Transcript slc.catalyst12 - Rensselaer Polytechnic Institute

Human-Level
Machine Learning
Selmer Bringsjord, Nick Cassimatis,
Kostas Arkoudas, and Bettina Schimanski
Department of Cognitive Science
Department of Computer Science
Rensselaer Polytechnic Institute (RPI)
Troy NY 12180 USA
December 9 2004 @ NSF
RAIR Lab Sponsors
Deontic/Doxastic
Reasoning
-Cracking Project;
“Superteaching”
QuickTime™ and a
TIFF (Uncompressed) decompressor
are needed to see this picture.
hypothesis generation;
AI in support of IA
“Poised-For” Learning
Slate (Intelligence Analysis)
test generation
synthetic characters/psychological time
advanced synthetic
charactrs
Overview
• The Problem:
– Machine learning is dominated by forms of learning that are
impoverished relative to the human case.
– Humans often learn by leveraging an ensemble of “preestablished” heterogeneous reasoning mechanisms and vast
amounts of prior knowledge.
• Solution/Goal:
– Formalize human learning and rich cognitive mechanisms that
underlie and enable it.
– Implement these formalizations to produce “human-level”
machine learning, and corresponding applications.
• Applications:
– Software and robotic applications; in our case, specifically
• Homeland defense/intelligence analysis tools
• Elder-care robots that are quickly adapt to their owners
– Improve learning in humans:
• Intelligent tutoring systems in math,/logic/computer science
• More precise understanding of learning disabilities for less
traumatic interventions
Formal Models of Human-Level
Learning Can Help Close Learning Gaps
• Learning gaps (esp in math) between:
– US and other countries
• The latest PISA and TIMSS point to an outright crisis!
– 12.7.04 WSJ
– High-achieving and low-achieving students within US
– High-achieving and low-achieving schools within US
• A precise, formal understanding of learning would
enable us to
– pinpoint the factors that enable rapid, explosive learning;
– build machines able to augment human teaching (which for
various reasons is failing) in the math/logic/comp sci area
Machine Learning Today:
Costly Trial and Error
•
Traditional machine learning:
–
–
–
–
•
Learn only after many repetitions of trial and error
Stuck on function-based model
E.g., Language: WSJ Corpus, 1987-1989, with 39 million words
Explanation-Based Learning uses only primitive reasoning/knowledge
Hurts with applications:
– Trial and error not good in cases where errors kill
• Medical robotics
– Thousands of learning trials can be expensive
• Acquainting a robot with a new hospital would take days
• Teaching people new software makes them less productive in the short-term.
Machines train us now instead of us training them.
– Learning trials often not available
• Homeland security: Not thousands of people in flight schools
– Robots and software therefore limited to narrow tasks and inflexible
– We are forced to assemble machine knowledge manually
• CYC has over a million facts and is not even remotely complete
Some Motivating Examples...
Millions of students are currently learning primarily by
reading -- and ditto e.g. for adult researchers like us!
Example 1: Suppose You Were Tasked
to Learn About Astronomy!
The scorpion lies between Libra and Sagittarius in the Milky Way.
It is not hard to imagine this pattern of starts resembling a scorpion,
with its claws and stinging tail. An arc of stars marks the curve of its
raised tail and the fiery red star Antares lies at is heart...
Example 2: Human One-shot Learning
(a simple example)
USB
CONVERTOR
CUP
QuickTime™ and a
H.263 decompressor
Insert
here
are needed
to see movie
this picture.
(Nick has a copy)
The traditional machine learning
approach...
Behavior of Micro-PERI
Implications of One-Shot
Learning and Learning by Reading
• Learning by reading and one-shot learning examples require:
– Rich set of representation and reasoning abilities early on
•
•
•
•
Where was speaker looking when he said “USB Converter”.
Social reasoning to track where speaker was looking.
Spatial and temporal reasoning to infer what he was looking at.
Diagrammatic reasoning
– Existing machine learning algorithms have no notion of space, time or human
attention.
– Statistical generalization just one of several learning strategies; also need:
•
•
•
•
Inference (deductive, abductive, inductive, ...) from single group of percepts
Analogy
Imitation
Instruction
– Learning much more socially and physically interactive.
• Ask questions: Why? How? What if? Physically test their own hypotheses about
the world.
• And, in learning by reading...
– the best learners are those who “pre-test” themselves, and hence acquire
“poised-for” knowledge that marks true learning
To Solve the Problem:
A New (5-step) Research Program
1
Without flinching, study the human case -- humans (including kids) who learn
rapidly, including learning by reading
– Developmental psychology has shown that even infants and toddlers have rich notions
of:
• Time, place, causality, belief, desire, attention, number, etc., and of inference over these
concepts
2
3
Develop formal theories that show how to use these factors to make learning faster
and more effective
Develop machine learning algorithms using these formalizations that learn by:
–
–
–
–
–
4
Explicit reading and instruction
Analogical reasoning
Deduction, Abduction, etc.
