Transcript Presentation - IBM Research

IBM Research
Fusing Animals
and Humans
Jonathan Connell
IBM T.J. Watson Research Center
© 2003 IBM Corporation
Criteria for perceived intelligence
human level
Communicative
Abstract
Social
Personality
Aware
Animate
Can express internal ideas and ingest
situational descriptions, true language
Can conceptualize situations
remote in space and time, planning
Aware of social order, use
other beings as agents
Individuals have different likes and
dislikes, preferences learned over time
Responds and changes actions
based on human-perceptible
environment change
Coordinated movement,
many degrees of freedom
The above seems to be the layered ordering in natural organisms.
Note that language is a uniquely human ability.
2
How to achieve AI?
 Silver bullet: Language
Human veneer on top of base system
→ Construct a language interpreter
• Focuses attention & partitions world
• Enables one-shot “learning”
 Core value: Motivation
Animal substrate underlying control
→ Build symbols & decide how to act
• Needs innate segmentation, comparison, and interest
• Bootstraps to more elaborate concepts
Artificial General Intelligence needs both parts
3
HUMAN LANGUAGE
 Scripting system for sensory-motor subroutines
 Linguistic interpreter needs grounding
 Objects – show examples and give same label (needed for others)
“horse” =
 Properties – show different named objects and give same label
apple
car
cup
house
flower
“red” =
 Relations – show configurations with named objects
 Actions – show temporal sequences with named objects
4
Language & cognition
 Can guide attention, which helps learning
Without – Show lots of contrasting examples
“moth”
“butterfly”
“moth”
With – “No, this is a moth
not a butterfly. Look at its
fuzzy antennae.”
“butterfly”
 Can impart procedures more directly
Without – Trial and
error until suddenly
“Hurray!”
With – “Hold the
jar in your left
hand, grasp the
top with your
right hand, and
twist hard.”
5
Internal dialog
 Sapir-Whorf revised:
– Simply remember speech verbatim
– Replay it through interpreter to actualize
– Cf. Vygotsky’s model of child development
Example:
new driver operating a car
“Okay, the stop sign is coming up.
Slow down and watch for other cars at
the intersection. It the car on the cross
street arrives first he gets to go first. It
looks like there is no one around, so
you can go now …”
6
Compiling patterns
Condense sequences by removing reliable intermediate steps
 Compiling out echoic situations yields direct encoding:
(see: shaggy animal → say: “It’s a dog”) → (hear: “It’s a dog.” → represent: dog)
see: shaggy animal → represent: dog
 Compiling in narrative patterns enhances perception:
(see: bird → hear: “It’s a bird!”) → (hear: “What shape is its beak?” → look: at beak)
see: bird → look: at beak
7
ANIMAL MOTIVATION
Needs:
 Underlying proto-symbolic system
•
•
•
•
Objects – representing spatial-temporal loci
Properties – characterizing the objects
Relations – between objects and places
Reflexes – for generating for motions
 Innate base cases
• Segmentation: color, depth, texture, motion, loudness, pitch
• Comparison: hue, brightness, template match, nearness, acoustic
spectrum
• Interest: Bright lights, loud noises, colorful objects, high motion
 Methods for extending each
• to produce and interpret more complex representations
8
Segmentation
isolated word
acoustic
spectra
HMM
model
GINGER
isolated object
new “isolated” object
located in clutter
visual
pixels
template
model
residual
error
9
Comparison
Checking if some situation matches a precondition:
 Measure intrinsic feature similarity
 Count number of exactly matching features
 Estimate compatibility of parts
straight
brown
black
?
“dog”
at-back
straight
at-front
white
at-back
curvy
“dog”
Brown is close to
black, but both
subparts mismatch
whiskers
X
at-back
“dog”
Brown is an intrinsic
mismatch to white,
but everything else
is exact
black
at-front
black
Decide whether the
beagle example
matches the poodle
or the cat
at-front
10
Interest
 Guides system’s overall activity
 Traditional goal-driven systems are brittle
– Many respond only to human-imposed goals
– Do not spontaneously take the initiative if stuck
– Sit idle if no goal (as opposed to exploring, etc.)
?
 Indirect control systems are more robust
Policies: sets of free-running situation-action rules
Interest: general “goals” in terms of desired situations
Directives: activate policies (K-lines) likely to achieve situation
Affordances: detect that environment offers certain opportunities
11
Autonomous control
<S, E, A>
Affordance
Detector
E
Persistent
Directives
I(x)
Interest
D(x)
K-Line
Policies
Sensors
S
Actuators
A
S = prevailing sensory context
I(x) = system is interested in x
E = exciting affordance sensed
D(x) = system intends to achieve x
A = action selected to take
12
Bootstrap rules
 Walking by pond and hear “splash”!
– Interesting event occurred (splash)
– So remember situation (pond) and event (walk-by)
S & E & A & I(E) → <S, E, A>
 See a pond and interested in splash noises
– Remembered pattern mostly matches current conditions (pond)
– And interested in remaining portion (splash)
– So set a directive to obtain that portion (splash)
<S, E, A> & S & I(E) → D(E)
 Want a splash and near pond
– Intend to obtain condition (splash)
– And current conditions match remembered context (pond)
– So do associated action in record (walk-by)
D(E) & <S, E, A> & S → A
 Want a splash and remember about dropping rocks
– Intend to obtain condition (splash)
– And condition happened in another memory
– So become interested in the context of that memory (drop-rock)
D(E) & <S, E, A> → I(S)
13
Summary
 Human-level AI requires language
– Makes learning classifications faster
– Makes learning procedures easier
– Internalized dialog can guide cognition
 An interpreter can be built on an animal substrate
– Use operant conditioning to obtain grounding
– Needs proto-symbolic structures to work with
 Self-motivation is essential to animals (& humans)
– Actions not necessarily driven by explicit or imposed goals
– Needs some innate segmentations, comparisons, and interests
– Bootstrap procedure can make progressively more complex
14