Download view presentation - The National Academies of Sciences

Powerful But Limited:
A DARPA Perspective on AI
Arati Prabhakar
Director, DARPA
Artificial intelligence
Three waves of AI technology (so far)
Handcrafted knowledge
Statistical learning
Contextual adaptation
The first wave is still advancing and solving hard problems
The second wave is amazingly effective, but it has fundamental limitations
New research is shaping the third wave
Intelligence is an ability to process information
perceive
rich, complex and subtle information
Intelligence scale
learn
within an environment
perceiving
abstract
to create new meanings
abstracting
learning
reasoning
reason
to plan and to decide
Artificial intelligence is a programmed ability to process information
First-wave AI technologies
Handcrafted knowledge
Engineers create sets of rules
to represent knowledge in
well defined domains
AI systems reason over
narrowly defined problems
Planning tools
Command Post of the Future
Cybersecurity
Expert systems
No learning capability
and poor handling of
uncertainty
Perceiving
Learning
Abstracting
Reasoning
First wave stumbles on natural data
DARPA Autonomous Vehicle Grand Challenge
140 miles of dirt tracks in California and Nevada
2004
# completed: 0
2005
# completed: 5
The problem in 2004
The difference in 2005
Vehicles were able to follow the
GPS waypoints very accurately
but either missed or hallucinated
obstacles ahead
A probabilistic algorithm interpreted
the flood of incoming sensor data to
learn the terrain and map out an
optimal driving surface
Second-wave AI technologies
Handcrafted knowledge
Statistical learning
Engineers create sets of rules
to represent knowledge in
well defined domains
Engineers create statistical
models for specific problem
domains and train them on
big data
AI systems reason over
narrowly defined problems
AI systems have nuanced
classification and prediction
capabilities
No learning capability
and poor handling of
uncertainty
No contextual capability and
minimal reasoning ability
Perceiving
Perceiving
Learning
Learning
Abstracting
Abstracting
Reasoning
Reasoning
Apple Siri
Text
analysis
Image
recognition
Farfade, Saberian, and Li 2015
AlphaGo
Key enablers of second-wave AI
Better machine-learning algorithms
Big data
Neural Nets
Graphical
Models
Deep
Learning
130
Statistical
Models
CRFs
HBNs
MLNs
2010
2013
44,000
exabytes
2015
2020
Decision
Trees
Markov
Models
SVMs
2005
Random
Forests
SRL
AOGs
4,400
Ensemble
Methods
Bayesian
Belief Nets
7,91
0
1,227
Many applications and big markets
Power-efficient processing
1000
KW for 1Bparameter
deep
learning
network
CPU
100
10
GPU
1
2000
2005
2010
2015
2020
Patents 2004-2013
Today’s artificial intelligence is powerful…
Operate highly autonomous platforms
Search the deep web
Manage cybersecurity in real time
Overcome spectrum scarcity
…but limited
Construction worker
in orange safety vest
is working on road
A young boy
is holding a baseball bat
Andrej Karpathy, Li Fei-Fei
Internet trolls cause the AI bot, Tay, to act offensively
Future third-wave AI technologies
Handcrafted knowledge
Statistical learning
Contextual adaptation
Engineers create sets of rules
to represent knowledge in
well defined domains
Engineers create statistical
models for specific problem
domains and train them on
big data
Engineers create systems that
construct explanatory
models for classes of realworld phenomena
AI systems reason over
narrowly defined problems
AI systems have nuanced
classification and prediction
capabilities
AI systems learn and reason as
they encounter new tasks and
situations
No learning capability
and poor handling of
uncertainty
No contextual capability and
minimal reasoning ability
Natural communication among
machines and people
Perceiving
Perceiving
Perceiving
Learning
Learning
Learning
Abstracting
Abstracting
Abstracting
Reasoning
Reasoning
Reasoning
Develop explainable AI
This is a cat:
• It has fur, whiskers,
and claws.
• It has this feature:
New learning
process
Training data
Explainable
model
Explanation
interface
• I understand
why/why not
• I know when it
will succeed/fail
Uncover causal relationships
among 1,000,000s of scientific observations
Early result: An automatically generated web of influences from 1000 research papers
Some third-wave AI technologies
Human-machine symbiosis
Automatic whole-system causal models
Continuous learning
Embedded machine learning
Explainable AI
R&D investments in foundational technology and applications
Broad Operational
Language Translation
(2012)
Domain: Natural language processing
Dragon (Nuance
NaturallySpeaking)
Global Autonomous
Language Exploitation
(2005)
USA CPoF,
USSTRATCOM,
NMCI
Speech Understanding
Research (1971)
Personalized Assistant
that Learns
(2005)
GALE: Over 35
systems deployed
worldwide
Military Foreign language
Translation System POR,
16 DoD organizations, FBI
Low Resource Languages for
Emergent Incidents
(2015)
Communicating
with Computers
(2015)
Big Mechanism
(2015)
Apple Siri
IBM Watson
Deep Exploration and
Filtering of Text
(2012)
DTRA terrorism
indicators
Machine Reading
(2005)
Law Enforcement,
ATF, FBI, NSA
DTRA,
NASIC
Artificial Neural Nets
(1991)
Memex (2013)
XDATA (2011)
DNI, CIA,
Treasury,
FinCEN,
FBI, DEA
Facebook AI (LeCun)
DARPA programs
Direct commercial impact
DoD transition
Deep Learning
(2010)
Baidu Research (Ng)
Robust Automatic
Transcription of Speech
(2010)
AF Rivet Joint,
SOCOM PUMA UAV,
IC, JSOC Traveler,
JIATF-S, Navy EP3