that’s exactly the purpose of my first post hoping to get some expertise to help me to refine and simply my definition.
I am curious on that. Let’s assume (since I haven’t gotten much consensus yet) the high level definition of intelligence is “learn” and “decide”, I think it would be super interesting already if I can build a software simulation of the brain and attach it with a very low resolution of visual sensory (10x10?) and audio sensory component, and that I can repeatedly teach the brain by drawing a circle and say “circle”. if I draw a circle on the visual sensory and the audio sensory will reflect the “circle” sound representation and vice versa, as a newbie I will be jumping up and down already. And without changing anything and just letting the simulation to grow itself (the same way it does biologically) that it can also learn square or other pattern and reflect the same that will be super super exciting. But I recognize this is not a easy task because I will need to understand how the brain grow and prune itself. But just thinking about super excites me already. And I can’t get that out of my head. But I don’t necessary want to think too much of a specific case otherwise I will shoehorn myself into doing non general intelligence. That’s why i want to focus on the definition first.
Then a crazy question since it is based on fiction. Not sure if you have seen the movie “Her” by Spike Jonze. Do you consider the character Samantha has intelligence? If so, she does not have any moving parts at all. Then can we conclude that motor skill is not part of the fundamental list?
Haven’t seen the movie but, from your question, I assume the director depicted “her” as intelligent.
To me, there is a more fundamental question than “is she intelligent” if we want to consider building a brain in silico. That question is: “is it conceivable that we can build a non-moving AI which qualifies”.
Imho, moving is maybe too strong a requirement. But interacting (in any way) with environment (rather than passively sensing) is already something which some people would claim necessary. To succeed at all.
In my view, “her” does speak, so… okay, she “could” qualify.
I’m not merely viewing these interactions here as necessary to “judge” externally if she is. Possibly necessary for her to develop intelligence in the first place. At all.
Continuing with this view, about your example of a born-blind, senseless and motionless individual, I consider it quite a bold statement already that we’d all agree on assuming intelligence (at least if the born-as-such hypothesis applies in all cases). In the case where there would never have been any significant input to the brain, nor any interactive loop happening, there are people here which would argue against this brain’s capacity to develop in any meaningful way.
I’m among the people who’d add “Ability to interact” to your checklist. It is possible that perception alone won’t cut it.
But since no one reached strong AI just yet, this is all theory.
Note that I’m for adding that requirement for early developmental phases. Once anything meaningful was learned by means of interacting, yes, I believe the brains of “locked-ins” can still function from remaining forms of perception, or imagination alone. But it would be hard in that state to learn anything brand new, imho.
Highly recommend it even if it is not in the intelligence context.
This is an interesting term and I am quite honestly struggling with that. While I am re-reading Jeff’s book for the third time (yes, slow learner), I started to pick up phrases to strengthen my understanding from the first two times (and it could form a bias). Related to “Ability to interact”, this is the excerpt from “On Intelligence” that causes me to struggle:
“Behaviour is a manifestation of intelligence, but not the central characteristics or primary definition of being intelligent”
“You can be intelligent just lying in the dark, thinking and understanding.”
If we agree with Jeff (I do, but I could be bias or I could misunderstand), then “lying in the dark” means motor control is not a “primary definition of being intelligent” and to a certain extent, motor control is just a behaviour.
For the same reason, “ability to interact” seems to be also a behaviour (I could be wrong, but I feel very strongly about this interpretation and that causes me to struggle)
But if we break down “ability to interact” to more on the fundamental level “ability to decide” then yes I would agree “ability to decide” is a fundamental element of intelligence. But the word interact conveys strongly about behaviour.
I would highly disagree with this. Such person can for sure learn language (which linguistic skill requires high level intelligence to achieve. but at the same time don’t want this human intelligence to taint the discussion since I am not building human brain as a starter) just through the auditory input.
But definitely want to see if everyone agrees with Jeff’s statement about lying in the bed since I think that is crucial.
I was thrilled by “On Intelligence”. And you’re welcome to agree with all of it. Personally I’ve come to question that interaction thing at least.
Jeff himself is currently researching about all sorts of “sensorimotor” models. Terms says it all: sense & act.
We have machine learning already (Deep Learning and all) which does not interact. It can classify stuff. It can learn to classify stuff. Yay! It does not think by itself in the dark, though. It is a quite slow learner, poor generalizer, and would have a very hard time learning about brand new topics.
Some will judge success towards “intelligence” (strong AI?) based on swiftness to learn new skills, generalization abilities, creativity, etc… which you can put in a list, but the thing is… artificial brains did not achieve good scores at those.
Now, maybe from that realization, some argue that the missing ingredient could be “interactions”.
And here we are now, exploring this vast space of possible moving/sensorimotor/interacting models
Not sure it would be that easy.
In one of your examples above, you proposed to tie the sound of the spoken name for something with another form of perception, to learn about a circle. One without the other possibly does not go very far towards learning language. Or shapes, for that matter. And I’m still unsure we can really learn to speak without interacting with our peers. Deep Learning (again) may learn to synthesize speech, already. But it does not have the semantic association with it, required to qualify for having mastered the language.
Please don’t let a fictional movie distort your understanding of how the real world works. As far as computer speech production, many of the motor tasks to produce sounds have an analog of the human process to produce natural-sounding speech. That said, a speech generator is usually a specialist program that is driven by machine learning approaches that do not use what one would normally consider intelligence.
This is a common misconception about the nature of speech production. It is first and foremost a learned motor task.
