ZIA brain building for the masses

It should be very instructional to allow artificial agents to send each other messages and by “random” variation and darwinian success evolve a language. The key things to pay attention to are: 1) as always, existence of other individuals and their cognitive abilities is too hard to infer, it has to be hard wired 2) there is no guarantee, for the agents or us humans, that these messages will constitute a good (and even less so a perfect) description of the world. Realistically, the opposite should be the norm, where a description regarding food and sex essentials should suffice, as it did with so many of our ancestors and contemporaries 3) Key research outcome would be whether grammatical forms would emerge that resemble verb, subject object etc. It is fair enough that the existence of individuals makes subject, object and a 3rd category, verb perhaps, obligatory, still the game is on when it comes to the use of the other grammatic elements. A clue from Pinker, is that language works in many cases “relativistically”, or to put it differently all the little words create the context. This is just an adaptation/concession to the fact that individuals are different, in size,shape, internal states etc. It takes a certain optimism for a communicator to use words and hope that the other party is not so different that communication breaks down. Then again, communication always breaks down, just to acceptable degrees, it is the most distinguishing characteristic of life and intelligence that organisms make constant adaptations to achieve, among other things, consensus and cooperation. Then again, we have so many books and films and experiences where misunderstandings drag on and on, and this should happen to artificial communicators all the time as well. It would also be interesting if these agents were in these sophisticated physics-realistic simulation environments of today(’s games), and whether they would move towards developing science! A certain ugly problem will reappear however, the “Alien problem”: we will have no easy way to figure out WHAT the agents have discovered, we would have to observe and analyse their system the way zoologists do with animals or exobiologists do with weird ideas, and try to see where they click. A certain danger lurks as well the agents surpass us, we would have no way to know this, unless of course they hack their simulation environment, break out into our computer and start talking to us in perfect english!

,

The distinction between left and right is one of the mysteries of our universe. Certainly it is related to a line (ray of light) dividing space in two parts from the point of view of an entity, but there is no way to objectively define left as left, for example the idea of dividing the visual field in 360 degrees and assigning 0-180 to, let’s say, left, presupposes we know where left is to start numbering the angles in the right direction. It looks like chirality depends on both an inbuilt recognition mechanism, as well as a physical asymmetry that “implements” it, perhaps going all the way back to molecular chirality. In fact it seems to be a phenomenon that stretches all the way from physics into biology, as it is quite unlikely the distinction could be made without a (bicameral) asymmetric brain. In some sense chirality is a sense, that little special part of the mental world where “you know when you know”, the taste of orange being the taste of orange (I have not made up my mind if/how important it is that taste is a composite sense combining different primitives of smell and taste).

Part of my thesis is that, for all practical purposes, there are cognitive primitives, things we just know and somehow all other things that we find out are decomposed to such primitives. Discovering them would be akin to Euclid’s axiomatic foundation of geometry, only infinitely more challenging and rewarding. Automating or somehow contributing to the search for such primitives is still very tough, and it could be that alternative formalisations will be discovered that are equivalent but radically different, somehow depending on the relationship between their different parts rather than any “absolute values”, resembling algebraic views of geometry and geometric interpretations of algebra and mathematical analysis. For example, “3d-space” could be or couldn’t be a primitive, ie people instinctively know about three dimensions, or perhaps they always think of 2 plus some height corrections, or perhaps they are “programmed” to do ballistics by using an horizontal angle around their neck and a vertical angle.
In any case, the idea here is that a computer program without such primitives would see a very fluid universe, akin to the mystical or psychedelic experience. I don’t want to make epistemological statements on whether the earth is alive or the Ganges is a demigod and things like that, I simply want to claim that there are great reasons why we see things as inanimate, plants, animals and humans. In other words, I am asserting these are cognitive primitives (or closely associated with cp) ,”for sure”. The primitives should be seen as dimensions in feature space, where different values of a “feature” allow us to differentiate between animal and plant, and treat them differently. Surely a person has a personal history, important someones, aptitudes and aversions, loads of other things. Which would qualify as “primitives”? Well, it is quite obvious that “object” should be a cp, and objects have personal history, which can be attached without much ado to objects as different as the earth, you, the word “marriage” and the concept “marriage”. But then there are other interesting things for the survival of animals and humans, for example their pleasure seeking or survival drive could lead to regular unexpected solutions to their problems, whereas plants and rivers seem to operate in a different time scale and concept space. So, pleasure seeking or survival is a possible primitive, with a value close to 0 for non-animals.

Now, let’s ponder that previous statement: is it possible that a general intelligence will decode an organism (eg a human) and list what gives it pleasure? Well, it is possible, but rather unlikely. It is more likely that there are some primitive (mathematically, not morally) pleasures that we could list, and then assign these primitives and their derivatives to different time-human vectors (as objects are changing, we are forced to make all the object’s attributes relative to a point in their personal history). In other words, we need a routine that searches (eg in a biography or a chat session) for a persons desires, not a routine that deduces that people have desires or a statistics function that creates the concept “X” when a person has spent more than 1% of its waking time with another object, let’s say their new cabrio.

Discovering (more of) these primitives could be a project similar to that of subatomic physics: a few of the primitives will readily suggest themselves, others will be lurking underneath the ones we took for granted, and still others may announce their presence only in an artificial, high computation/high energy environment. As with physics, the absence of primitives may make itself felt by little discrepancies between Zia’s expectations from the world and Zia’s measurements of the world. (I am reminded here that concepts such as a measurement and experiment are still looking for good definitions and explanations, let’s just take them for granted now, I have the suspicion we will come back to them). In all cases, it is vital that Zia keeps measuring and simulating the world, and, for the purposes of the Loebner prize and Turing contests, that world is the world of persons!

