Continues from Frontal Lobe Activation Patterns in Pessimistic & Optimistic Brains, and in Infant Brain Before and After Understanding of Object Permanence
1. Higher-to-Lower level feedback is less efficient than Lower-To-Higher level feed-forward generalization.
Example: Image/Object Recognition vs Image/Object Rendering, performed by humans.
Every healthy child can recognize human faces, understand emotions and react accordingly, but it's not that easy when intentions are involved - feedback/output to act over environment. Realistic drawing or painting of faces or good acting in a film require time, talent and practice.*
(*I guess I can get the critics that regarding painting/drawing it's just "precision vs scope", some autistic people are great in copying inputs. However drawing by memory and creative drawing of imaginary subjects require both scope and precision and capability to keep consistent mapping between all the levels, from the highest to the lowest.)
2. Generalization out of specifics (rich sensory input) is simpler/more efficient than specification down from generalizations ("decompression")
Generalization is selective lossy compression. Decompression requires reconstruction of the lost data, or requires that data is preserved or the knowledge how to extract it from external memory and incorporate it is kept. Here is the point of "precision vs depth", however I'll emphasize that sometimes a lot of both is needed.
Besides, I suspect depth has more severe limitations than precision for the lower levels. The cognitive hierarchy can add generalization levels at the expense of wider scope of input data and/or lower detail, but these both are problematic - if the scope is extended at the expense of detail (to keep computational complexity under control), then generalization levels will run out shortly, because there won't be meaningful details remaining. On the other hand, if the scope is extended with more modest detail lost, then learning will go computationally out of control.
We don't know where a machine can go in generalization levels with more computing power, but I think brain is very limited.
3. Higher level sensory inputs have less of impact over the cognitive hierarchy, because the feedback is less efficient than feed-forward.
That's the reason why captions "Smoking kills" or "Speeding kills" usually have no effect to make people stop smoking or stop speeding and why first-hand experience - to see with your own eyes, to hear with your own ears and to touch with your own fingers - have more dramatic effect in transmitting any message, than relayed experience of others.
Seeing your friend smoker dying of lung cancer or seeing your friend smashed in his car because he drove drunk - that's a pretty different sensory input - and not only because of your personal involvement with the sufferers.
Text is too abstract and distant - the low level physical representation of text is meaningless, it serves only to encode a higher level representation - that's where the input starts having a meaningful impact to the brain, and the message has to go down in the hierarchy to have actual impact on the behavior.
Another example is acting and film. Film as media demands providing motion pictures and rich sound - the physics of the action at the lowest level possible with a lot of details/high resolution of the input. If there's too much of a dialogue and too much of self-explanations and declarations by the characters, especially of obvious things - then brain is fed with high level generalized input, it can't activate the lower levels from the top-down, and they're idle or "bored". On the other hand if the lowest sensory input is rich, brain can induce up generalizations and engage the entire hierarchy. (Besides the effects of the balance of unpredictability etc., see Schmidhuber's works on Creativity)
4. Rationalization is playing with high-level patterns to explain lower level patterns, which the higher level cannot access, because of the fact that feedback is worse than feedforward, or because lower level patterns are unknown
Higher levels in the cognitive hierarchy are derivatives of the lower levels - the lower levels induce the higher ones, not vice verse. In a sense (a bit simplified), higher level patterns in the cognitive hierarchy are a delayed expressions of the lower level ones. However once a higher level emerges out of the stable regularities in the lower level, it starts to mess with the lower level business - adjusting lower level input, selecting data to keep attention on; adjusting coordinates, resolution, location; and the higher level does in order to maximize its own "success" - match, prediction, reward.
Higher levels usually cannot explain and trace back how they are created, what their lower level patterns are and what are the lower level drives.
Similar situation is with bad philosophy and other fields* where lower level conceptualization, patterns and input are wanted for a conceptual progress, but practitioners deny it and keep blah-blah-ing with concepts which are too high a level, too general, too unrelated to the problem they're trying to solve.
That's also one of the reason for researchers such as Boris Kazachenko and myself to suggest bottom-up approach - it allows for the maximum possible abstraction, while keeping maximum possible resolution and keeping the traces of the abstraction.
