Comment of mine on the article "AI Is Harder Than We Think: 4 Key Fallacies in AI Research" https://singularityhub.com/2021/05/06/to-advance-ai-we-need-to-better-understand-human-intelligence-and-address-these-4-fallacies/
Posted on Real AGI FB group
The suggested fallacies are:
1. Progress in narrow intelligence is progress towards general intelligence
2. What’s easy for humans should be easy for machines
3. Human language can describe machine intelligence
4. Intelligence is all in our heads
(See also the article)
The title reminded me of a conclusion of the "AGI Digest" letters series where after giving the arguments I noted that: "AGI is way simpler than it seems". See the message from 27.4.2012 in "General algorithms or General Programs", find the link to the paper here:
https://artificial-mind.blogspot.com/2017/12/capsnet-capsules-and-CogAlg-3D-reconstruction.html https://artificial-mind.blogspot.com/2021/01/capsnet-we-can-do-it-with-3d-point-clouds.html.html
Summary
In brief: I claim
it is the opposite: AI is easier than it seems (if one doesn't unerstand
it and confuses herself, it's hard, right). Embodiment is well known
and it lays in the reference frames and exploration-causation, stability
of the coordinates and shapes and actions, repetitiveness etc. not in
the specific "material" substrate of the body. The "easy for humans..."
is well known and banal, also the point against machines is funny: in
fact humans also can't "apply their knowledge in new settings without
training" (see the challenges in the article) etc. IMO progress in
"narrow" AI actually is a progress towards AGI and it was so even in the
2000s, as current "narrow AI" ML methods are pretty general and
multi-modal and they give instruments to do processes which were
attached to "AGI" at least since the early 2000s, such as general prediction
and creation, synthesis. Current "narrow AI" does Analysis and
Synthesis, but not "generally enough in a big enough and "integrated
enough" and "engine-like-running" framework which connects all the branches, modalities and knowledge together, however the
branches and "strings" are getting closer. Practically, one can use as
many "narrow" NN with whatever glue code and other logic in a system.
Discussion
1. "Progress in narrow intelligence is progress towards general intelligence" [are not progress towards GI]
— IMO it actually is a
progress, because the methods of the "narrow" become more and more
general, both in what they solve and in the ambitions of the authors of these solutions. After a problem or a domain is considered "solved" to one degree or another, the intelligent beings direct themselves to another one, or expand the range, or try to generalise their solutions of several problems so far and combine them etc.
One of the introductory lectures in the first university course
in AGI back in April 2010, which I authored, was called "Survey of the Classical and Narrow AI: Why it is Limited and Why it Failed [to achieve AGI]?": http://research.twenkid.com/agi/2010/Narrow_AI_Review_Why_Failed_MTR.pdf
While
wrapping up the faults as I saw them, one of the final slides and
others in the lecture, matched with one of the main message of the
course - hierarchical prediction and generalisation, - suggested
that the methods of the advanced "narrow AI" actually converge to the
ideas and methods of AGI. Even image and video compression for example
share the core ideas of AGI as a general sensory-motor prediction
engine, so MPEG, MPEG2, H264 - these algorithms in fact are "AI".
"Motion compensation", the most basic comparison, is related to some of
the primary processings in the AGI algorithm CogAlg, all
"edge-detections" etc. are something where any algorithm searching for
shapes would start or reach in one way or another. Compression - finding
matches ("patterns), which is also "optimisation" - reducing space etc.
Two of the concluding slides (translation follows):
"The
winner of DARPA Urban Challenge in 2007 uses a hierarchical control
system with multi-layer planing of the motions, a behavior generator,
sensory perceptions, modeling of the world and mechatronics".Points of a short summary, circa early 2010:
What's wrong with NLP? (from articles from 2009) [and "week" AI]:
* The systems are static, require a lot of manual work and intervention and do not scale
* Specialized "tricks" instead of universal (geneal purpose) systems
* Work at a very high symbolic level and lack grounding on primary perceptions and interactions with the environment
*
The neural networks lack a holistic architecture, do not self-organize
and are chaotic and heavy.
Overall: A good "physics" is lacking, one that would allow creation of
an "engine", which to be turned on and then to start working on its own.
The systems are instruments and not engines.
Note, 7.2021:
The point regarding the NN however can be adjusted:
Many NN can be
stacked or connected with anything else in any kind of network or a
more complex system - we are not limited to use one or not use any glue
code or whatever. The NN and transformers are actually "general" in what
they do and are respectively applied for all sensory modalities and
also multi-modaly.
