Yes to Freedom

A TL;DR SanCon post on AI --> http://www.sankakucomplex.com/2009/04/1 … ent-104701

You are a remarkably intetesting person.
And I have huge respect for anyone that is willing to walk down the path you chose.

Nice tl;dr post as well: suprising to see a sancom comment worth reading. (this makes 2, your's and atma's)

LunarSD wrote:

(we tend to function in overload-mode, and are not long for this world, innately)

Better to die young after a busy life than to live for ages doing nothing, in my opinion.
And I'll extend my previous comment; Remarkably interesting family it sounds like you belong to.

I live in the country of 'tall poppy syndrome' sadly, so its rare to see a person really shine.

I'm not selfish about promoting AI research ( more like obtrusive? ^^; ).

"...it can begin with something small. If one smarter-than-human intelligence exists, that mind will find it easier to create still smarter minds."

- Ben Goertzel

There are ambitions which exist, open to anyone who is willing to chase them... which are simply too nourishing to human curiousity not to incinerate one's self speed-racing into.



So I'll just edit together a generalized text-dump linking together fragments of correspondences I've had with a few other cyber entities on the AI topic here ("for the lulz"?).

-----

"I'm curious as to why ur using a fully comprehensive physics engine as opposed to a specially designed game environment that simulates the complex real world environment effectively without using quite so much processing power, which I thought would be the best way to teach an ai."
-- Because language is based on the types of cause/effect a lifeform has learned (discovered are invariant) in their environment. I'm interested in making an AI that can eventually speak intelligently. Say you made an ideal, intelligent AI architecture... if it has nowhere to exist/interact to learn invariances that are similar to our own, it wont be able to communicate intelligently with us. Innumerable concepts would exist to it only in theory (like soaking/absorption/wet/dry/drying/hot/heating/friction/cold/blowing/suction/coalesce etc.etc.etc.; subtle differences such as between scrape/prick/slice/stab/pinch/push/massage/etc would be lost on them). Any verbs they are to be capable of learning, you would have to write specific code in the Physics engine to account for those circumstances. And time-based concepts (like spoilage, growth, being able to dip a cup into a basin of actual water and notice the water level decrease proportionally to the accrued liquid, etc.), which are the most important for learning cause/effect, would be left unknowable. I prefer a Cellular Automata (http://en.wikipedia.org/wiki/Cellular_automaton) sort of system, where everything that happens, always happens for a solid, trackable reason.

A real-life robot is several orders of magnitude more difficult to make than a software AI, and as technology stands it is impossible mechanically to fit enough sensors on one to even remotely approximate the number a simple housefly has. As you know, Bayesian systems are more comprehensive and intelligent not just from the more priors they have access to quantitatively, but also from the variety those priors have to offer the system as a whole. If you put emphasis on just quantity, or just variety, it logarithmically reduces the value of the system. No matter how one may try to adapt, say, the Half-life engine for an AI project, it wont be possible to make it complex enough to allow an AI inside of it to learn the experiential differences between the sort of concepts I stated above. (note: see section marked " *** ")

Also, in any modern 3D engine, collisions are reduced to simple geometry rather than specific locations. The whole concept of a modern Physics engine is to make things *look* like they're actually happening, rather than building a system where everything that occurs, occurs for a solid learnable reason. "Learnability" being key there.

Processing limits are a problem for every AI researcher, whatever field/strategies they've chosen. 4 Teraflops goes a long way (http://www.pcmag.com/article2/0,2817,2347889,00.asp), especially if you aren't concerned at how good the Physics+AI engine looks, but rather how well it functions. Though I'll be trying to get away with significantly less than that -_-;

-----

*** This brings up the debatable existence of qualia. "Examples of qualia are the pain of a headache, the taste of wine, or the redness of an evening sky". Though I personally think the debate is silly. Qualia could only exist both conceptually and materially, as it is a term describing the sapient experience of subjectivity itself. Cutting out the sophistry, qualia is better described as whichever material algorithms are necessary to construct a network's ability to conceptualize as a form of information processing.