*Search the blog with "What's Wrong with Natural Language Processing
5. There are Two Reward Systems which are Messed Up
There's another issue - two reward systems run in parallel in brain. A cognitive and a physical. Cognitive system aims at maximizing predicted match of pure data, while the physical system aims at maximizing desired match - input sensations must match hardwired target sensations loaded with value - food, warmth, water, sex etc. The physical is way more primitive and crude, it relies a lot on the more primitive brain areas and on dopamine and other neurotransmitters/neuromodulators/hormones, while cognitive system is based on finer processing, even though the former participate also. Both systems interact and overlap, the physical system can override the cognitive and make it a slave - for example higher level cognition of drug addicts is a slave of the primitive need to take the drug. Generally these systems are messed up and tangled, so it's hard to trace where starts which in real behavioral record.
6. Rationalization is also explaining physical motivation with cognitive means
Ask somebody why she loves her boyfriend. She's likely to tell you "because he's smart, funny, kind, blah, blah and because he'so soooo blah!", while the real reason is much simpler - the way he makes her feel. That's why she loves him, where "to love" has also more basic meaning than the societal one - it's about quantities of neurotransmitters and about "imprinting" of addiction-cycles of generating such neurotransmitters by physical-cognitive conditioning, inter-association.
It's true that the abstract reasons do play certain role in creating the inter-associations between physical and cognitive sensations - everyone has some preferences and favorites, - however this can be reduced to:
- I love him because he's the type I wanted him to be!
- I love him, because he's my perfect match!
- He's the best match I could find so far...
This is a match between desired and input, which is the type of match of the physical reward system - apparently this selection is driven by the physical system, overriding the cognitive.
That's why I think the abstract reasons of "smart, funny..." are "rationalizations", but rationalization is not strange at all. Everybody would say "yeah, socially acceptable explanations", but that's a cheap answer.
There's one more appropriate reason for rationalization - it's the cognitive system which is asked the question (it's asked in a natural language); the highest levels in the cognitive hierarchy are ruling this area, and yes - the society has taught this higher cognitive system how it should act in such situations etc.
If you ask the question in a lower level language, the answer is different - that's her body language and her behavior when she's with her boyfriend alone, when they're kissing and making love. The answers are also in the amount of oxytocin, dopamine and other chemicals in her brain, the release of which is conditioned with the certain cognitive patterns initially generated by perceiving her beloved one.
In general, for love and attraction it's true by the definition of "emotions" that the physical reward system kicks in. One may be "just a friend" with someone because he's "smart, funny, blah-blah", but even then if one is a human being with a brain in tact (not a sociopath/psychopath), his friendship relations would be messed up with the physical reward systems - emotions, crude emission of certain chemicals and activations of primitive brain areas, which is associated/recorded/conditioned with cognitive patterns, and both are inter-twined.
Purely cognitive "friendship" is business and if it's such, it's not really a friendship.
*Higher/Lower level drives - Passionate feelings demand for "lower" drives, where "lower" in this context has different meaning than lower in the cognitive hierarchy. The meaning here is driven by evolutionary and physically "lower" brain modules, which maps to areas different than the neocortex, archicortex (hippocampus) and thalamus. However physical reward system is not in the same hierarchy as the cognitive system and the levels regarding tracing back, cognitive and physical reward systems are entangled and the physical system can manipulate and "short circuit" all levels of the cognitive hierarchy.
Continues... - On the apparent inconsistency of the goals of a system with cognitive hierarchy and More on rationalization and the confusions of the higher levels in the cognitive hierarchy.
Suggested reading:
Analysis of the meaning of a sentence, based on the knowledge base of an operational thinking machine. Reflections about the meaning and artificial intelligence - T.A. 2004
http://knol.google.com/k/cognitive-focus-generalist-vs-specialist-bias
http://knol.google.com/k/boris-kazachenko/executive-attention/27zxw65mxxlt7/11#
http://knol.google.com/k/intelligence-as-a-cognitive-algorithm
http://research.twenkid.com/agi_english/
Slides on T. Arnaudov's "Teenage Theory of Universe and Mind"
(C) T. Arnaudov 2011
15 коментара:
Todor,
> Higher-to-Lower level feedback is less efficient than Lower-To-Higher level feed-forward generalization.
