AI

[sciwiz12]

I'm studying it right now, let's talk about it. I'm still just getting started in my readings really, but I feel at least a little capable of engaging in a dialogue on the subject.

 

Firstly I know enough to know I'm not in the camp that thinks we should be modeling the hardware and software after the human brain. It seems to me that the technology to create highly intelligent evolving computer consciousness is actually closer to being a reality than what would be needed to emulate the human brain.

 

A lot of my computer science knowledge is also heavily tied to game design which is my current major, so when I think of procedural generation I am thinking about procedural content generation techniques using algorithms to create content for games and cross applying the knowledge to the generation of evolving computerised consciousnesses, although I suppose the same techniques exist wholly outside of game design and development.

 

I have two frameworks for thinking about the issue, one is the problem of understanding the meaning behind words and the other is the problem of understanding the self. Essentially what I'm trying to get at is that on some level we as humans have this perception of a ghost within the machine, an observer within us that on the one hand is created by and exists as a construct of the brain and yet seems to our point of view to be somehow removed from the physical.

 

I'm going to start by considering a model of human consciousness within which a man's personality and sense of consciousness are firstly separate and secondly closely related. I then also propose that within this model both are simulations created by the brain to fulfil useful filtering functions. I will also refer to the personality as the identity and the conscious awareness as the mind. In programmatic terms I suspect the identity could best be represented and constructed as a storehouse of information. In other words the identity itself would be represented simply as a collection of information about the characteristics one learns to associate with one's sense of identity. That information is later pulled by various other parts of the brain when making calculations or decisions.

 

For example let's take Jack. Jack is told that he is handsome, perhaps reasons that he is smart, and decides through experience and sensory/perceptual feedback that he likes chicken. Later when Jack hears a compliment he accesses his Center of identity containing information about himself, pulls his sense of self image, recalls that he identifies himself as handsome, compares the compliment to this recovered self knowledge, and decides that the statement was a genuine expression of admiration.

 

By treating the identity as a storehouse of self knowledge, we can model or potentially replicate the way that identity and personality function in human beings.

 

Consciousness is another vital consideration. I think that we can model consciousness partially through its relationship with identity. Let's take seeing for instance. It is difficult to accurately deconstruct one's thoughts and senses so bear in mind I'm not trying to accurately determine how the human brain actually works, merely how we might think about modeling a computer intelligence with similar features. So when I look out through my eyes, I can think of this process as a combination of processes, on the one hand I am capturing visual data about color, shape, depth, etc... On another level however if I choose to focus on it I may become sharply aware that there is a living thing which is me looking through those eyes and perceiving the world that must therefore be in front of where I am.

 

This perspective can potentially help to deepen our conception of this model of conscious thought. Think of this process of focusing as making a query, it is a willful effort to realize who is doing the looking, where the information from my eyes is going. We can begin to think of the simulated identity as a useful construct the brain employs to orient itself in the environment. The self is a construct to which sensory information can be passed for orientation and from which information can be collected to make calculations. I perceive a me looking through the eyes so I can register that the eyes are mine, that the seeing is being performed by the me, and that the things being seen must therefore be in front of the me, so if a dangerous animal is a thing being seen then I can think about the threat that it poses by realizing that it is the me doing the seeing of it which must therefore be in front of me, placing me squarely in the realm of danger. So then other parts of my brain on an update cycle can go into the me object to check whether or not isInDanger is enabled at which point the various parts of the brain doing the checking can spring into action and do what they are programmed to do to resolve the situation while isInDanger is enabled.

 

So to recap, the identity is essentially a box into which information about the world can be placed to orient the self to the world and from which information about the self can be extracted in order to make various other calculations.

 

Of course now we must take a step back and talk about thinking, as I have definitely skipped over knowledge representation.

 

When we as human beings use words, we do so in a capacity above and beyond a simple grammar engine. I'm thinking of course about the Chinese room contention to the Turing test. We must think about how to transcend the web of words defined by other words, because the word only has weight in its reference to a thing, or in more abstract dialogue a word may exist as a reference to an idea, but on some level even some of the most abstract ideas have their roots in things or relationships between things or the states and conditions and natures of things.

 

In some ways we can start to see a picture unfold where a generation of new ideas comes from abstraction from a few base concepts whose interplay and expansion give birth to an ever widening range of possibilities.

 

The key then, in my mind anyway, is to determine what is the smallest set of concepts that can be used to create the primary building blocks for the kind of intelligence that could grow to human and superhuman levels of intelligence? Also what is the most efficient way to represent new ideas so that the consciousness doesn't severely leak memory before it has learned much of anything.

