I’m trying to find as succinct as possible definition of “intelligence.”
an intelligent memory structure has built into it functions that create reinforce it’s ability to assimilate information better and better by being able to interpret new data in the light of old data. Vague I know, but it’s a start?
Let me know your thoughts on the following observations:
Is the essence of intelligence nothing more than generalized transfer learning?
And is “generalized transfer learning” really nothing more than outsourcing all decisions around the association of patterns in the environment to the environment itself?
For me, these “behaviouristic” definitions of intelligence miss something. Isn’t intelligence somehow more flexible than mere stimulus-response? (Maybe @bitking’s “choosing the best” has something to do with this?)
I would say that intelligence is: The ability to flexibly adapt perception and behaviour to changing circumstances and to extrapolate from known situations to unfamiliar ones.
I base the definition on main features of any trial and error learning system, and “scientific method” we have been using since we were born to develop models for how things in the world around us work.
I think there’s a nuance in my definition of generalized transfer learning here that isn’t clear enough. I used the term decision, but I’m actually not talking about an organisms behavior at all.
What I’m talking about is the allocation of attention. How should an organism attend to patterns and thereby associate them in memory?
The intelligent answer to that is the most meta answer, the most potentially dynamic. And where does the answer come from? In part it comes from the memory structure already built and the most meta answer it gives is, “you should attend to whatever the environment says to attend to according to previously constructed memory, that is, previously associated patterns.”
So when I use the term, “decision” I’m referring to the brains decision in how it should make new memory which informs it’s behavioral decisions, but is actually previous to them.
Thus I can’t call my model a behavioral model because that isn’t what I’m describing, nor is that how the model is evaluated.
Sorry for splitting hairs, it just sounded like i hadn’t been clear enough. If you’ll indulge me a little longer I think I’ve got another ramble coming on.
I see things like the brain or any hypothetical intelligence as a layering of structure and algorithm; of memory and innate implications on how that memory should be changed (computation).
For instance chemistry has certain rules, indeed, chemistry is nothing more than a set of observed and assumed rules which things of a very tiny size adhere to. Whatever those rules are, since our brains are made of cells and cells have tiny molecular structures, that is to say they have functional moving parts at the molecular level, then the structure of chemistry has a guiding and constraining effect on the operation of the brain, on how it’s more macro structure evolves, changes, moves, mutates, even ‘computes’ over time.
This is an analogy and example of the repeating pattern in intelligent structures: the layering of structure and structures embedded algorithms: the structure itself is the source of, has implicitly embedded in it, the rules of change of that very same structure. Wild.
So moving up a level of structure, from the tiny to the small we reach the level of cortical columns and that structures’ effect on its own mutation; the general algorithm of the neocortex which HTM tries to make explicit.
Somehow, through biological evolution - largely informed by the structure of chemistry as bigger patterns (organisms) interacted with each other in the world - (sorry for continually defining common terms but I’m trying to describe the form-is-function framework and how it relates to intelligence as we go) somehow evolution changed a memorial structure, evolved a better cortex, created a form of repeated structure that is the function of some kind of more generalizable, more programmable computation (rather than the a slow computation of evolutionarily iteration).
So then the change of this structure is the computation that attempts to answer the question, “how should this structure change?”
It’s as if you could ask God any question and you asked, “which question should I choose?” The neocortex has an element of default for what to do with new sensory data but that default tends towards producing structure that leverages patterns in the data itself to know how to encode memory. The grand insight of HTM is to recognize some of the default mechanisms by which that tendency is resolved, mainly through prediction. Thus it naturally associates patterns occurring at nearly the same time together.
Consider the computer, the instructions for how to do computation is not integrated into the CPU, rather they are separated. Not so in the brain. the feedback loop to determine how structure should change the future is much tighter. In the computer changes to the hardware are determined by how we use it. The feedback mechanism goes through us, our economic structure and our engineers. There is an intelligent evolution of hardware, of structure, but it’s a large strange loop instead of a small one.
Yet in the computer changes in memory, which are changes in structure, can be programmed to be anything, save entirely self referential. The code doesn’t change when the compute memory changes, they are separated.
In the brain the memory structure is the computer. That means the data that flows through it, and itself, both working together, become the ‘code that computes’. And that so called ‘code’ changes it’s very structure for how to ‘execute code’ in the future.
So then in this context, when there is nothing other than patterns, is not the association of patterns with each other the very definition of ‘memorial code execution’? Is not that all that the structure technically is and technically is doing?
It’s such a strange dichotomy because “form is function” really just means the feedback loops are tight. The only way to describe changes in form, algorithm are to use structure, memory, and the only way to make structure, memory is to make changes according to some embedded design, some set of inherent-to-the-structure-itself instructions. It’s like how matter is made of energy - well if course it is, what else is there to make matter out of?
Intelligence wants to know what structure is like what other kind of structure, it wants to associate patterns together, but not all patterns that look similar are similar, and not all patterns that look different are different. So the intelligence needs to ask the question, “how can i tell what patterns belong next to what other patterns?” To answer that it observes changes in patterns over time and evaluates it’s pattern association maps, inherent to the way I’m which patterns are associated in memory, through the algorithm of prediction. Then it is faced with a deeper problem, “how should those implicit maps for associating patterns change?” To answer this it repeats itself into hierarchies. And forms patterns within the context of other patterns. Maps of pattern association become patterns themselves to be associated. And you get this growing structure were everything in it learns incrementally what it should be in relation to it’s surrounding structure. It learns incrementally as in it has discovered generalized transfer learning. According to it’s particular environment, it had developed the perfect form of analogous structures and approaches the perfect meta form to describe how those analogous structures should change over time and according to it’s particular meta environment.