Imitation
Visual reasoning
Build applications from these algorithms that have broad impact
– Elder care
– Homeland security
5
Trace out the implications of these algorithms for better teaching/learning in the
human sphere, particularly in mathematics/logic instruction
– address “Math Gap”
– including intelligent tutoring systems and synthetic characters
Our Approach Forges a Bridge
SBE
Behavioral &
Cognitive Sciences
CISE
?
?
Artificial Intelligence and
Cognitive Science
The Right Time:
Resurrection of Human-Level AI
•
Recognition of need for human-level AI and integrated cognitive systems growing:
– Dedicated issue of AAAI’s journal of record (AI Magazine) to be devoted to human-level
AI
• Cassimatis editor, Bringsjord, Arkoudas, Schimanski contributors
– AAAI Fall Symposium on Integrated Cognition (Cassimatis led)
– “Grand Cognitive Challenges” under discussion @ DARPA’s Learning-Focused IPTO
• “Psychometric AI” a candidate
– Hundreds of studies in infant cognition give us a good idea of what the right substrate is.
•
•
Integrated cognitive models exist and are advancing every day
Computational infrastructure there:
– Abundant computational power for multiple methods in one system
– Formal methods exploding with new power (e.g., Athena)
– Robot and machine vision infrastructure in place:
•
•
•
•
Object recognition
Face recognition, eye-tracking
Mobility and navigation
Robot manipulation
Applications
Some Applications
•
High-stakes applications where trial and error too dangerous.
– Homeland security.
– Hazardous waste removal.
•
Robots and software for less sophisticated or learning-challenged humans use them.
– Disabled.
– Elder care.
• Elder-care robots easier to use by the older set.
• Emerging Robotics Technologies & Applications Conference Proceedings, March 9-10, 2004,
Cambridge, MA
– Rodney Brooks mentioned Elderly Care as one of the current future trends in robotics:
» Currently: None
» Future: Robotic Assistants in Millions of Households
•
•
Less brittle, more general, easier-to-learn and use robots and software.
Better learning environments:
– Direct/instruct robots (PERI)
– More accurate pinpoint causes of problem learning.
A catalyst grant for ...?
•
•
Carry out proof-of-concept version of entire 5-step research agenda
Build team to implement this sequence
– part of team that would presumably power full SLC on Human-Level
Machine Learning
•
Build proof-of-concept
– p-o-c would run all the way through our proposed 5-step R&D sequence,
start to finish
– application/implementation:
• homeland defense
• Elder care robot
• ITS for math/logic/comp sci
•
•
•
•
Workshops/Symposia
Conference presentations
Publications
Web site from the very start
END
Objection
• How is this an improvement over GOFAI? i.e., Why isn’t this the 1970s all
over again?
– Less knowledge of human learning then
– Formal methods in their infancy
• Nothing like Athena (used to prove a good part of Unix sound)!
• Like two-layer neural networks compared to bigger ones
– Formal infrastructure was fragmented. Not known how to combine logical and
probabilistic knowledge?
– So researchers were either using no representation and reasoning substrate or
they were using the wrong one.
– Integrated cognitive models for combining methods not developed,
• Polyscheme, ACT-R, ...
– These techniques were not interactive.
• No question asking
• No tracking or reasoning about human intent
• No experimentation
PERI
Psychometric Experimental Robotic Intelligence
• Scorbot-ER IX
• Sony B&W XC55 Video
Camera
• Cognex MVS-8100M
Frame Grabber
• Dragon Naturally
Speaking Software
• NL (Carmel & RealPro?)
• BH8-260 BarrettHand
Dexterous 3-Finger
Grasper System
Our Assets
• Background in intersection of reasoning and
formal methods, and learning
– Bringsjord, Cassimatis, Arkoudas, and Schimanski
• Prior R&D in logic-based machine learning.
– Bringsjord, Arkoudas
• Background in child development.
– Cassimatis
• Integrated cognitive models
– All four
• Background in robotics
– Cassimatis, Bringsjord, Schimanski
Prior Related Work on One-Shot
Learning
• There isn’t anything that maches up perfectly.
• But, related, we have:
Prior Related Work on Learning by
Reading
• Ask for pointers from Ken Forbus...
Impact on Machine Learning and AI
• More flexible and resourceful learning and reasoning
algorithms
• Intellectually flexible robots (again, e.g., PERI)
• Quantum leap in machine learning
• Learning in situations that were impossible before
• Integration of reasoning community back into
learning community
• Impact back on education, including machine-assisted
education (e.g., intelligent tutoring systems &
synthetic characters)
Impact on Study of Human Learning
• Existing empirical work hampered by vague theories
that make results of simple experiments controversial.
– Formal theory should help this
• Develop better understanding of which instruction or
learning techniques are best in which circumstances.
• More specifically:
– Will produce new pedagogy linking learning to reasoning
(mathematics/logic a beneficiary)
– Will produce revolutionary advances in intelligent tutoring
systems, synthetic characters/simulation)