torsdag 26 juni 2014

Mönsterigenkänning och vägen till superintelligens

Linked In "frågade" idag om jag ville skriva en "artikel"... så det gjorde jag. På engelska, jag vet inte riktigt varför, det bara blev så eftersom "inbjudan" var ställd på engelska. Den handlar om hur vi steg för steg ska ta oss från idag till en framtid med artificiella superintelligenser.

Håll till godo. Den är rätt snabb och rough så ha överseende med stavfel och tankekrumbukter.


Some anticipate an artificial intelligence (AI) to be able to pass the Turing test within as little as 15 years - and an additional 15 years thereafter trigger the technological Singularity (a kind of threshold after which the pace of technological evolution is so rapid that no human can keep track of it let alone understand it).

I wouldn't hold my breath for 2029 as the key date; however, I am convinced that a super human AI will emerge at some point and then evolve exponentially from there. It is not merely possible, it is inevitable. Resistance is futile.

First we need a good theory of how the brain processes information. Then we need to measure brain activity on the finest scale possible and use the measurements to improve on the process theory until it is approximately correct. Then we build a bigger brain and use that to build even bigger thinking entities. Size means more memory and information processing capacity, in short a genius, that could help develop even bigger, faster and sooner or later qualitatively more competent brains. And then it is out of our hands unless we merge with the AI:s.

1.    First, the human brain managed to emerge spontaneously from the primordial soup. It would surprise me if no one ever could reverse engineer and first just replicate and then enhance the brain’s functions, perhaps first biologically and then in a much more efficient and robust substrate. The level of intelligence (pattern recognition and hierarchical symbolism in the neocortex) is limited in size by the cranium, whereas an identical structure outside the brain could be expanded by orders of magnitude

2.    Second, intelligence is not magic; intelligence seems “simply” to be based on recursive pattern recognition. The brain survives by correctly observing patterns in its environment and anticipating and avoiding lethal threats. One important aspect of the environment is other people, another is itself and its body. By modelling people (including itself), awareness emerges.

3.    Third, whatever intelligence is, it is made of matter (the brain is made of matter – if not, all bets are off) and it is not likely that the current design and material are the best conceivable. Signal speeds, e.g., in an ordinary computer is one million times faster than in organic matter like the human nervous system. Just transferring the brain as is to a computer substrate of 2014 would make the brain one million times faster. By improving the actual information processing architecture, completely new orders of speed and capabilities are highly likely to emerge, given empirical evidence from the recent history of hardware and software technology.

4.    One exciting avenue of explaining, exploring and evolving past the human level of intelligence is Kurzweil’s theory of hierarchical pattern recognizers working from hidden Markov model principles.

a.    On the first level, specialized simple basic pattern recognizers (PR) are triggered by external stimuli, e.g. a straight horizontal line, or a vertical one, or curved, or some other arbitrary fundamental visual pattern.

                                          i.    E.g., a horizontal straight line prepares the next level of PR:s, excites them, for anything that usually contains a straight line – like the letter “A”, or the horizon, or a stick… there are A LOT of stuff containing horizontal lines. If a “diagonal line”-PR is triggered simultaneously, the likelihood of an “A” being seen is increased and corresponding PR:s are being excited and extra prepared to detect cues pertaining to an “A”.

                                         ii.    The likelihood of detecting a diagonal line is also increased since a horizontal line often comes with one of those.

                                        iii.    Other letter-PR:s like “E” and “B” also get excited, since they also have horizontal straight lines in them.

b.    Higher level PR:s also get excited, such as word-PR:s with the letter “A” or “E” or “B” etc. in them (“Apple”, “Ape”, “BANANA”, “Adam” etc)

                                          i.    Since everything happens in parallel in the brain, a horizontal line stimulates, to various degrees, all levels of PR:s from letter and word PR:s to smell PR:s (apple, banana) and even childhood memories of old relatives and apple pie. The PR:s get ready to detect this stuff without you knowing it

                                         ii.    If an “A” is more or less firmly established, the threshold to detect stuff that usually comes with an A is lowered, such as “B” (in the alphabet) or “P” and “E” in (apple or ape). That also explains why we can see or understand a word that is at an angle or partly covered

c.    Once the word “Apple” is detected (actually in parallel of course), it becomes easier to detect “fruit”, “pie”, “oranges”, “vitamins” or whatever historically has occurred next to the word “apple” for the brain in question

d.    Even higher up in the hierarchy, other PR:s get ready to detect biblical stories of “Adam” or other tales of knowledge, shame or whatever has occurred in connection with apples and Adam before. Simultaneously more letter-PR:s on the more fundamental PR-level get ready in an appropriate cascade of recognized patterns and excited PR:s to read the whole sentence or page, if that is what was detected.

e.    Depending on how intelligent the brain is, it has a certain high end limit of hierarchies, where very complex patterns like jealousy or love resides, but there is no reason to assume that it has to end there. A future person or AI could have an arbitrary number of levels and correspondingly complex prepared patterns or “emotions”.

5.    If the theory above, or something similar, turns out to be close to the material truth, an iterative process of modelling the brain and comparing the model with the brain could commence. Gradually as the models get more and more accurate and the resolution of brain scanning and imaging improves in both the spatial (room) and temporal (time) dimensions, nothing seems able to prevent a future point, within a handful of decades, where we know how and when a neuron fires, and how different neurons interact to form pattern recognizers, and how these in turn are organized in hierarchies to manage a symbolized representation of the environment.