Complete or powerful enough for a complex
simulated/real world sensory-motor multi-modal frameworks are not good
enough and these algorithms may be not the fastest to find the
correlations and have unnecessary brute force search which can be
reduced by more clever algorithms (and they should), however these
models do find general correlations in input.
2. "What’s easy for humans should be easy for machines"
— Isn't
that banal, also it is vague (easy/hard). Actually some of the skills
of the 3 or 4 years old are achieved by 3 or 4 years long training in
supervised settings: humans do not learn "unsupervised" except basic
vision/low level physical and sensual stuff (language learning is
supervised as well; reading and writing: even more).
Test
people who didn't attend school at all, check how good they are in logic
for example, in abstract thinking, in finding the essential features of
the objects or concepts etc. Even people who have university degrees
could be bad in that, especially BAs.
There are no machine learning models with current technology from the
"narrow AI" which are trained for that long yet, an year or years with current compute. We don't know what they
could achive, even with todays' resources.
On learning and generalising:
"If, for example, they touch a pot on the stove and burn a finger,
they’ll understand that the burn was caused by the pot being hot, not by
it being round or silver. To humans this is basic common sense, but algorithms have a hard time
making causal inferences, especially without a large dataset or in a
different context than the one they were trained in."
That's right about training if you use a very dumb RL algorithm
(like the ones which played for 99999 hours in order to learn to play
the basic games on Atari 2600), however overall the "hardness" of
learning this by a machine is deeply wrong and not aware of what the
actual solution could simply be:
"An algorithm" would
have sensors for temperature which will detect "pain", caused be
excessive heat/temperature, which happened at the moment when the
coordinates of the sensor (the finger) matched coordinates within the
plate of the stove. Also, it could have infrared sensors or detect the
increment of the temperature before touching and detecting that there is
a gradient of the measurement. The images of the stove when the finger
was away didn't cause pain, only the touch. This is not hard "for an
algorithm", it's trivial.
4. Intelligence is all in our heads
— Wasn't that clear at least since 20 years? (for me it was always clear) However, taking into account, that the embodiment can be "simulated", "virtual". The key in embodiment are the sensory matrices, coordinates ("frames of reference" in Hawkins' terms) and the capability to systematically explore: cause and perceive/study the world; the specific expressions of the sensory matrices and coordinates could vary.
3. Human language can describe machine intelligence
"Even “learning” is a misnomer, Mitchell says, because if a machine truly “learned” a new skill, it would be able to apply that skill in different settings"
+
1. "a non-language-related skill with no training would signal general intelligence"
— I challenge these "intellectuals": can you make a proper one-hand backhand with a tennis racket with "no training"? (Also how long will you train, especially before delivering a proper over-the-head service with good speed or a backhand, while you are facing back to the net; or a tweener (between the legs shot, especially while running back to the base line etc.)
You're not supposed to need explicit training, right? You did move your hands, arms, wrists, elbows; legs, feet... You've watched tennis at least once on TV sports new, therefore you should be able to just go and play against Federer and be on par with him, right?. If you can't do that even against a 10-year old player, that means "you can't apply your knowledge in new settings"...
Can you even juggle 3 balls: by "applying your knowledge of physics from school and sense of rhytm from listening to music and dance - even the simplest trick.
Can you play the simplest songs on a piano : by applying your understanding of space and motion of the hand and find the correlations with the pressing of the keys and the sound of each of them etc. - can you do it especially if you lack musical talent.
Well, "therefore you lack GI", given your own definitons... I'm sorry about that...
(In fact the above is true for many humans; humans really lack "general intelligence" by some of the high-bar definitions which a machine is expected to met before being "recognized")
...
Слайдът на български (4.2010):
* Какво не е наред в обработката на естествен език [и слабия ИИ]?
●Системите са статични, изискват много ръчна намеса, не се развиват и мащабират.
●Специализирани „трикове“, а не универсални системи.
●Работят на високо символно ниво и нямат основа от първични възприятия и взаимодействия със средата.
●Невронните мрежи нямат цялостна архитектура, не се самоорганизират и са хаотични и тежки.Липсва добра „физика“, която да позволи създаването на „двигател“, който дасе включи и да заработи от самосебе си. Инструменти, а не двигатели.
