ADAMI, Chris, A new path to Artificial Intelligence, May 5, 2015
<p class="MsoNormal" >ADAMI, Chris, A new path to Artificial Intelligence, May 5, 2015</p>
<p>
intro by ... Kuipers (from Complexity Group)..</p>
<p>
see videos on Complexity Group (eg dbl pendulum expt);</p>
<p>
intro by Aaron...</p>
<p>
ADAMI did post-doc with Mike Brown at Caltech...
studied A-Life...</p>
<p>
fellow of AAAS</p>
<p>
he wants to come back to Calif cuz of the great weather!</p>
<p>
he is from Michigan State Univ...;</p>
<p>
Dept of Microbio and Genetics;</p>
<p>
HAL supposed to be operational in 1997 from HAL Labs in Urbana, Illin. (from 2001 movie)</p>
<p>
HAL = heuristical programmed ALgorithmic comptuer.</p>
<p>
mischievious with grudges...</p>
<p>
he shows terminator... but what we have is a Roomba
and ASIMOV.</p>
<p>
soccer playing mini-robots...</p>
<p>
12M robots... 11M are Roombas.</p>
<p>
self-driving cars; Watson;</p>
<p>
But playing chess and Jeopardy is not intelligence... it is just special-purpose know.</p>
<p>
just brute force with some heuristics.</p>
<p>
cannot put IBM Watson at helm of a car.</p>
<p>
Where are We?</p>
<p>
why don't we have intell. machines?</p>
<p>
what is missing?</p>
<p>
he talks about a 50 year reunion....</p>
<p>
1956 Claude Shannon; John McCarthy, Ray Solomonoff... Minsky etc.</p>
<p>
1958 Simon Newell within 10 yrs...</p>
<p>
will discov impt new math theorem...was the claim</p>
<p>
1967 Minsky... the problem of AI will be solved... soon,etc</p>
<p>
Two types of intelligence...</p>
<p>
Ken Thompson correctly predicted chess. UNIX developer.</p>
<p>
type 1: behave approp given the enviro...
have all the answers ready...
if X then Y...</p>
<p>
he shows AIBO, Watson; self drive car</p>
<p>
Type 2 reason about the enviro...
understand, predict,</p>
<p>
what's the diff...</p>
<p>
type 1... reactive machines...</p>
<p>
brightenberg vehicle....</p>
<p>
has light sensors.
photosensors connected to wheels... it moves toward the light.
just an I/O machine.</p>
<p>
for those who are computer scientists
(he jokes... he shows brain with a label and arrow "brain")</p>
<p>
dorsal stream is the where pathway...
connects directly to motor neurons...</p>
<p>
say if a rock was falling toward me... jump to side...</p>
<p>
ventral stream... the what pathway.
uses memory... try to classify the sensory info.</p>
<p>
(consider daniel kahneman... poss. difft factorization of intell: fast and slow);</p>
<p>
what pathway... LT stored reps...</p>
<p>
a repres is a standin-in for somethng else.</p>
<p>
a charicature...</p>
<p>
R. Magritte (his fave artist... born in same town as Magritte)</p>
<p>
he said "this is not a pipe" Leci n'est pas unepipe."</p>
<p>
reps are internal models of the world.</p>
<p>
enables predictions and plan actions...</p>
<p>
programming reps into computers has proven fiendishly difficult...</p>
<p>
AI in 1970s spheres that you could move around... toy probs.</p>
<p>
Rod Brooks...</p>
<p>
Intell without representaiton...</p>
<p>
do away with reps!</p>
<p>
in AI 47 in 1991.</p>
<p>
situated, embodied agents do not need reps.</p>
<p>
in evol robotics, agents intect with enviro via phys sensors, they are situated in the world.</p>
<p>
sensors go to controller, go to actuators and changes world (enviro) as perceived by sensors.</p>
<p>
Repres-free AI...</p>
<p>
but these are like the robots with photocells.</p>
<p>
The return of repres...</p>
<p>
<b>what evolved is the capcity to buld up internal models... (key idea)..</b>
from Darwin...</p>
<p>
(Nicholas sits in front row... used to work on evol of robotic bodies)</p>
<p>
can evolution create intellig? (again)</p>
<p>
how he WILL try to do this...</p>
<p>
<b>Markov networks...</b></p>
<p>
networks of Hidden Markov Gates (HMGs <b>SLG =stochastic logic gates</b>)</p>
<p>
want to evolve reps...</p>
<p>
what are they?</p>
<p>
Logic gate is just an I/O table.</p>
<p>
see his paper in PLos Comp Bio in 2011 (with J Edlund, Koch, Tononi, Adami)</p>
<p>
fuzzy logic gate...</p>
<p>
with probab table)</p>
<p>
it has 0.5 in some entries of table (in logic table)</p>
<p>
probab finite state machine.</p>
<p>
can SLGs have memory?</p>
<p>
yes... they can if gate's otuput is fed back into the input...</p>
<p>
and if gates are hidden call those HMGs hidden markov gates.</p>
<p>
MWUA... multiplicative weight update algo.</p>
<p>
Markov network can be trained via a reinforcement learning algo called MWUA.</p>
<p>
How are Markov networks evolved?</p>
<p>
rank the algos by how they did... those get to have more babies.
