MITWorld Marvin Minsky Emotion Machine 2007
<p>Dec 27, 2007 4:45 PM
 
MITWORLD Marvin Minsky
 
Emotion Machine  (most of the book is on Marv's home page).
 
Will need AI in 20-50 years cuz will not have enuf young helpers.
whole population is aging.
 
Marv liked projects from 60's and 70's:
Slagel  SAINT:  automatic integrator  (ran on IBM 701 at 20Khz)
Danny Bobrow:  high school geometry
Terry Winograd:  blocksworld  Shrdlu
Pat Winston
Gerry Sussman:  program debugger.
 
AI lost its way in the past 20 years.
 
around minute 20 perhaps
Push Singh PhD Thesis  Multiple Realms of Everyday Thought
 
(girl contemplates boys stack of blocks)
physical:  what'll happen if I push that block
social:  shall I help him or hinder him
spatial: can I reach it
mental:  I forget where I put that block
visual:  can't see it
architectural: can I build an arch.
 
Common sense requires multiple views and interpretations.
 
25:48  FADS in AI
                                  PROBLEM
reinforcement learning:    does not know how to assign credit (to which feature?)
rule-based systems        limited resources
neural nets                    opaque to analysis
statistical inference        no causal reasoning
formal logic                    no analogical reasoning
genetic algorithms           cannot recall failures
insect-like robots            only create more of same
baby machines                do not ever develop into anything
 
HUMANS DO NOT LEARN MUCH FROM EXPERIENCE  (most of what we know we have learned
by being TOLD, SHOWN, or TAUGHT.
 
Feral kids are idiots  (best tale is from 1750  The Wild Boy of Averon   1750.
raised by wolves.
 
Reinforecement learning ... reward  or punish a specific trial.
All 8 above have some small limited applicability:  like to include all
but then  HWO and WHEN to apply each of them.
 
min 30 :Marv:  We Need MEMES to complement our GENES.
 
34:  Evolution only knows what succeeded.
 
Carl Jung ... ludricous popularity... many cultures have similar legends
(Marv... these are just little stories about social blunders that can get you killed.
Same stories will arise whenever you have large number of people living together)
 
How to represent this knowledge?
First guy:  Well,  use logic
Next guy:  NO ... too inflexible
Use neural nets ... No ... too opaque
use predicate calculus ... no;  can only quantify with all, one, some
neural nets ... can't quantify at all.
neural nets ... cannot keep track of instances:
eg is 12 the sum of 5 + 7 pr 54 - 42 etc.  must keep infiv records and
records of counter examples and failures.  Need to know rule and all its important exceptions.
 
Note:  most of time just roll along with inertia;  ONLY THINK when Stuck...
 
CRITIC:Recognize a problem type:  and  then SELCTOR:  activate a  "Way to think"
 
eg:  Try to use ANALOGY to find a previous solution and modify it.
ASK SOMEONE for help
Try to Redescribe problem perhaps in simpler terms or break it down (decompose it)
Try to show opposite is impossible (reductio ad absurdum)
 
Marv had a visitor to the lab and Alpine Guide who put a rolling office chair on a desk to change a lite bulb...
spent 6 weeks in infirmary.
How many mathematicians to change a lite bulb  .... N.
 
Neurologists only know something abt the I/O of brain (vision and motor) and nothing in between.
Have NO knowledge repres to guide them.... just blindly collecting data.
 
eg mirror neurons... they got it wrong ... he thinks these represent goal states.
 
Emotions just activate (and TURN OFF) some resources  Anger (one subset); Love (another subset)
Hunger etc.
 
Anger turns off LT goals;  turns off social diplomacies, turns off high level reasoning
 
Infatuation:  I am so in love with her.  She seems perfect.  I would do whatever she wants...
 
counter each claim:  I have turned off my mind;  I will over-ride the things that I normally need to do.
 
Superego: values, consors, ideals, taboos...
 
Self-conscious emotions
self-reflective thinking
reflective thinking
deliberative thinking
learned reactions
instinctive reactions.
 
Analogy
Planning
Simplify
Reformulate
Simulate
Emotions
 
part of your brain that watches what you're doing and summarizes it.
(we don't do much logic)
 
Stanford Robot: Stanley ... a lot of instinctual behavioral systems. (bottom)
 
He had the book on the web for ten years..
 
Psychology is not Physics (400 brain centers) Occam's razor is wrong for psychology.
Newton could desribe it with 3 laws.
 
Understanding is having multiple ways to deal with multiple aspects of things.
 
Each type of learning theory or knowledge
<b>what kind of problems does it help solve</b>
goals that it might serve
other similar ideas
typical cases for which it is useful
story-like narrative depicting its use
contextual cues to suggest when its relevant
parallel interpretations to other realms.
 
Don't need to emulate the lowest levels of cns
 
synapses have not changed since jelly fish.
higher levels made of cortical columns
 
computers are made of flip-flops (either on or off)
intermed levels of cns try to INSULATE us from the low level props of neurons and synapses.
 
Picture with scripts at top, time series, neural nets, prob infer nets
 
Dreams: like garbage collection;  Crick - some concepts share neurons (get sorted out during sleep); 
 
Why are people smarter than chimps?  if made bigger brain it would get slower (cuz of garbage accumulation)
he thinks credit assignment is a major ability (perhaps that evolved and was decisive) ... some key development was an enabler.
perhaps sl. better credit assignment algorithm.  (perhaps there was this and perhaps 4 or so other key changes).
 
either train it or have 5,000 people on the web teach it by typing stuff in (see Push Singh's work).
 
transcranial magnetic stimulation can turn off brain centers for awhile.
 
Developing AI theories since 1960 and no invitation to Brain Sciences depts.
 
 
 
 
 
 </p>