Qualia is conceptual in the form of a self-perceiving mirror network (you could also loosely call it self-reflecting, for fun really) - described more computatively in the section below this one - comparing the relative relationships between invariant predicates. Predicates are any isolatable properties you can extract/deconstruct from a concept. i.e. Predicate categories of an apple could be the color red, a top-heavy curvaceous structure, etc. These are properties which you can also use to partially construct other concepts. In a higher layer of abstraction, a predicate of a human is that it can eat apples. "In first-order logic, a predicate can take the role as either a property or a relation between entities" (source). Predicates are a kind of Russian Doll hierarchical concept, though their emphasis is on being shareable/re-usable elements in the formation of all concepts. Again, a predicate is a relative property, and the relationships they form as shared/shareable properties constitutes the fundamental substance of any concept. You can view most predicates as themselves concepts, which can house/deconstruct into ever smaller predicates, until you hit the basic percepts that construct the most fundamental invariance categories in a sensory input (like the colors and lines in the early visual hierarchy architecture, linked again a few paragraphs below).

So we've established how qualia is conceptual via the above relationships interacting as a network. But it is also material simply because it is accomplishing said interactions by physically representing percepts, predicates and concepts in the form of designed (whether by evolution, or by a researcher's hand) interactions between neurons and synapses.

The qualia of a mental state must have a material grounding for it to exist.

See Kohonen nets to get a primitive graphic depiction (and functional algorithm) of how related, isolatable elements from a scene can map into a simple Hebbian-esque neural network.

Here is a concrete example of an AI researcher putting these sort of ideas into actual practice.

-----

As for my own philosophies, I'm keeping my project entirely in the language of Neurons and Synapses. What it amounts to is a language inside a language. For me, it is a CUDA & .NET-based GUI parallelized programming language application that allows the user to "swim" inside a selected virtual brain from inside a cube-partitioned virtual universe that obeys ALL the laws/algorithms of Physics (chemical reactions, chemical bond assimilation/breakage, malleability/stretchability at the molecular level, friction, bendability, etc.) as closely as I have been able to program them.

The mechanics are not simple Nodes and Links as generic Networking concepts run, but cell protein wall encryptions, nested dendrite branching logic, simple cube-partitioned "brain lobe/sub-lobe" essential data-type boundaries (really no different than how modern computer architectures require various areas reserved for processing components, storage locations, data traveling, etc. except in this case it is an innately parallel system, and constantly adapting), all organized and evolving as a data ecosystem in 3 dimensions.

So that, for instance, when the AI's physical "eyes" observe a snapshot frame from inside a DirectX-based physical world, you can watch in the architecture editing window as the visible cells of the brain adapt to the hierarchical invariances of the pixels in that scene (see Jeff Hawkins, and the cell architecture of the human eye), abstracting the data up the hierarchy, dispersing it into the forward/backward flow of stored memories and present imaginings one could describe as the lifeform's Cortex-level Ego.

By Cortex-level, I mean that it is affected by lower instinct/impulse, but not symbiotically linked to one's foundational behavioral architecture. "Survival" impulses are hard-coded, but so is the forward/backward flow of invariances that causes the emergence of an Ego.  The degree to which impulses drive behavior is inextricably linked to how the Ego mutates and adapts to said impulses over time. The complexity of an Ego depends on the number of simultaneous percepts and concepts (recall concepts are made of property relations, or predicates) the cortical flow can maintain in working memory at any given moment, as well as the longevity and accessibility of salient memory storage.

Essentially, the storage bin is the processor. Its parallel processing rulesets begin as incipient self-regulating occupants whose input comes from both sensory organs and the growing contents of the storage bin itself. This is a viable form of computation because all forms of percept and concept are broken down into invariant hierarchies, represented as concurrent patterns across time in signaling frequencies, stored entirely in cause-effect logic. Concurrent = less than ~20 milliseconds apart (i.e. >50Hz), the "coherent thought frame rate" that signaling trains occur together to make a "scene" of intrinsic and extrinsic conformations of the current momentary stored&perceived thoughts ("downward hierarchical flow" concepts) and sensed information ("upward hierarchical flow" percepts).

Initial growth is primarily hormonally-guided (hormones are the most conveniently computable, as their signaling venue isn't localized, and can be applied en masse to entire neuronal colonies via a simple on/off, present/not_present flagging system) with the contents of this "incipient storage bin" being mostly survival animal impulses .