use evolution.
encode connectivity and logic of markov networks into genes</p>
<p>
the expression of these genes creates the network.</p>
<p>
gene 1 has start codon... Nin then Nout and which neurons is it reading from 0,2,6
and to which neurons 6,7
and then the probas that need to be encoded.</p>
<p>
he shows a causal diagram that translates this.</p>
<p>
does it work at all...</p>
<p>
proof of principle</p>
<p>
robot has 6 sensors.</p>
<p>
actuators are wheels 2 actuators (2 wheels)
retina... front, left, and right sensors.</p>
<p>
needto get thru this maze...</p>
<p>
importance of memory...</p>
<p>
starting from a random brain...</p>
<p>
he shows it after 2,000 generations of evolution...</p>
<p>
then shows after 14K generations...</p>
<p>
recall firing pattern is probabillistic...</p>
<p>
it knows to run thru a door... it jitters</p>
<p>
but after 14K does a purposeful seatch (goes entirely to R and to L</p>
<p>
but after 49K gens...</p>
<p>
it remembers where the doorways are.</p>
<p>
so, early on it learns evol of vision and contact snesor utilizaiton</p>
<p>
but then around 15K gens it evolves door sensor utilizaiton...</p>
<p>
it gets info from door sensor...</p>
<p>
only need to remember proprioception in one wheel (to know whether to move L or R)</p>
<p>
no human design involved.</p>
<p>
but do these have reps?</p>
<p>
from paper 2 years ago...</p>
<p>
a new task...</p>
<p>
similar to space invaders...</p>
<p>
the invasers are two difft size blocks.</p>
<p>
goal is to catch the small ones and avoid the large ones....</p>
<p>
define repr of info abt the world goes beyond the image of the worlde</p>
<p>
a pic of the world is not a reps.</p>
<p>
can u think abt it with the eyes close...</p>
<p>
the world is E and the brain is M.</p>
<p>
S are the sensors.</p>
<p>
R = what is in the brain that is NOT in the sensors.</p>
<p>
R = what the brain knows abt the world with eyes closed.</p>
<p>
World States.</p>
<p>
E0 has a sensor been a ctivated</p>
<p>
block is to the left of the agent</p>
<p>
block is moving to the left or to the right...</p>
<p>
does R increase (in our brain model... doing this task)
our networks are v sparse (as are brains)
repres goes up as fitness gos up...</p>
<p>
how do the repsw work?</p>
<p>
we gave this 10 internal neurons...
but only used 6 neurons...</p>
<p>
created edges</p>
<p>
this brain has perfect fitness
(works perfectly on the invaders task (two incoming blocks of difft sizes)</p>
<p>
he convert to it digital logic circuits...</p>
<p>
, these reps are cryptic...</p>
<p>
cannot understand them.</p>
<p>
compressed reps are inherely not redundant...</p>
<p>
reps are internal models of the world
that enable deicisons on state other than sensors.</p>
<p>
How much info, and which concepts are reps... and can be measured.</p>
<p>
so... what is the new path to machine intell and machine consc...</p>
<p>
want to go from simple behavior (like a fish) to us.
must be evolved.</p>
<p>
shows algo evolved in comptuer and used in robot that traverses a circle... inside areas.</p>
<p>
future: want to evolve progressivle larger brains...</p>
<p>
future : our goal is to evolve a brain than can control
a physical robot to explore the world on its own and learn.</p>
<p>
(evolve it in computer then apply it to phys robot)</p>
<p>
he jokes... it take 10 to 15 years to have interesting thots.</p>
<p>
have a backup copy with robot (but not with your adolescent)</p>
<p>
rod brooks... path to evolved intell is NOT easy...</p>
<p>
A Hintze... co constructd this learning algo.</p>
<p>
he is now bldg a new institute at Mich State Univ.</p>
<p>
he says saccading an image is how we learn to recog an image....
saccading around a numeral (is v different from image recog using deep NN)</p>
<p>
</p>
<p>
</p>