Comment on:
From "Trends in Cognitive Scinece", vol.22, April 2018.An opinion article
Hierarchical Active Inference: A Theory of Motivated Control
Giovanni Pezzulo, Francesco Rigoli, Karl J.Friston
https://doi.org/10.1016/j.tics.2018.01.009
It's an excellent paper, would be insightful and accessible for beginners in AGI, psychologists and for seeing the big picture like "On intelligence" etc. and for readers who like divergent thinking and seeing mappings to real agent behavior and "macro" phenomenons. Good questions, huge set of references, mapping to brain areas and nueroscience research.
However as of architectural, philosophical ideas it sounds too similar to my own"Theory of Universe and Mind", published in articles mainly in the Bulgarian avant-garde e-zine "Sacred Computer" (Свещеният сметач) between late 2001 and early 2004. Its ideas were presented/suggested also in the world's first University course in AGI in 2010 and 2011.
Thanks to Eray Ozkural who was familiar with Friston's work and we had an interesting discussion in Montreal.AI FB's page, see a recent post regarding his work in "Ultimate AI" and "AI Unification".
The term "active inference" sounds pretentious, it means using a model of the world in order to act, I assume in opposite to being simply reactive as in simpler RL models. However IMO that's supposed to be obvious, see below.
Theory of Universe and Mind
The terminology and content of that 2018 "opinion" paper strongly reminded me of the teenage writings of myself from the early 2000s. The term "control" (the cybernetics influence), the need of prediction/reward computation at different time scales/time steps, cascade increment of the precision (resolution of control and resolution in perception); specific examples of "behaviorintrospective" analysis and specific selection of the actions etc.
"Theory of Universe and Mind", or "my theory", started with the hierarchy and "active inference" as obvious requirements (not only to me, I believe).
Mind is a hierarchical simulator of virtual universes, it makes predictions - controls ("cause" is a better term, though) at the lowest level. The hierarchical simulations are built from the patterns in the experience. Highest levels are built of sequences of selected patterns at lower level ("instructions", correlations) which are predictive.
At the lowest level all combinations of instructions are legal, the system shouldn't hang.
However at the higher levels, only selected ones work, not all combinations of low level instructions are correct which makes the search cheaper. That implies reduction of the possible legal states, which as far as I understand in F.'s terms is called "reduction of the free energy".
So the mind, the hierarchy of virtual universes, makes predictions about the future in the Universe - as perceived at the lowest level virtual universe - and causes desired changes in the input, by aiming at maximizing the desired match.
Through time it aims at increasing the resolution of perception and causality-control while increasing also the range of prediction and causation. That's what a human does in her own personal development, as well as what the humanity's "super mind", the edge of science and technology.
My old writings were also explicit about the predictions at different time-scales, precisions and domains - for a sophisticated mind there's no one single "best" objective behavioral trajectory, there are many, because there are contradictory reward-domains (like not eating chocolate, because it may lead to dental cavities, or eating it, because it's a pleasure now). There's also a prediction horizon, uncertainty.
In the domain of Reinforcement learning, there are two types of reward, called "cognitive" and "physical". Cognitive is about making correct predictions, that is "curiosity", exploration etc., while physical is about having the desired input in the senses, implying a desired state - or "pleasure".
There must be accord between these domains and a sophisticated enough hierarchy and various time-space-precision ranges, otherwise the system would fall into a vicious cycle and have an "addiction".
In the paper, they have called my cognitive reward/drive "cold domain" (choice probability, plans action sequences, policies) and my "physical" one - "hot domain" (homeostasis, incentive values, rewards).
Etc.
...
The "Theory of Universe and Mind" works and the 2010's slides could be found in this blog, in the online archives of "Sacred Computer" (Свещеният сметач - the original texts in Bulgarian), and on twenkid.com, http://research.twenkid.com/agi/2010/
From "Trends in Cognitive Scinece", vol.22, April 2018.An opinion article
Hierarchical Active Inference: A Theory of Motivated Control
Giovanni Pezzulo, Francesco Rigoli, Karl J.Friston
https://doi.org/10.1016/j.tics.2018.01.009
It's an excellent paper, would be insightful and accessible for beginners in AGI, psychologists and for seeing the big picture like "On intelligence" etc. and for readers who like divergent thinking and seeing mappings to real agent behavior and "macro" phenomenons. Good questions, huge set of references, mapping to brain areas and nueroscience research.
However as of architectural, philosophical ideas it sounds too similar to my own"Theory of Universe and Mind", published in articles mainly in the Bulgarian avant-garde e-zine "Sacred Computer" (Свещеният сметач) between late 2001 and early 2004. Its ideas were presented/suggested also in the world's first University course in AGI in 2010 and 2011.
Thanks to Eray Ozkural who was familiar with Friston's work and we had an interesting discussion in Montreal.AI FB's page, see a recent post regarding his work in "Ultimate AI" and "AI Unification".
The term "active inference" sounds pretentious, it means using a model of the world in order to act, I assume in opposite to being simply reactive as in simpler RL models. However IMO that's supposed to be obvious, see below.
Theory of Universe and Mind
The terminology and content of that 2018 "opinion" paper strongly reminded me of the teenage writings of myself from the early 2000s. The term "control" (the cybernetics influence), the need of prediction/reward computation at different time scales/time steps, cascade increment of the precision (resolution of control and resolution in perception); specific examples of "behaviorintrospective" analysis and specific selection of the actions etc.
"Theory of Universe and Mind", or "my theory", started with the hierarchy and "active inference" as obvious requirements (not only to me, I believe).
Mind is a hierarchical simulator of virtual universes, it makes predictions - controls ("cause" is a better term, though) at the lowest level. The hierarchical simulations are built from the patterns in the experience. Highest levels are built of sequences of selected patterns at lower level ("instructions", correlations) which are predictive.
At the lowest level all combinations of instructions are legal, the system shouldn't hang.
However at the higher levels, only selected ones work, not all combinations of low level instructions are correct which makes the search cheaper. That implies reduction of the possible legal states, which as far as I understand in F.'s terms is called "reduction of the free energy".
So the mind, the hierarchy of virtual universes, makes predictions about the future in the Universe - as perceived at the lowest level virtual universe - and causes desired changes in the input, by aiming at maximizing the desired match.
Through time it aims at increasing the resolution of perception and causality-control while increasing also the range of prediction and causation. That's what a human does in her own personal development, as well as what the humanity's "super mind", the edge of science and technology.
My old writings were also explicit about the predictions at different time-scales, precisions and domains - for a sophisticated mind there's no one single "best" objective behavioral trajectory, there are many, because there are contradictory reward-domains (like not eating chocolate, because it may lead to dental cavities, or eating it, because it's a pleasure now). There's also a prediction horizon, uncertainty.
In the domain of Reinforcement learning, there are two types of reward, called "cognitive" and "physical". Cognitive is about making correct predictions, that is "curiosity", exploration etc., while physical is about having the desired input in the senses, implying a desired state - or "pleasure".
There must be accord between these domains and a sophisticated enough hierarchy and various time-space-precision ranges, otherwise the system would fall into a vicious cycle and have an "addiction".
In the paper, they have called my cognitive reward/drive "cold domain" (choice probability, plans action sequences, policies) and my "physical" one - "hot domain" (homeostasis, incentive values, rewards).
Etc.
...
The "Theory of Universe and Mind" works and the 2010's slides could be found in this blog, in the online archives of "Sacred Computer" (Свещеният сметач - the original texts in Bulgarian), and on twenkid.com, http://research.twenkid.com/agi/2010/
0 коментара:
Post a Comment