"efficient" is a junk word (most are), all you have is "selected" (lossy) & "compressed" (lossless).
> Decompression requires reconstruction of the lost data,
"lost data" is lost, decompression is simply copying preserved compressed inputs into multiple lower-level coordinates.
> There's another issue - two reward systems run in parallel in brain. A cognitive and a physical.
This is very crude, purely "physical" motives don't matter much beyond infancy. Far more important is the intermediate stage: conditioned motives. These three levels of motivation loosely correspond to the levels in "triune brain" model: http://en.wikipedia.org/wiki/Triune_brain . The specifics of the model are obsolete, but it's still useful as a high-level taxonomy.
> Generally these systems are messed up and tangled, so it's hard to trace where starts which in real behavioral record.
They are "tangled" in a sense that a value (relative importance) of any concept is combined from all three systems, & people generally don't bother to untangle them.
> That's also one of the reason for researchers such as Boris Kazachenko and myself to suggest bottom-up approach
That "suggestion" is utterly ignored in this post.
> - it allows for the maximum possible abstraction, while keeping maximum possible resolution and keeping the traces of the abstraction.
Only first & last parts are true.
Hi Boris,
B: "efficient" is a junk word (most are), all you have is "selected" (lossy) & "compressed" (lossless).
T: OK about selected and compressed, but I don't think "efficient" is meaningless in the context I've used it & I think examples tell WIM. My AGI preferences use to be less lossy than yours, and also I think the best AGI will be a hybrid of general cognitive hierarchy and a computer (it is now - a human with a computer).
B: "lost data" is lost, decompression is simply copying preserved compressed inputs into multiple lower-level coordinates.
T: Sometimes yes, but sometimes it's preserved coordinates of locations/patterns of search of how to find the inputs which are lost or how to induce/record them again with higher resolution.
B: This is very crude, purely "physical" motives don't matter much beyond infancy. Far more important is the intermediate stage: conditioned motives. These three levels of motivation loosely correspond to the levels in "triune brain" model: http://en.wikipedia.org/wiki/Triune_brain . The specifics of the model are obsolete, but it's still useful as a high-level taxonomy.
T: I know of the "Triune Brain". BTW, do you know:
-- CITE -- BEGIN --
"Surprise! A unifying model of dorsal anterior cingulate function?"
Tobias Egner
Nature Neuroscience 14, 1219–1220 (2011) doi:10.1038/nn.2932
Published online 27 September 2011
Few brain regions' functions have been debated as intensely as those of the dorsal anterior cingulate cortex. A computational model now suggests that seemingly diverse cingulate responses may be explained by a single construct, 'negative surprise', which occurs when actions do not produce the expected outcome.
-- CITE -- END --
(My hypothesis from the "Guided evolution of ..." - newer brains actually do the older brains job and are supposed to have a lot of functional similarities. Newer brains intrinsically are "tweaks" of old ones, have significantly more computational resources/room for scaling complexity and extend older brain design to allow for higher resolution and scope.)
Regarding my "physical" - it's not literal mapping to brain, it's my dichotomy to cognitive, slides 20-22 in: http://research.twenkid.com/agi_english/Todor_Arnaudov_Theory_of_Hierarchical_Universal_Simulators_of_universes_Eng_MTR_3.pdf
B: Far more important is the intermediate stage: conditioned motives.
T: Conditioned motives and crystallized conditioned patterns are "physical" rewards in my dichotomy, they're conditioned because they had physical component, either forgotten or known.
-- Continues --
Boris: They are "tangled" in a sense that a value (relative importance) of any concept is combined from all three systems, & people generally don't bother to untangle them.
Value has many meanings, there can be conditioned motives and conditioned patterns, one of the entanglements is that crystallized conditioned motives/patterns may short-circuit the cognitive hierarchy at any cognitive level, some levels may be short-circuited (biased), others - may not.
