39. boris kazachenko Afficher le profil Traduire en Français Autres options 2 mai 1998, 09:00 I agree that to build an Intelligence it is helpfull to define it, & that Minski & Co are full of shit. I also agree that Intelligence is a learning ability, more specifically an ability to build a predictive model of both subject system & its environment, to the degree that it may impact the subject. Since prediction, including planing, is an interactive projection of past Spatio-Temporal Patterns to the degree of their recurrence, main task of Intelligence is to find such recurrence within input, primarily video & audio data flow. I am working on a technique that would perform search for recurrence, directed by past recurrence & constrained by overhead, compressed coding of recurrent patterns by forming new variables such as length of recurrence, & elevation of sufficiently compressed (restorable input per record) patterns to higher levels of generality with expanded range of search, greater overhead & syntactic complexity. . This is how my technique can by applied to earlier steps of Image Compression/Image Recognition: Search starts from the begining: brightness comparison between ajacent pixels within one line. In case of match 1D Patterns are formed adding Length of recurrence for each derivative of original Brightness: L,B,(L,dB,(L,ddB))), where higher derivative is formed with match of a lower one, & top deriv. represents 1D Contrast. On the next level of search 1D Ps are assigned Horizontal Coordinates & compared with those of overlaping Coordinates & ajacent lines, in case of match forming 2D Ps with full complement of new V-types. The next step is comparison among 2D Ps with Vertical Coordinates to those of overlaping H & V Coords from Co-Focused 2D Frames with a given Base & Angle, to form 3D Ps. Continuous Search is exhausted by forming Temporal Patterns from matching 3D Ps of sequential frames of video with overlaping H+V+D Coords. Subsequently search proceeds through multiple levels of discontinuous 4D Ps with exponentially growing degree of compression, syntactic complexity, & range of Spat/Temp Coordinates. This is a rough outline of my approach. Let me know if you find it promising. Boris Kazacenko ## 04:00 21.12.2010 г. modlin Afficher le profil Traduire en Français Autres options 2 mai 1998, 09:00 In <6if6h4$8b...@supernews.com>, "boris kazachenko" writes: - Afficher le texte des messages précédents - [snip] Thank you for your support, though I think "full of shit" is wrong. I disagree vehemently with Minsky about the proper direction for AI, but he has done much interesting and valuable work for many years. If he were simply "full of shit" it would be easy to ignore him. Unfortunately he isn't, and his influence in pushing the field in unprofitable directions has been and remains large. You and I do seem to share related views of the need for multi-layered pattern recognition and predictive modelling, but our methods are so different that I can't really see how to merge them. Your algorithms are closely tied to the specific kinds of data you are processing and the nature of the patterns you expect to find. You therefore need different algorithms for audio and video data, and different algorithms for each level or layer of your processing. In that respect you are no different from Minsky, with his Society of Mind vision of uncounted layers and networks of individually specialized agents, each doing some small task of its own with methods evolved or designed somehow to be right for that task and no other. That's what I'm trying to get away from, by casting all the pattern recognition and predictive modelling functions into the same abstract space of statistical relationships among discrete events with mutable defining functions. I expect that the end result will be the emergence of just the sort of specialized interacting functions posited in Society of Mind, so that Minskys description isn't wrong... merely irrelevant. We can't in practice start with a description of all those millions of agents and purpose-make a design for each, since the result would be a rigid and brittle network able only to do those things we anticipated in our design. Instead, we must start with mechanisms common to all the agents and the principles whereby each becomes specialized to a purpose appropriate to the role it defines for itself in the larger network. These self-defining roles can then continue to evolve over the life of the network, following the same rules and principles which brought them into existence in the first place, and enabling the whole to adapt to changing conditions as they are encountered. Only with such an approach can we achieve true intelligence and mind. Bill Modlin Transférer 41. boris kazachenko Afficher le profil Traduire en Français Autres options 4 mai 1998, 09:00 From: "boris kazachenko" To: Newsgroups: comp.ai.philosophy,comp.ai References: <6i8vd9$s0r@examiner.concentric.net> <6if6h4$8b...@supernews.com> <6iffhi$...@examiner.concentric.net> Subject: Re: intelligence as a learning ability Date: Sun, 3 May 1998 09:29:34 -0400 X-Unsent: 1 X-MimeOLE: Produced By Microsoft MimeOLE V4.71.1712.3 mod...@concentric.net wrote in message <6iffhi$...@examiner.concentric.net>... - Afficher le texte des messages précédents - I believe the function of intelligence is discovery, compressed record, & interactive projection of recurrence, so what you call functions & Minsky calls agents I call patterns, which are extracted from input rather than programmed manually. >That's what I'm trying to get away from, by casting all the pattern >recognition and predictive modelling functions into the same abstract >space of statistical relationships among discrete events with mutable >defining functions. I lost you here, could you explain it please? - Afficher le texte des messages précédents - I believe I am suggesting a universal method, at least to the extend that spatio-temporal continuity is universal. The basic property of this universe is a degree of order: probability of recurrence corresponds to proximity. The only alternative would be Random Universe, in which patterns are accidental, carry no predictive value, & Intelligence is useless. Hence the expectation that probability of recurrence would decrease with the distance, & search must start at max.resolution (pixel is only an example) & max.proximity, & expand only for patterns with critical past recurrence, which results in the degree of non-destructive compression, or concentration of reference. It is true that dimensionally sequential search expansion that I proposed is empirically specific, but, again, 4D Space-Time is common accross the known universe. Besides, reduction or expansion of dimensionality can be independently discovered & accomodated with no change in the method. This method can be used to process data from any analog(sensory) sources. The difference between sources in resolution, dimensionality, distance, & direction of inputs can be programmed or learned by comparing patterns accross sources at various levels of generality. The method is also the same for search accross levels of generality, although patterns have to be syntactically synchronized so that comparison proceeds between variables of the same type, which is the same for analog data, except that there is only one V-type. As for processing linguistically coded (symbolic) data, it seems obvious that all such data is ultimately derived, by wetware or hardware, from analog sources, as as it can be done using my method. So, if the system posesses generalised model of environment that overlaps the one that is described by symbolic data, & this data is elevated through sequential levels of generality, at one of these levels it will match, perhaps partially, symbolic patterns generated internally. It does seem that we agree on principles, which is a rarity in the field, & we also agree that the problem is not empirically specific, which means that the solution must be derived from these principles. If we are consistent we can't arrive at different conclusions. I do admit to a complete ignorance of your methods though, so perhaps you could enlighten me. Boris Kazachenko