Comments by Todor to johntcm in his AI group in Discord ~ 17:xx h in room #memory, 16.6.2024 group: https://discord.com/channels/1095629355432034344/1133760747437051964/1251904578677637171
See also: l https://artificial-mind.blogspot.com/2014/08/the-super-science-of-philosophy-and.html
https://artificial-mind.blogspot.com/2013/08/issues-on-agiri-agi-email-list-and-agi.html
(The multi-inter-intra-disciplinary blindness and the actual obviousness of "everything" creative and cognitive, but the lack of proper understanding, knowledge and access to the structure.)
Nobody asked me, but I am not so impressed with these quotes and explanations at this level of generality (including the previous book, the ones which I checked from the "Why We Remember:Unlocking Memory's Power to Hold on to What Matters", 2024)
[The quotes from the other book were from some of https://en.wikipedia.org/wiki/Nicholas_Humphrey I am not sure which exact one, about 30 years ago]
IMO we must be way more technical and the required technicality can't be expressed in that kind of NL; a proper language of thought and code, and referred data, are required.
As of being imaginative, IMO Kant and Schopenhauer were, and quite more insightful than many current "stars", they were 200 years earlier. 30 years ago is not that much time either (or 0 years, current texts), there were pretty advanced computers already, hot discussions on connectionism even since late 1960s/early 1970s, and very hot in mid-late 1980s, there were also "hidden Markov models" (and "hidden variables").
Kant and Schopenhauer in particular explained a lot of material that, it seems, many of current AGI/AI experts still can't really understand or rediscover it and reexpress it as fresh with a gazillion of flops and bytes and with all data and knowledge manually collected, classified and preprocessed already, again and again rediscovering trivial wheels. Everything should be already obvious for the blind, we can "touch" every pixel, every possible transform, every formula, everything.
Kahneman's "System 1 and System 2" are, as far as I understand them, well known and investigated concepts from the German philosophy at least from 200-250 years ago, apodictic/intuitive and discursive knowledge.
"Symbol grounding", the basics of "predictive processing", Understanding [it's a concept, a faculty of mind] as physical simulation of the sensory input/world modeling, intelligence inferring the causes from the effect, is explained by Schopenhauer even in his PhD dissertation in 1813. The cognitive hierarchy or "latent spaces", the weighing of motives and their relation to the Will is explained again at least by Schopenhauer in his subsequent major work "World as Will and Idea", starting in 1818 and later parts and editions, which is translated also as "World as Will and Representation" etc.
It is actually Kant in late 1700s and continued by Schopenhauer, not Alan Turing who first defined the abstract computers: theirs, mainly Kant's, refined by Schopenhauer, definition is the centrality of the a priori conceptions for any thought process (or "computation"): time, space and causality, and the medium "matter". Time, space, causality and matter is the minimum definition of a computer, where:
Space is the memory, time is the process of change, the reading of the next step/instruction/address; causality is the rules, the specific instructions per every possible current state - how current state is transformed to the following; and matter is the substrate which can hold the states, it is the "type", the set of possible values within the memory cells within the space, the capability to have properties and their possible values. The matter is "eternal" within the running simulation or computer, only the "forms" (the current content, the "accidents") change by the laws of the causality (the "principle of sufficient reason") which in computers is the chain of executed instructions, or if it is represented as a state automata - the changes of their states etc. A 2014 article about that etc.: https://artificial-mind.blogspot.com/2014/08/the-super-science-of-philosophy-and.html
Re reward vs goal, I've participated once in such a discussion here: essentially it's the same, it matches to "match" also to any kind of "optimization" (variation calculus). Defined as a path of reward or as a goal and subgoals, it can do the same, each of them can be defined as the other, finding subgoals can be a reward, maximizing prediction can be a reward in the cognitive space.
Some define RL as getting the reward "only at the end of the episode", but even then the episode could be reduced to the shortest step, and be on each step. There could be and there are different "rewards" in different domains, modalities, resolutions, "abstract spaces" (yes, they are multiple: multimodal and there could be branches within the modalities) and they could interact in a multi-agent way. Reaching the goal can be counted as "reward", or achieving "highest reward" for a selected span of steps, time, whatever can be counted as a "goal". All of them is about prediction and minimizing "prediction error", i.e. "will", or matching: matching some target, which can be expressed as "reducing the error" or "maximizing the match".
Also there are at least two layers and types of rewards/goals/prediction in mind, one is sensual, the other one is cognitive, that was discussed here (also taught during the AGI course in 2010,2011), and it seems it was also defined in Schopenhauer's AGI theory, it's even in the title. The Will (see his concept) maps to the sensual reward, which is about matching the desired state, being near it, reducing the error to the *desired* representation. While the Idea (representation) is the Cognitive "Reward" or "Goal" which are the same as conception, and it is about maximizing *prediction*, or prediction *progress*, maximizing the *knowledge*, "epistemic reward", which may be *against* and contradicting to the sensual, survival, preservation reward/goal. The lower the speceis/the individual in the cognitive ladder, the more his cognitive part, his Idea, representations, or Reason in humans are ruling his behavior, and more he is a slave of the sensual goals or rewards, which are the same as in the animals.
John: "Multiple abstraction spaces to process image - position, movement, shape, colour etc.I believe our brains process images via multiple abstraction spaces, not just with one transformer like what we see right now."
Yes, the concepts per se are "abstraction spaces" by definition, concepts/generalization is inducing the spaces for different features and classes. Also different resolutions, different maps to different views/aspects, selections of "items" within the spaces.
0 коментара:
Post a Comment