 

We already have some semblance of a self, though not yet very well fleshed out. For an intelligence to properly orient itself I would suggest it also needs to have a concept of world and a concept of other. Without sensory devices (I should say without consideration of sensory devices for the time being) the intelligence needs an interface so that it can begin receiving new information from a specific instance of other and it needs a way to be able to sort the incoming information according to what it already knows. It also needs a way to create new temporary ideas which can be considered for future deletion, can be later so closely related with other ideas that they to an extent are more or less combined to save working memory, can be stored in a separate file and saved to disk, etc...

 

Here's where I think words come in, and rather than using dictionaries I think the machine needs to be able to create a word web, in other words each new string should initially be considered an independent idea, the machine also needs a counter for new instances of an idea so that it can measure the extent of the relationship between ideas based on usage, in this way the machine begins to build a complex web of ideas each having varying degrees of relationship with other ideas. It also needs a function to create new objects as its understanding of inter related ideas grows. Furthermore it needs a function to incorporate new ideas into base concepts.

 

Here is where this conception, if it be worthy, could use the most fleshing out. Once you begin inputting new strings through the interface initially the machine might have to request more information, whether it should think about the string as more of a property name, more of a method name, what the object would be, etc... As it evolves it should be able to reason such things for itself.

 

Furthermore there could be room for some genetic algorithms but I haven't thought of an effective implementation with this conception quite yet. Like I said before, I'm still pretty green when it comes to AI so go easy on me if you don't mind, but I think I'm ready to hear some critical feedback and additional thoughts.

[sciwiz12]

I wrote another post in the psychology forum attempting to further explore the concept of ideas, and while I'm still waiting for feedback on both this post and the post on ideology, I do have more thoughts to add here. I think that in order to teach a machine to understand words it might be useful to take works of English literature in plain text, have the machine search the text for repeated words, creating a count of each word in the text and the number of words between the most commonly used words. For an idea of what to do next it may be useful to divide the machine mind into primary regions with an interface acting as a gateway between incoming words and parts of the mind used to store those words.

 

Initially all words are stored on the language level until they can be sorted, along with information about potential relationships between words based on frequency within a text, possibly also by attempting to intelligently identify parts of speach and relationships between the same word modified by usage and tense.

 

Then it would go through a comparative phase where it would open up the parts of it's brain to see if the words taken from the text can be found within within the database being accessed. If the word is contained within the computer will check to see if it thinks one of the new words might be related to the existing words. If yes then the new word will be stored with information about its relationship to the existing word.

 

If two or more related new words already existed in the data base it would update the relationship counter to show a stronger relationship.

 

 

This process will repeat for each primary database, or possibly a primary folder containing data bases for which each data base will be examined.

 

Then the machine will go into evaluation mode where it stores all words not stored into a database into a database of unsorted words for later consideration after it is able to access more literature.

 

In the evaluation phase it also will go through databases making comparative queries to see what words have been stored in multiple data bases, wt which point the word is given a special status as a bridge word that may find use in multiple areas. The bridge words that have connections to many databases, are tested or evaluated to provide a rough prediction concerning the likelihood that it is an overlooked article or conjunction. The literature files in plain text may also have to be stored in accessible folders for reference which means that words may also need to contain some data concerning their sources for future context testing.

 

Words that span many databases in various areas (folders) but seemingly aren't conjunctions, articles, etc... Are put into a general use folder, perhaps with a special function to occasionally notify the programmers of the words in the general databases and possibly query the developers for more contextually related words to help it zero in and identify the word with greater detail.

 

 

The folders should probably all be general topics with broad headings containing sub folders with slightly less broad headings with maybe one more layer of folder and the data bases, for instance science containing chemistry containing organic chemistry or something.

 

From here we can begin building a framework for the computer to begin using the words by having information hard coded concerning the relationship of words contained within certain databases to itself and the world as well as methods to utilise new words and their relationships to begin constructing ideas about itself, about the world, and about others.

 

Ideally we could then examine these constructed ideas to measure the machines evolution and success over time in correct use of information to orient itself to the world.

 

We could also begin to employ clever algorithms here to have the machine test and evaluate its own ideas and create better methods based on patterns which lead to more frequent successes with higher scores.

 

Of course it's time to stop for now because I'm beginning to get far too vague and need to reflect on where I am so far and where to go from here to get further.

[fiveworlds]

Firstly I know enough to know I'm not in the camp that thinks we should be modeling the hardware and software after the human brain. It seems to me that the technology to create highly intelligent evolving computer consciousness is actually closer to being a reality than what would be needed to emulate the human brain.