Sometimes people just can't untangle it and go to psychotherapists to unveil their subconscious problems.
The difficulty comes partially because of the rationalizations, which is because of bad higher-to-lower level reflection ("worse efficiency") that I discussed - higher level plays with high level patterns.
"Little Albert experiment", phobias, and I suspect many apparently "meaningless" or "random" personal preferences might have roots in such type of conditioning at some point of live.
One of the assumptions in my "theory" in one of the POV was there's no "random" or "inconsistent" goals (a discussion in AGI list reminded me that I've discussed on their poblems) - there's wrong model of the goals, or/and insufficient resolution of the definitions of the model of the (virtual) control/causal units - ones which have rewards/goals and are building blocks of behavior.
B: That "suggestion" is utterly ignored in this post.
T: The topic of the article is in the title...
T: ... it allows for the maximum possible abstraction, while keeping maximum possible resolution and keeping the traces of the abstraction.
B: Only first & last parts are true.
T: Maximum possible resolution (MPR) given some constraints, and MPR at the lowest level of input/output.
-- END --
T: My AGI preferences use to be less lossy than yours,
It's meaningless unless you have a way to quantify it.
T:"seemingly diverse cingulate responses may be explained by a single construct, 'negative surprise',
It's implausible that such a large region can have a simple function.
T: Regarding my "physical" - it's not literal mapping to brain, it's my dichotomy to cognitive
You do need to map to the brain, - every region affects the value of a concept.
T: they're conditioned because they had physical component, either forgotten or known.
Original physiological values no longer matter, only the mechanism of association does. "Cognitive" patterns are also derived from "physical" inputs.
T: The topic of the article is in the title...
Right, the topic is "confusion"...
Todor:
T:>"seemingly diverse cingulate responses may be explained by a single construct, 'negative surprise',
B:> It's implausible that such a large region can have a simple function.
T: I thought you get the same criticism for your definition of intelligence.
T:> Regarding my "physical" - it's not literal mapping to brain, it's my dichotomy to cognitive
B:> You do need to map to the brain, - every region affects the value of a concept.
T: Yes, in brain there are projections from input to different regions, projections between different brain areas within neocortex and within different modules, projections from different brain areas to the motor output (motor pathway).
T:> they're conditioned because they had physical component, either forgotten or known.
B:> Original physiological values no longer matter, only the mechanism of association does. "Cognitive" patterns are also derived from "physical" inputs.
T: I explained what I mean with "physical" - it's not "nuclear physics", neither body-building, but a complementary to cognitive and related to control/causality impact initially originating from lower brains and related to maximization of desired match, rather than predicted match.
T: I thought you get the same criticism for your definition of intelligence.
B: OK, reacting to "negative surprise" really means more planning. That's what prefrontal cortex does too, & the difference might be in temporal scope, which would be person & situation -specific. As I speculated in "Executive Attention", ACC may mediate "primarization" via long-range feedback (especially through its spindle cells). But that mechanism is superfluous for AGI.
T: I explained what I mean with "physical"... maximization of desired match, rather than predicted match.
B: It's a POV, predicted match is also "desired". The difference is in where these "desires" came from: innate, vs. conditioned by coincidence, vs. recognized by comparison. And all this talk about "physical" neurotransmitters is pointless, - they mediate all of these processes.
B:> OK, reacting to "negative surprise" really means more planning. That's what prefrontal cortex does too, & the difference
might be in temporal scope, ...
T: Thanks for elaborating.
B:> which would be person & situation -specific. As I speculated in "Executive Attention", ACC may mediate "primarization" via long-range feedback (especially through its spindle cells). But that mechanism is superfluous for AGI.
T: I suspect (guess) such projections may be correlated with intuition or "presentiments". Cingulate cortex as a lower resolution version of the neocortex, more strongly connected with "physical" reward-system. (In which POV emotions are part of.)
I suspect hippocampus is supposed to do prediction/matching and planning either (trajectory, motion) if there was not a neocortex, starting with matching proprioception sensory data with other modalities and the "physical" rewards, and projecting trajectories.