As growth proceeds, Ego/sapience formation is primarily a result of mirror network dynamics. Parroting both the storage bin's own memories associatable (association = related thoughts are always naturally amplified in signal via the "map"-like coursework of invariances as they had settled via cause&effect in past experience) to present percept/concept concurrences in working memory, and -more primitively- applying thought&behavioral mirror dynamics to thalamic and early cortical levels of the sensory-->hippocampal hierarchy.

-----

Attempting this sort of project alone is a war of attrition. I've been at it in one form or another around 4 years now, though admittedly I had quite a few basics to learn as well back when I first began, even now frankly, considering the scope of the damned thing. I regularly exercise, fantasize (why I'm here @ yestofreedom ^_^), and ingest nootropics (chocolate seed extracts, Piracetam, lecithin, Bacopa Monnieri, Huperzine A), all to help maintain concentration and motivation. I tend to average between 0->5 hours of sleep a day, then periodically crash and sleep entire days away.

The core code of the AI module is only about 10,000 lines (in .NET abstraction, using GASS for the CUDA lines, and I wouldn't call it optimized even remotely yet), but the architectures are basically a nested genuine-parallel language of literally just various placements of neurons and synapses and type-rules (how will the engine handle things when NeurType#013 is docked by SynType#042 at DendriteType#011 which excretes -only when sensorily or cortically signaled into- NeurotransmitterType#009 into ClusterType#101)...

There are as many as 100 Trillion synapses in a human brain, all firing in parallel with as little as 10-20 milliseconds between signals (but only in extreme cases, it's more commonly at 10 or 20 Hertz or so). Naturally, each individual one when viewed on its own, is mostly entirely meaningless ( see: http://yudkowsky.net/rational/bayes ). Synapse firings gain layers of probabilistic abstraction the farther up the Cortical hierarchy they travel. Much the same way as a .NET implementation of C++ is an abstraction of native C or MIPS assembly.

Granted, a significant chunk of these organic human synapses are for mundane processes outside the scope of the Central Nervous System, and are firing inside areas of the brain that provably don't contribute to Sentience or Sapience (learn the distinction). It's still enough to make one think twice about designing even a Mouse-level intelligence via their own ground-up architecture.

Mind you, to some extent you can definitely consider just making a simple cortical invariance-processor such as Jeff Hawkins details in his book and white papers, and then set up a Genetics engine inside your Physics engine, forcing typical animal instincts to emerge after thousands of generations of survival of the fittest (Obviously this requires something in the Physical world they must be forced to evolve against to "ancestrally", progressively outwit; inclement weather, simple AI predators, etc.). "Nature does it best", as they say. "Emergence" is a very common thing to hear when reading up on this kind of thing.

And then study the resultant Neuron/Synapse clusterings and Type-mutations that ended up responsible for the new behaviors in your evolved organisms (you could "dye" various ClusterTypes and track their progress as they mutate and settle into "punctuated equilibrium"). That's the kind of thing I'd prefer not to resort to, but only time and more experiments will tell if it's really required.

Making sense of it all comes down to what you were thinking when you first began laying down Neurons and Synapses in 3D Cortical space. It's like looking at the night sky and trying to find logical patterns in how the star's gradual pulsings relate to each other (assuming a universe where there actually was a logical pattern there). It would take an entire book to explain one's reasoning behind just one architecturing rationale in particular.

Last edited by LunarSD (2009-08-30 10:46:12)

Just wow... it's mind boggling how you managed to get this far all on your own.

In fact, I'm right now preparing for my final exam in Neuroinformatics. Nothing we did was even remotely on the same level of ambition, though the knowledge I gained at least allowed me to follow and comprehend your words. You have my utmost respect.

And suddenly my year studying Business Management really seems very unimpressive .

You are indeed a remarkable individual LunarSD, and I look forward to the opportunity to work with and get to know you. 

(Also, I suppose I should post an introduction topic at some point - I'll get on that after work.)

Unfortunately, I must admit I have nothing near your level of understanding on the AI topic (although to say otherwise would likely be an insult to your impressive intellectual accomplishments thus far), however it does sound very interesting and I support you fully.

On a more mundane level , your introduction makes me think two things, in particular:

1) Despite how much I dislike phones (to say nothing of employment), I kind of envy your job, as it really doesn't sound all that bad.

and, 2) You've got me kind of paranoid, and now I'm thinking I should go see a doctor to make sure I don't have any serious pressing conditions, that may only be getting worse over time...