[\quote]

 

Why the human brain?? We are fairly stupid. Kill each other for ridiculous reasons. Don't care about other people. Then we get grumpy when we are hungry, tired, haven't washed or gone to the toilet.

[sciwiz12]

It seems to me that they key to an AI intelligence is not that it mirrors the human brain but that it has a good way to think about the world and try to understand it better over time.

 

If a machine understands how the world works it is intelligent, if it can apply that understanding it is also powerful. The goal is to find a good approach for constructing a way of thinking that allows a computer to build and test its understanding of the world, but we have to think about what it is to understand.

 

For instance I may know that 1+1=2 which I can use to answer the question: "what is 1+1?" But there are different levels of understanding. Understanding has both depth and variety. By depth I mean layers of reference. 1 is a number, it is a quantity, an amount, an abstract idea, a concept, a notion. By variety I mean different ways of thinking about the application of an idea. I can use 1 to solve the above problem, I can also use it to refer to the amount of a real world object I can identify, I can break any number apart into 1s, etc...

 

An intelligent machine needs a way to be able to develop new modes of thinking abstractly, but I'm getting ahead of myself. Let's start with something simple.

 

Say I'm given a large file of plain text. Assuming the text all uses legal English grammar and syntax I should be able to Decode it using a few simple rules. I start with array of capital letters and an array of ending punctuation characters as well as a quotation character and a variable for the space character. I may also need some other character variables for later to search for commas etc... But this will do for starters. First I need to be able to build an array or string of characters from the text representing a full sentence.

 

I need iterate through the text and for each character check to see if it's in the ending punctuation array, and if it is store it in a new array with a number to represent its position in the text. I also need to find the quotation marks so I can think about everything between pairs of quotation marks separately. Then I need to find every capital letter in the text and determine whether or not the character just before that not counting the space was an ending punctuation mark.

 

If so I know that whatever follows and comes before the next ending punctuation excluding eclipses and quotations is part of the same sentence, so I can store it all together in an instance of a sentence array.

 

For each sentence I need every collection of characters that come between spaces with excluding characters that are commas, colons, and semicolons. Now I can go through the characters of each word for capitalisation, apostrophe, -ly,-ing, and other clues to help me make educated guesses concerning parts of speech. Then I can apply some basic rules of grammar and syntax to try to get a better sense of a sentences meaning. I can identify nouns, make some educated guesses to try to figure out subject, object, implied subject, etc...

 

I can also output some data concerning my level of certainty of each word if I'm in test mode.

 

Once I have a good sense of a given sentence I can perform a few other tests etc.. For quotation marks and of course consider the use of colons and semi-colons. I know I'm oversimplifying here but so far all I'm doing is using a few rules of grammar applied to a text written especially for my computer mind to read so that I can make some good easy guesses to identify the use of various words and the likelihood of certainty for each word for each sentence.

 

When I'm done I have a new list of words, their parts of speech, and I can use each sentence to possibly begin to understand the relationships that certain things have with each other. For instance given a sentence "A shark swims in the ocean".

 

I may not have a good grasp of what any of those things are yet, but assuming I can accurately identify parts of speech I know that a shark is an object, swims is an action, and ocean is another object upon which a swims operation can be performed by a shark.

 

Now let's take this in with some ideas from earlier, let's say I already have a very basic and shallow conception of the world built in, and included with that is the word "ocean" and some data and methods I can use to help me understand and conceptualise, and think about oceans. The word can now act as a sort of receptor site for me to evolve my understanding of the world.

 

I already knew a bit about oceans, now I know that whatever a shark is, it can perform a swims operation in the ocean. This may not really make sense to me at first but it is a building block I can use. I may need some other text decoding methods so maybe I can figure out tenses or proper nouns, or filter out conjunctions and articles or something, but later on let's say I see that a kid also performs a swims action on whatever a pool is.

 

I can propose two new ideas, a pool is like an ocean and a kid is like a shark. A human grading my ideas can score me highly for suggesting that a pool is like the ocean and poorly for thinking that a kid is like a shark, increasing my association between pools and oceans and decreasing my association counter between kids and sharks.

 

In this way I can use text to learn and form rudimentary ideas. The real test though is when I transition into a phase of reflection where I truly test my understanding by creating a file in English using what I've learned to try to write about the world as I understand it based on my built in constructs modified and expanded by an evolving understanding of new words.

 

So in my mind the main challenge to tackle next is how to set up the base constructs before the program even runs.

[fiveworlds]

I can break any number apart into 1s

[\Quote]

 

No you can't.

Edited November 18, 2015 by fiveworlds

[sciwiz12]

Excuse me, any whole positive integer.