T:> I explained what I mean with "physical"... maximization of desired match, rather than predicted match.
B:> It's a POV, predicted match is also "desired". The difference is in where these "desires" came from: innate, vs. conditioned by coincidence, vs. recognized by comparison.
T: My "desired" has different meaning than cognitive "desired" by definition. Conditioned by coincidence is coincidence of physical reward with cognitive pattern - the cognitive pattern turns "dirty" then - physical. Cognitive hierarchy however doesn't seem to have good reflection and it often looses track of the dirt, and some of the connections just don't have feedback, the very high level consciousness may analyze it and understand it.
The higher the level, the better reflection and tracking-back features, however there's a "rationalization catch" here - people think in high level terms and trace back their thought process and origins of it at this high level. They believe and are reassured this is the "truth", and since they cannot trace back (worse feedback than feedforward), they get easier stuck in "fluffing" than in failing to generalize to a certain level.
The ones who are more intelligent and are more creative in writing and story telling, movie directing, seem to me to have better feedback, so that they can understand and access the lower level processes and apply them better, so that stories are more fluent and realistic, involving more complex dynamics (the higher the level, the more inarticulate representations). The higher the level, the more "declarative" patterns.
B:>And all this talk about "physical" neurotransmitters is pointless, - they mediate all of these processes.
T: I admit it's not articulate enough, but not that it's pointless, e.g. it's about:
neurotransmitters and rations between them which switch brain to different emotional "Modes".
Brain Architecture Lecture - Slides 23-24.
I'm tired and distracted now, will continue on these topics later.
T: Conditioned by coincidence is coincidence of physical reward with cognitive pattern
B: It's a "response", not "reward" (another junk word). Responses are not just positive or negative, they include specific reaction patterns & emotional or physiological states. And a stimulus is conditioned by coincidence with a response-driving stimulus, not with a response itself. Originally, animals only developed senses that directly drove responses. The first such sense was "smell", initally on a single-cell level. Latter, they developed senses that *may* become conditioned to drive responses by basic coincidence: binary local presence | absence. This "divorce" of a stimulus from immediate response is a major advance, - the stimulus became neutral by default. Even later, they developed recognition by hierarchically & syntactically differentiated comparison. That's another advance of a similar order: now it is the *association* among stimuli that became neutral by default, - these differentiated patterns *may* become conditioned latter. These are similar steps of incremental abstraction of representation from action.
T: I'm tired and distracted now
B: As opposed to all other times... You are a mess, Todor, your thinking is a mess, & will stay that way until you sort out your priorities.
B:> It's a "response", not "reward" (another junk word).
T: Actually in my dichotomy and in the first slides I cited: http://research.twenkid.com/agi_english/Todor_Arnaudov_Theory_of_(...) and the complete words.
T: There are control-causality units defined, and physical rewards are inputs to such units which have predefined (more static) "goals" to match with input.
B:>Responses are not just positive or negative, they include specific reaction patterns & emotional or physiological states.
T: Of course, who said it is *at global level*. Reactions can be postive, negative (or none) in the micro-level. I explained in the AGI list (and in one old article on "meaning" cited in this one) that human goals are not inconsistent, and behavior model should be divided to little enough and many enough control-causality units, if modeled that way. Their interaction causes macro-behavior.
B:>And a stimulus is conditioned by coincidence with a response-driving stimulus, not with a response itself.
T: This is again a terminological discussion and taking sentences and terms out and fighting them. My view is expressed in a more clearly explained way in the article(s), and models will have unambiguous clear meaning if read in one sentence once they are implemented in a formal system. The dichotomy is in order to make the artificial "brain" having more clearly distinguished sub-systems.
B:> Originally, animals only developed senses that directly drove responses.
T: Yes.
B:> The first such sense was "smell",
T: OK, olfactory cortex.
B:> initally on a single-cell level.
T: Logical direction.