Most things about language are still puzzling for science, but I hope to be able to outline the AGI (Artificial General Intelligence) essentials for the sake of ZIA here. So, if you followed the previous post, the brain has developed an amazing multiresolution representation of the world, and could seek to externalise it, write it down perhaps as a practicality for one’s own sake. This is where the plot thickens, as the brain’s solipsist, solitary universe meets a world teeming with potential friends and foes. Luckily, or by definition, the most interesting lifeforms share a lot of your genetic make up, meaning you do not have to work on developing a language from scratch (this “from scratch” refers to the problems of one alien intelligence trying to interpret the signals of another, rather than simply making up new words and rules around concepts we have mastered, as with the creation of Esperanto and Klingon).

Again from an earlier discussion, you wouldn’t expect the brain to use flowery language, to settle for linguistic ambiguities, or use imprecise or altogether meaningless verbalisations “people are good”, “people are evil”, “what is the purpose of life”- ultimately the brain wants to work out things like “there is a banana there, and the banana-eater closer than me”. It is likely that the secondary characteristics of language serve the social animal very well, and have, perhaps accidentally, allowed the mind-language structure to abstract and generalise ad infinitum. For my purposes however, the most salient feature of language is that it MUST be a good universe representation. “There is a bad man trying to break the door” must be something of an ideal compression of a myriad of Newtonian Physics facts such as “he is holding a 5 kilo axe and swings at 5 rpm at a typical speed of 60 km/h - and he did spent 5 years in a mental institute”. It could also be that this man is a true hero trying to save somebody behind the door, which suggests we are dealing with lossy compression, which by no means detracts from the value and power of such linguistic/cognitive compression.

The corollary of the previous two paragraphs is that language suggests itself as a cognitive representation for artificial brains. When “thinking”, we may want to make use of decompressors, expanding statements to something of real-world relevance just like the mp3 player does, and we would love to have compressors that take such expanded forms and produce common, imprecise, illogical English. In fact, a “real” machine translation should work exactly like that, balooning input statements from one language and deflating them into the other. The use (and invention) of intermediate languages for machine translation is not new of course, what is new (or, anyway, newer) is the implicit, embodied intermediate layer through compressors and decompressors. This intermediate layer will have to be highly tuned and, perhaps, intelligent to deal with the vastly asymmetric mapping between the linguistic and physical spaces, not to mention that for human affairs the “physics” involves very slippery internal worlds, style, emotion etc.

Let there be mind! I have finally decided to proceed with my brain building project, the Zero Intelligence Architecture (zia). This blog should serve as a biography for zia, as well as some kind of program documentation. The advantage over program documentation is that there is space here to discuss the alternatives that were rejected, as well as allow comments and, ideally, a fruitful idea exchange with the community. There is of course a multitude of very deep and tricky questions pertaining to this attempt, and many of them may never even have occurred to me so far. Many others I have answered to my satisfaction, while others still have troubled me for years and I have made little progress so far. In any case, my ambition is to RAD the project without any particular software engineering or project management approach, coding in the easy parts in all kinds of top-bottom, bottom-up, midway-out sequences. In a little more detail, I envisioned a software system with monstrous properties, with distributed, migrating, dynamic, freezable processes in all kinds of configurations, compiled and assembled from multiple sources and languages and frameworks. In short, “my” brain is more of an operating system than a program, and could end up packaged in a virtual Linux machine pretty soon. Until then, I will be proceeding with some scripts that would tie OpenCYC with some (3rd party) NLP code and my own “worlddb”, an unspecified as of yet structure that would encapsulate all the brain knows up to a point.

Without further ado, I am asserting that a brain is just a predictive structure (an interpretation) for the micro-environment of a localised physical configuration (also known as an organism), and the controlling loop of the brain should be:
L0 forever: use interpretation for predictions, refine interpretation

I am also asserting that what makes (part of) an organism a brain is continuity: survival and reproduction means the brain has time to become interesting and relevant. There are interesting properties attached to continuity, for example strategies that have proved very promising: treating other physical configurations as brains, distinguishing between animals and plans, and the zillion others cognitive sciences study. It is conceivable that all of these have “evolved”, but I will take as many shortcuts as possible to allow zia to evolve from a common-sense starting point. It is most interesting project to go for “total evolution”, ie have a brain deduce from god-knows-what inputs that it lives in a 3d world etc etc, but zia aims to look more like a 4 year old at birth. Two other research projects that are relevant and related but I do not want to address (yet) is an analysis of why societies progress technologically when the members do not necessarily progress intellectually, as well as the development of a generic method to distribute computation, eg a good way to covert a minimax chess algorithm to stream processing code.

What about softer aspects of the brain such as “who am I”, “why am I here”? Well, assuming other brains will have to be evaluated as friends or foes for the purposes of the loop L0, there could be a reflective use of the relevant functions, but I do not expect much from it at this stage. Consciousness? Well, Igor Aleksander’s 5 axioms are unimpressive: Imagination and planning are matter-of-fact attributes of the scheming brain, sense of place is not only key in the darwinian struggle but is arguably optional for less localised intellects, decision/emotion is a behavioural mode that can contradict the “conscious” qualia (I mean, it is too much of an automation), and directed attention may just be a computational limitation of the human brain.