I should have mentioned that the earlier shown circuit is representative of coexisting emergent systems. The same model works at multiple levels. If the molecular (RNA world) level system is working properly then cells should on their own emerge, followed by these cells on their own learning how to develop into increasingly complex intelligent multicellular systems as happened during the “Cambrian Explosion”.
The explanation below was written to at the same time help answer big-questions the general public are most interested in, while using the phrase “intelligent causation” in scientific context needed for the biological scientific theory I had, before the ID movement made such a phrase controversial. It would be easier for me to avoid the issue by renaming things using next best words, but for the sake of those who only wanted to see what happens when a hypothesis for “intelligent cause” is objectively developed into a scientific theory it’s at least in part a personal victory, instead of shame for having seen no harm to science and science education in at least trying. To be useful to those who do not want or are ready for HTM theory scale detail I found that the core model needs to start with David Heiserman basics that for good reason have long been of fascination to robot builders of all ages, and ~25 years later still best applies at all fundamental levels of biology. In this case what’s most important are the very first neural self-learning brain circuits, not what 600 or so million years later developed. There is then a high school appropriate model and theory with a very real “It’s ALIVE!” moment for young scientists who can later at home successfully code/create their first virtual critter. Being able to go from there on into modeling the (assumed unintelligent) behavior of matter “which does not necessarily need to be intelligent to be the fundamental source of consciousness” keeps that scientific and for some religious mystery going on into the biggest of big questions including those relating to afterlife. This makes the model and theory inherently “faith friendly” enough to be welcomed change, better that than nothing. To also be as precise as possible to all readers and stay in spirit with what started in Kansas the terminology is the same as it has always been even though this puts the model and theory in context of theory that according to Wikipedia is “pseudoscience”. So I hope that this long paragraph disclaimer made it clear that I do not represent what gave itself a bad reputation. In my case there is a model where the illustrations for it would remain the same regardless of words used to name things, where you only have to not mind the paradox.
Reciprocal cause/causation between levels goes in both the forward and reverse direction. These communicative behavioral pathways cause all of our complex intelligence related behaviors to connect back to the behavior of matter, which does not necessarily need to be intelligent to be the fundamental source of consciousness.
(1) Molecular Level Intelligence: Behavior of matter causes self-assembly of molecular systems that in time become molecular level intelligence, where biological RNA and DNA memory systems learn over time by replication of their accumulated genetic knowledge through a lineage of successive offspring. This intelligence level controls basic growth and division of our cells, is a primary source of our instinctual behaviors, and causes molecular level social differentiation (i.e. speciation).
(2) Cellular Level Intelligence: Molecular level intelligence is the intelligent cause of cellular level intelligence. This intelligence level controls moment to moment cellular responses such as locomotion/migration and cellular level social differentiation (i.e. neural plasticity). At our conception we were only at the cellular intelligence level. Two molecular intelligence systems (egg and sperm) which are on their own unable to self-replicate combined into a single self-replicating cell, a zygote. The zygote then divided to become a colony of cells, an embryo. Later during fetal development we made it to the multicellular intelligence level which requires a self-learning neural brain to control motor muscle movements (also sweat gland motor muscles).
(3) Multicellular Level Intelligence: Cellular level intelligence is the intelligent cause of multicellular level intelligence. In this case a multicellular body is controlled by a brain made of cells, expressing all three intelligence levels at once, which results in our complex and powerful paternal (fatherly), maternal (motherly) and other behaviors. This intelligence level controls our moment to moment multicellular responses, locomotion/migration and multicellular level social differentiation (i.e. occupation). Successful designs remain in the biosphere’s interconnected collective (RNA/DNA) memory to help keep going the billions year old cycle of life, where in our case not all individuals must reproduce for the human lineage to benefit from all in society.
The combined knowledge and behavior of all three intelligence levels guides spawning salmon of both sexes on long perilous migrations to where they were born and may choose to stay to defend their nests “till death do they part” from not being able to survive for long in freshwater conditions. Motherly alligators and crocodiles gently carry their well guarded hatchlings to the water, and their fathers will learn to not eat the food she gathers for them. If the babies are scared then they will call and she will be quick to come to their aid and let them ride on her head and body, as they learn what they need to know to succeed in life. For humans this instinctual and learned knowledge has through time guided us towards marriage ceremonies to ask for “blessing” from a conscious part of us that our multicellular intelligence level (brain) may be able to sense coming from the other intelligence levels we cannot directly experience, which at the genetic intelligence level has for billions of years been alive, and is now still alive inside of us…
This is how the (origin of life from behavior of matter) interaction relates to chemical/molecular and biological speciation, which on rare occasion adds another intelligent behavior level similarly able to learn by trial and error:
All of the behavior levels between the molecular level system and environment are at work at the same time. When our cellular level is busy battling invading cells “we” at the multicellular level feel ill and may respond by staying home from work or school instead of usual activities. We also all started as a single cell zygote that takes nine months to develop the mind of a newborn, which further changes over time to become experimentative then flirtatious then usually develops an adult mind needed to lovingly raise a newborn.
I agree that to be complete a model for human intelligence requires the molecular level detail you mentioned. Where successful you’re then in a realm that has religious implications and how things are defined can make all the difference in the world in regards to how useful and well received it is by the average person.
In my case I ended up where a problem pertaining to something “intelligent” was waiting to be solved, where you would probably have to be there to sense what was needed. It seemed that this was a good time for me to explain what I came up with, even though that task is not easy.