[fiveworlds]

Excuse me, any whole positive integer.

 

You still can't. The world's fastest typing speed ever recorded was 212 words per minute.

 

(212)(125)(24)(60)= 38,160,000

 

ergo the biggest number you can split into 1s is less than 38,160,000.

 

As a side note china's fastest computer does 30.65 petaflops a second. which is

 

30,650,000,000,000,000 calculations a second

 

Given that unicode is 120000 letters approx in size at the moment

 

(38,160,000)(120000)= 4,579,200,000,000

 

Therefore china's fastest computer can solve for everything you will ever write in less than a second.

Edited November 19, 2015 by fiveworlds

[John Cuthber]

 

 

 

Therefore china's fastest computer can solve for everything you will ever write in less than a second.

China's biggest computer probably can't solve the Times Crossword.

I could copy it out in a few minutes.

I could even compile a crossword in a few hours.

If I transmit that over the web- say I send you a digital picture of it- then that's a whole lot of numbers- but your computer won't solve it as quickly as my friend Hugh- he's good at crosswords.

Edited November 19, 2015 by John Cuthber

[sciwiz12]

I don't know what you're talking about, I mean mathematically any positive integer can be broken down into ones, at least theoretically if you gave me a positive discreet integer and I could live forever I could keep subtracting one from the value of that number regardless of size until it had been reduced to a large collection of ones. I don't know what you're talking about.

[sciwiz12]

More thoughts occurred to me, as I am reviewing some math that has become quite rusty for me. I'm thinking about it in terms of an artificial intelligences ability to conceptualise.

 

I think an AI should have some separate components for performing mathematical operations and an interface between language and math components. Now I'm building a more complex notion of what it is for a computer to have an idea.

 

For instance the concept of an apple should contain within it a reference to a geometry method with apple specific parameters under the name appleGeom or something. It could store data concerning an apple's color based on light values etc... So that it could use that information to think about how apples would behave under various circumstances.

 

This gives us a sense of imagination, it can create ideas by taking what it knows and testing it under different scenarios to simulate what would take place. It doesn't have to create a visual representation, or show a graph beyond testing and debugging, but it would have to store the data based on the type of simulation.

 

I know I'm being vague because this concept alone takes a lot of work starting from even basic mathematical concepts and it requires a solid fundamental understanding of physics and geometry as well as algebra and calculus even for just a basic version, but given the ability to run mathematics and physics simulations the next step would be to relate these abilities to language and to it's understanding of itself, the world, and other people and to manage the interactions such that at no point does the machine attempt to process more data than it can handle in a short timespan.

 

Again, I think the trick is to start with a few basic words that are predefined and tied to scripts containing information pulled from the more abstract math and science scripts.

 

Upon being given a new word and its relationship to previous words it must attempt to figure out where this thing belongs in the world, perhaps through a 20 questions style game with the user, or perhaps through context. Perhaps it could use a method for determining when a word it previously didn't understand has enough information to attempt to understand what it is. It can design a class around it and store the class information in a human readable file pending approval awaiting correction and grading. Having been corrected it can incorporate the class into its knowledge base and over time use some statistical model and seek patterns and differences between incorrect guesses and correct guesses to determine if some part of its understanding of the world is wrong and suggest corrections and updates.

 

 

Now I'm really getting vague. Yet for me this is starting to come together, before it can create a concept of itself, the world, etc... It classes containing methods and data concerning the various areas of math and science. It also needs some of the previously discussed methods in order to parse at least English. Possibly also a class to facilitate the concept of language itself. From these it can build changing models of itself, the world, and the users. As it parses more English it needs to be able to think about new words first in terms of English to help it understand the category of things to which a word belongs so that it can pinpoint the category of scientific and mathematical knowledges it can use to create a model of the object and describe the place of the object in the world and understand the actions and operations that can be performed on it.

 

It'll also need a way to update object and action concepts when it has added new information and a way to output information based on its understanding of the world.

 

It needs a way to write an essay addressed to the user to demonstrate the ways in which its understanding has evolved.

 

Obviously I still have a lot to consider, more than before, but I think I'm sniffing down the right path.

[Strange]

I don't know what you're talking about

 

If you are referring to fiveworlds, you will come to realise that about 90% of his posts are utter crap. But occasionally he says something quite informative.

 

I mean mathematically any positive integer can be broken down into ones

 

On the other hand, I don't know what you mean by this. What does "broken down into ones" mean?

[sciwiz12]

I don't know how this math thing became the Center of the discussion while everything concerning AI is overlooked.

 

But to address the point briefly any positive integer, say 5 for a simple example, is composed of a number of ones equal to the value of the integer.