In other areas of the forum we have explored delayed production of mating songs in finches. The model we have converged on is that we hear and learn speech sounds first. Then we sequentially learn to control the mechanism of sound production and then to use this new skill to train ourselves to reproduce these learned sounds.
Many of the behaviors that we express that are considered parts of human intelligence are in fact learned during the acquisition of language. These language linked behavior form a long list but they include things like object naming, imagination, play, some mental manipulations, and theory of mind.
Piaget’s theory of cognitive development works through the stages of bootstrapping human-level intelligence. Note that the earliest parts are the Sensorimotor stage as has been mentioned by others in this thread. You asked about the basic intelligence foundations and most of the things we call intelligence are built by combining these basic building blocks of running a body into the more advanced things we generally recognize as intelligence. By these latter stages, we are very far away from the basic building blocks. As has also been stated before - these early stages are often bypassed as uninteresting.
I would strongly suggest that you reconcile your laundry list against the early stages of development. Note that these boring bits have to be firmly in place long before you add things like symbol manipulation. The memory mechanisms that are the basics of a brain are strongly tied to the sensory-motor learning that is used to run a body. Understanding this should shape your original list of:
Perceiving Information
A starting point triggered by sensory input or “awareness”
Memorizing Information
In volatile memory then selectively to non-volatile memory
De-memorizing Information (recall? forgetting?)
In non-volatile memory (1, 2, & 3 = short and long term memorization and recall, all in the same mechanism!)
Predicting Information
When you let go of your common bias of what intelligence is or is not and get to the point were you can conceptualize that this does not work anything like a symbolic computer program you will be in the right place to understand how HTM (cortical column computation) organically performs the underpinnings of your laundry list.
It’s already not a simple case and you have to decide what approach you are going to use, because vision works completely differently in, let’s say, bacteria, fish, and rat.
The simplest example of something with the ability to “learn” and “decide”, would be just a binary switch with the programmable reaction for the input.
I wouldn’t use the word “distort”. Since no one in the world seems to know for sure how the brain works, or even what intelligence is, anything including fictional movie would become an inspiration. And often in difficult task, inspiration is quite necessary.
I will take your advice to heart. And I will study the links you provided. Thank you very much for your help. It has been very informative.
I want to stress that I am not saying by defining intelligence is the only step. Similar to what Chris Eliasmith suggested, the What and How have to be discussed together. Machine learning is just statistics with zero biological reference. It is absolutely necessary to find out how the biological brain able to achieve each fundamental list of elements of intelligence since that is my whole goal.
Could you elaborate more on this? Particularly around the word “programmable” reaction. The word “programmable” sounds against the idea of intelligence.
Just did a quick read on the links, those theories seem to be very human specific. Not sure at this stage I would like to get distracted by the human ones since human intelligence is multiple magnitudes higher than other form of intelligence. But thanks for the links. I will still study them.
Yes, my links have moved far past the basic steps as I can see how they build to more intelligent behavior. I can see how the basic building blocks are combined to do more interesting things. I don’t feel that I need a strict test to see that is or is not intelligent. If my model exhibits the behaviors I am looking for I am satisfied that it is working.
To your original question of defining intelligence this paper gives some good ways to think about the problem:
I didn’t say hard-coded, programmable by environment == trained.
But the hard-coded option is not against the idea of intelligence either. The intelligence of many simple organisms is hard-coded, the adaptation part is external for them. It can be evolution, it can be your own guidance, for an organism’s intelligence it doesn’t make any difference.
The evolution itself definitely can be considered as a form of learning, but it works only at the level of population and involves environment.
However, from the perspective of a bacteria, or a plant, its intelligence is hard-coded.
Plus, our day we create genetically modified organisms, which a product of engineering. So, evolution can be substituted, it doesn’t make much difference at the level of individual intelligence.
A wrap-up from my previous post to summarize what I have learned from others before I bring up the next question for my brain building journey.
Thanks for over 90 replies I have learned a lot especially from the paper published by Francois Chollet. And I have revised my “Define Intelligence” as follows:
Objective
Learning from Francois Chollet’s paper and also first hand from the replies to my post, the objective now is not trying to get everyone to agree what intelligence is. But instead the objective is to have a fundamental list of elements to satisfy the most basic form of intelligence as a guideline for me to build a simulation for the starter brain. The simulation of the starter brain can form the basis to evolve to a more complicated one in the future with eventual goal to match the human intelligence. But for now, the full focus is on building the simplest form first.
Define Intelligence
Based on various feedback and a lot of study, my conclusion on the definition of intelligence for the objective on building the simplest software simulation of a brain is the ability to learn, adapt, and decide in extreme generation.
In order to establish a more refined guidelines for the development, I further break the definition down to the following fundamental list of abilities:
Receives real-time input from sensory inputs
Forms memories (Learn)
Derives new associative memories (Learn)
Performs prediction and compares with incoming inputs. (Learn, Adapt)
a) When prediction is the same as the incoming input, strengthens existing memories
b) When prediction is different than the incoming input, weakens existing memories and forms new memories based on the incoming input
Makes decision based on the strength of the consensus of the existing memories
I used memories in plural to emphasize similar types of inputs will be recorded as multiple memories.
I think for now this list is a decent guideline for myself. I will prepare the next question on brain building to focus on the basic biological element as the building block for the simulation.
Thanks everyone for your help. With my first post hitting over 90 replies is really encouraging and I am very confident I can get a lot of help in this forum. Thanks!
I am not entirely sure if the memory which based on my understanding is represented in the synapse will be totally degenerated. But I include this “weakens existing memories” in the learning and adapting process.
