156983bc43cf2c7e2504427751c534

B.K.: "...In 2D image processing, basic comparison is done by edge detectors, which form gradient and its angle. They are used as first layer in the proposed model, same as in CNN. It then segments image into blobs (2D patterns) by the sign of gradient deviation, which is also pretty conventional. But these blobs are parameterized with summed pixel-level intensity, derivatives (initially gradient and angle) and dimensions. I don’t know of any model that performs such parameterization, so the algorithm seems to be novel from this point on...."
Exactly the same - I don't know, - but similar ideas of incremental evaluation of whatever measurements (derivatives) - I don't think so. For instance the pre-CNN methods for Object recognition in Computer vision and Shape analysis in topology (3D-reconstruction, fixing broken laser scans, medical imaging) use parameters - "derivatives" - which are coordinates/paths, "dimensions"(boxes), curves (contours), curvatures, lengths, distances, angles, ratios, areas, normals. The "pixel-level intensities" of a topology map could be vector fields with curvature, "heat maps"etc., set of derivatives and operations on them.

RAZDEL 2004

In published works of mine, see e.g. that one from my fresh months at the University, working in sound domain, done in late 2003 - early 2004: http://eim.twenkid.com/old/_5/31/analiz_na_zvuk.htm

Notice the data structure which encompassed the basic patterns (almost-periodic functions):

typedef struct DATA
       {
         int begin;
         int end;
         char type;
         unsigned char cycles;
         Zvukk max;
         Zvukk min;
         Zvukk absmax;
         Zvukk absmaxcycle;
         int period;
         char position;
         int changes;
         int accumulated;
         int average;
         Zvukk primer[MAXPRIMER];
       };

BK's CogAlg, which started to have any code just a few years ago, and its actual function started to become intelligible just about an year ago, has similar basic parameters like "ave" (average, filters); it has summation of the "pixel-level intensities" (lowest level input) as one of the basic operators (accumulated here), it records explicit coordinates/"dimension"(begin, end - location in the record), starts with basic types of classification - 2 in CogAlg, 3 here - silence, noise, tone; corresponding to the binary positive and negative patterns and their sign.

RAZDEL is also a bit more bottom-up at that lowest level, because it searches for correlated sequences and adjusts the first filter (ave) correspondingly to the actual sampled input. Also it doesn't do a comparison of the whole record ("frame" in current CogAlg) with one fixed ave, which CogAlg does, i.e. it is more granular. The usage of fixed filter in the latter is a binary thresholding, which you could see illustrated in the pictures from SuperCogAlg. Razdel has a separate "same-filter-span" in per pattern, in CogAlg terms.

The article introduces incremental derivative ...
"Sacred Computer" #31, published 4.2005. The article is written and first published in 3.2004 internally to the Plovdiv University Research Institute in Computational Linguistics:

(...)

The correlations which we could use are for example:

(...)

Etc. other later works to be reviewed later (...)

* Thanks to C. for a recent discussion and comments