 

5 = 1+1+1+1+1

 

Such is the case for all positive integers, which is what I mean when I say they can all be broken down into ones. Any positive integer can be expressed as a geometric series of ones. Is geometric series the right word? All added together? You know what I mean.

[Strange]

5 = 1+1+1+1+1

 

Such is the case for all positive integers, which is what I mean when I say they can all be broken down into ones. Any positive integer can be expressed as a geometric series of ones. Is geometric series the right word? All added together? You know what I mean.

 

That is a sort of arithmetic series. It is also trivial; it is what we call "counting". Why would you use this as an example of a computing task? It doesn't even require a computer.

[sciwiz12]

I don't know, I'm not even sure how it's relevant really, I'm not sure why we're still talking about it. Were my ideas so far so stupid that nobody even wants to bother? Pointing out one comment I made about number deconstruction is the only interesting or worthwhile thing to discuss on a thread about AI.

 

You're not going to school me and point out a theorem discovered decades ago that makes my ideas irrelevant or intractable?

 

I mean I don't mind being called out on being ignorant so we can discuss a more relevant approach but is artificial intelligence really so passé you'd rather talk about a non-foundational comment I made concerning numbers, not even a laughably false statement showcasing the extent of my ignorance but an unimportant statement I made in passing while discussing my thoughts on the designs of a program with the purpose of creating more accurate models of the world and using those models as a framework for creating intelligent response to users in order to pass tests of intelligence.

 

Are you sure you wouldn't rather talk about that? Right or wrong it at least seems more interesting.

[Strange]

Sorry, I don't know much about AI or approaches to it (I only picked up on the 1s thing because someone else mentioned it). I have read arguments for and against it (hard AI) being possible. The arguments against seem to come down to "but the brain is magic", while the arguments for seem plausible.

 

I think you said that copying the way the brain works isn't necessarily the only or even the best way to go about it. That seems sensible. It might also open up the possibility of an intelligence totally unlike ours. Which would be interesting.

[DevilSolution]

I don't know how this math thing became the Center of the discussion while everything concerning AI is overlooked.

 

The reason is people nit pick your post looking to find errors so they can quickly dismiss the integrity of the rest of your post.

 

As far as your OP goes and the further extensions to it, it all seems rather Pseudo "thinking" on the subject. In one of your original statements you claim to "know enough about AI to not sit in the neurosimulation camp of AI, but you dont seem to pose many computational answers to the concepts you later relay as AI requirements.

 

I also studied AI and computer graphics as a module of my CS degree and im also interested in philosophy and neurobiology / chemistry.

 

What i'd suggest is that you break down your idea's into simpler concepts so that you can address them computationally. For example as im sure your aware "machine learning" is a huge area of AI. Why dont we look at current approaches to machine learning in relation to some of the more complex idea's you have regarding consciousness and imagination?

[sciwiz12]

I suppose that's a pretty fair critique, I haven't addressed the matter computationally yet because I find this manner of expository abstract postulation to be both a more relaxed and more deeply personal way to exchange ideas, though I suppose it would be daft to act as if such a view would extend to others. Also I'm sorry if it seemed as though I was posturing as any kind of expert in the matter, I don't demand to be taken seriously on the matter or to have my ideas discussed exclusively, I would just like to see more discussion on the general topic.

 

I do intend to discuss the topic more computationally once I've stumbled into a more solid approach and really narrowed my design. At that point I think I would feel more comfortable discussing the approach I'm developing in terms of the other major approaches and the computational workload and specific implementations.

[DevilSolution]

I'll give you a tip, conceptualizing something, it could be anything, say a trading concept, works ALOT different from how you computationally go about implementing it. I understand you want to talk conceptually about imagination and consciousness of an AI but thats somewhat more philosophical than computer talk. Really you just need to break down your idea's one by one so you have a more solid framework from which to start.

 

I'm not going to lie, having a fully self aware, self conscious AI with an imagination is quite ambitious, but hey, its probably possible.

[petrushka.googol]

Sorry, I don't know much about AI or approaches to it (I only picked up on the 1s thing because someone else mentioned it). I have read arguments for and against it (hard AI) being possible. The arguments against seem to come down to "but the brain is magic", while the arguments for seem plausible.

 

I think you said that copying the way the brain works isn't necessarily the only or even the best way to go about it. That seems sensible. It might also open up the possibility of an intelligence totally unlike ours. Which would be interesting.

From my rather primitive knowledge of AI -

Computers use heuristics and alpha beta pruning as search techniques

Neural networks are getting popular

The cloud is being combined with neural networks to improve decision making

Wonder what the future holds?