B:> Latter, they developed senses that *may* become conditioned to drive responses by basic coincidence: binary local presence | absence. This "divorce" of a stimulus from immediate response is a major advance, - the stimulus became neutral by default. Even later, they developed recognition by hierarchically & syntactically differentiated comparison. That's another advance of a similar order: now it is the *association* among stimuli that became neutral by default, - these differentiated patterns *may* become conditioned latter. These are similar steps of incremental abstraction of representation from action.
T: OK, do you really believe this is so new and enlightening to me I don't get it?
B:> As opposed to all other times... You are a mess, Todor, your thinking is a mess, & will stay that way until you sort out your priorities.
T: Sure it's supposed to get better, but I don't think it's that messy as you're trying to present it with out-of-context criticism of terms.
T: I think I use more words for stuff for which you use less, and I imply other stuff which you use more words, where I use less because it seems to me not relevant or obvious/logical derivative.
Edit/Add:
B:> It's a "response", not "reward" (another junk word).
Control-causality unit (CCU) has a "goal", which at the lowest level is matching an input value to a template - the simplest is binary, the more complex is by a continuous value).
In the POV/philosophy I use, "reward" at micro-level is reinforcing/messaging a control-causality unit that it's doing what it's designed for; it's let know that it's located close its "purpose of existence" and it aims to keep staying there.
"Negative" reward, "punishment" is data which pushes a micro CCU away from it current coordinates, it notifies it that the input is far from the goal one, so the CCU needs to make a correction.
There's competition between CCUs, etc. These are micro-units, emotions or complex reactions are chains of interactions of "satisfaction", saturation, competition between many micro CCUs for long period of time.
"Emotions" and particular basic physical reactions are ensembles of CCUs and specific connections to actuators. Some of the CCUs are connected to lowest sensory inputs.
Etc.
T: OK, do you really believe this is so new and enlightening to me I don't get it?
B: It should be enlightening because:
- I got rid of the junk. That's what generalization is all about, right? Your terms & statements are ambiguous, misleading, redundant, or just plain contrary to the meanings you latter ascribe to them.
- I did incremental bottom-up analysis. True, it's just a version of what I have in meta-evolution knol, but that didn't seem to get through to you.
- I showed that conditioning is as removed from response ("physical" is just another junk word) as cognition is from conditioning. The action - representation dichotomy is meaningless until you analyze specific mechanisms that translate the later into the former.
T: Here I'm not bidding for your prizes or fitting to your theory, and I'm using my terms, not yours, and more disambiguation of their meaning and that POV is given in my works, as well, which you don't know, because "they are junk". Discussion on them is meaningless, you assume you understand all of my confusions, perhaps because you assume I'm always fitting to your terms and your definitions.
B>The action - representation dichotomy is meaningless until you analyze specific mechanisms that translate the later into the former.
T: It's not meaningless, there are "" in more "declarative" for a reason, it means the higher level patterns get more ambiguous in terms of final feedback to lowest output, and what I mean is disambiguated in the works.
T: you assume you understand all of my confusions, perhaps because you assume I'm always fitting to your terms and your definitions.
B: No, I only assume that you are speaking English.
I didn't mean to question your knowledge of English. It's just that your raw knowledge doesn't seem to translate into conceptual cohesion. That'll probably take a lot more time than you can stay focused for.
T: English is quite an ambiguous language, I don't claim I'm using the right nuances in all the cases and that my definitions are clear enough if you don't have my intuitive understanding (they're not); and my English is supposed to be less "native-like" than yours, the misunderstandings are not surprising.
The original definitions of mine of which this translations are derivatives are in works in Bulgarian, and I often use many words to explain one term in Bulgarian.
A short presentation of a little part of the meaning of that physical "reward" is there, also 20-22.
Slides for my theory, starting from slide 11, stop as later as you wish.
The dichotomy is in order to integrate "physical" CCU and cognitive CCU.
B:> It's just that your raw knowledge doesn't seem to translate into conceptual cohesion. That'll probably take a lot more time than you can stay focused for.
T: I don't know, I'm a not generalist, my POV of the concepts is different than yours.
I'll probably focus on creating fully explicit and formal model as soon as I start designing a real AGI prototype having cognitive and physical reward-systems, where "reward system" is what I've defined in the works.
Post a Comment