6.    Once having a functional model, behaving exactly as the brain, depending on the relative state of biotechnology vs computing vs nanotechnology, it will be possible to expand a brain by:

a.    Transplanting more neocortex to a surgically enlargened cranium

b.    Fusing the biological neocortex with an artificial, computer based hierarchy of pattern recognizers

c.    Replicate the brain’s functions stand-alone from a human in a computer/robot

7.    After that, it is only a matter of mechanically expanding the number of PR:s (from the current around 300 million) and the number of levels of PR:s to create an entity with more memory and more and higher-level processing capacity than the un-enhanced human brain. That entity would be a genius surpassing the information processing and pattern recognition capacity of e.g. Einstein’s and Newton’s. If we create enough of those, sooner or later they would be able to improve the brain model more than an un-enhanced person, thus triggering an exponential intelligence evolution.

8.    Once the functional model is there, and once the cycle of one AI creating the next level of AI, creating an even higher level of AI etc is in place things will go very fast. Different AIs my compete for the lead or they may merge to evolve even faster. Why would two half-witted AI:s stand by to see somebody else take the lead if they could simply merge and take it themselves. And why would three other AIs stand idly by to watch that process instead of merging themselves…

9.    And why would any sane human being not seize the opportunity to expand his own intelligence by fusing or merging with as high an order of intelligence as possible?

10.  Will we thus become the Borg collective of Star Trek? Would that be bad?

7 kommentarer:

  1. Du som verkar mer eller mindre ekonomiskt oberoende, har du tänkt alls på att ge upp din nuvarande karriär för att istället börja studera eller syssla med något annat du är intresserad av (exempelvis ämnet du skriver om ovan..)? Du borde ju trots allt ha åtminstone 30 aktiva år till ifall du får vara frisk, folk har väl uträttat stordåd på kortare tid än så. :)

    SvaraRadera
    Svar
    1. Du är inne på helt rätt spår och jag trodde att jag redan hade gjort det ganska klart här att jag ÄR helt ekonomiskt oberoende och att jag redan HAR gett upp min nuvarande karriär (jag har sagt upp mig men inte slutat helt, bara slutat förvalta men sitter kvar som VD så lång tid det tar att skola in en ny).

      Det är också helt sant att folk uträttat stordåd på mycket mindre än 30 år. Givet inlägget gäller det också som KRAV för att inte vara helt onödig när AI ändå tar över

      Radera
    2. Aha, okej! Har inte följt bloggen så länge, så den informationen har förmodligen blivit delad före min tid.

      Efter det här inlägget har jag dock börjat tvivla på din existens... Tänk om AI-utvecklingen har kommit så långt att du är en cyborg som börjat bedöva dina överbelastade kretsar med alkohol? Hmm... Men en cyborg hade knappast lagt på sig extra kilon...

      Radera
    3. Ingen fara. Jag är glad att någon öht läser något inlägg alls, för att inte tala om reagerar på dem och kommunicerar reaktionen.

      PS: Om en robot ska klara ett utvidgat Turingtest, som i Bladerunner, så är nog några extra kilon en perfekt maskering, precis som det gäller att inte svara för snabbt på komplicerade matematiska problem.

      Radera
    4. Blev precis varse din underhållande blogg! Kanonintressant artikel du skrivit. Som av en händelse såg jag precis Ray Kurzweils framträdande på Google I/O - https://www.youtube.com/watch?v=MG_nOddk01E där han berättar om sitt arbete med nästa steg mot AI.
      Angående borgerna hoppas jag att någon form av estetiskt och etiskt sinnelag hänger med i intelligensexplosionen.

      Radera
    5. Jag blir alltid överraskad av Rays CV. 20 hedersdoktorstitlar, 3 presidentutmärkelser...

      Radera
  2. Makrillmannen29 juni 2014 22:58

    Kudos, som ungdomarna tydligen säger nuförtiden. Eller om det var för 10 år sedan...

    SvaraRadera

Träna effektivt; minimalt för maximala resultat

Det här är en träningsblogg (men även finans, teknologi, futurism, filosofi och life hacking). Jag lyfter skivstänger varannan dag, inte alltid utan alkohol fortfarande rusande genom ådrorna. Efteråt dricker jag en liter mjölk. Inga dyra tillskott som ändå inte fungerar.

Jag kör få och enkla övningar, alltid med fria vikter. Jag tror på att träna mindre än alla andra, men smart, och på att njuta av all mat jag tycker om. Cardio är begränsad till 3-4h hundpromenad om dagen samt 15 minuter på löpbandet varannan dag.

Basen i min träning är något som liknar 5x5 i knäböj, marklyft och bänkpress, men jag kompletterar med lite fler och lättare i t.ex. militärpress, dips, rodd och chin-ups.

Mina stats: 44 år, 90kg, 183cm, 140kg bänkpress, 180kg (och 5x170), 155kg knäböj.

Utnyttja "Resurserna" här intill för inspiration. Själva bloggen nedan är annars bara en träningsdagbok samt lite filosoferande om hur man blir en bättre människa.

MOTTO 1: Aldrig kötta - Work smarter, not harder
MOTTO 2: Alltid vila - Träna kort och effektivt 1%, vila 99%
MOTTO 3: Cardio valfritt - Det behövs inte för deff, men en del gillar det. Hoppa över cardio och PW om du tycker att det är tråkigt.

TOO WEIRD TO LIVE, TOO RARE TO DIE