Upload
alton
View
26
Download
0
Tags:
Embed Size (px)
DESCRIPTION
“ The question of whether computers can think is like the question of whether submarines can swim.” - Edsger W. Dijkstra. Growing Intelligence - Looking beyond year one. Gadi Singer VP and GM, IDGz Architecture Group GM, Israel Development Centers (IDC ) May 7th, 2013. - PowerPoint PPT Presentation
Citation preview
Gadi SingerVP and GM, IDGz Architecture Group
GM, Israel Development Centers (IDC)
May 7th, 2013
Growing Intelligence -Looking beyond year one
“The question of whether computers can think is like the question of whether submarines can swim.” - Edsger W. Dijkstra
Why are we here today?
Page 3ICRI retreat
Almost 1 year ago – ICRI InaugurationTel Aviv Museum of Arts, May 22, 2012
What were we trying to achieve ?
Computational Intelligence
Bringing together ML and heterogeneous architectures to
deliver next generation of intelligent devices that are
efficient, adaptive and always-learning
Page 4ICRI retreat
Were we the first to think about it ?
Page 5ICRI retreat
בן גוריון מתנבא על העתיד
Page 6ICRI retreat
ICRI-CI Year One Retrospective
• Impressive group of researchers, impressive set of projects• Cross-domain / Cross discipline research • High match between the grand vision and institute themes /
projects – ML 2020– Intelligent Agents– Brain Inspired Computing– Accelerators
• Intel Internal - Growing interest regarding the selected ICRI themes
Page 7ICRI retreat
Year Two and Beyond• Computational Intelligence - one of the greatest frontiers– Academia, Industry, Societies
• Requires World-Class research• Impactful research with path to deployed
solutions• Increased collaboration among researchers, and
with Intel• Continuously refine program – e.g., adding
projects– NN based architecture– Agent assist in discussion
Characteristics of “Truly Intelligent” Computing
Page 9ICRI Retreat
Elements of Intelligent Systems
Judges a course of action:
• Rules to govern decisions
• System recommendation
• User approval• Surface relevant
options
Understand the user needs and goals:
• User’s general intent
• Specific goals may be explicitly defined with corresponding actions
Acts on user’s behalf:
• Autonomous action
• Proactive decisions
• Enable users to track actions
Adapts to experiences over time to improve the system:
Collects and synthesized userdata to gainawareness:
• Knowledge of the user
• General Knowledge
• Sensing of the environment
AWARENESS ALIGNED GOALS ACTION LEARNINGDECISIONS
Principle #1 – Brain Inspired Computing
Page 11ICRI retreat
Theories of Perception and Cognition Approaching Viable Implementation
Brain Inspired Computing
Principle #2 – Modular and Open Platforms
Page 13ICRI retreat
“Open & Horizontal” is live and kicking!Source: Bain. *Other brands and names may be claimed as the property of others .
Architecture
Other RISC(IBM)
SPARC
OtherCISC(IBM)
Power
‘90 ‘92 ‘94 ‘96 ‘98 ‘00
75%
Data Centers100%
50%
25%
0
Tablets & PhonesOther
75%
100%
50%
25%
02009 2010 20122011
25%
15%
60%
Platform where capabilities come from modules
provided by individuals, companies, or Academia
Example – Intel’s Perceptual Computing
The Rise of Natural Intuitive ComputingNow Near Future The Vision
Providing Human-like Senses to Computing
Intel® Perceptual Computing SDK Beta
Providing Infrastructure to build on• BETA SDK: Free for Evaluation • Perceptual Modes Support:
– Face Analysis, Tracking– Finger Tracking– Close-Range Hand Gesture Recognition– Voice Processing – 2D, 3D Augmented Reality
• APIs:– High-Level API: For fast, easy
programming– Low-Level API: For innovation and programming control
Principle #3 – Development for a Learning (Evolving) Machine
Page 17ICRI retreat
Design a machine for unforeseeable scenarios
Validate a solution that will evolve and change in the field• What does “correctness” means?• Validate to ensure people’s safety, security, and
privacy
Development for a Learning (Evolving) Machine
?
Open
ing
a do
or
Principle #4 – Significantly Improved Power Efficiency
Page 19ICRI retreat
Efficient Architectures for Perceptual/Cognitive Computing
Watson: Ninety IBM Power-750 servers (plus additional I/O, network and cluster controller nodes in 10 racks)Total of 2880 POWER7 processor cores and 16 Terabytes of RAM. Each Power-750 uses a 3.5 GHz POWER7 8-core processor, 4-threads per core.
Calculation vs. Cognitive
“Invisible” and seamless
Re-imagine power efficient computing
Principle #5 – Structuring for Ethical Choices
Page 21ICRI retreat
“Ethical Computing”• Autonomous == Making Choices• Value system weights on options• High-impact opportunities for good, also imply
risks
“Ethical Module[s]” needs to emerge
Page 22ICRI retreat
Example - Three Laws of Robotics (Asimov, 1940)
1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
2. A robot must obey the orders given to it by human beings, except where such orders would conflict with the First Law.
3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws.
0. A robot may not harm humanity, or, by inaction, allow humanity to come to
harm
Year 2 and Beyond - The Quest
Page 24ICRI retreat
Computing EvolutionThe Decade Cycle• 1980’s – Compute
revolution• 1990’s – Network revolution• 2000’s – Sensor revolution
• 2010’s –
Recognition/Cognition?
The Usage Evolution• Productivity and
Entertainment- Von Neumann Arch has worked well• Interactive computing :
– Traditional Devices struggle to fit the needs (e.g. ASR, Object Recognition etc…).
– Dedicated platforms lead (Gestures, Voice…)
• Relating Computing –– Currently available architectures fail to
cope with the new tasks (NLU, AGI etc…)
Source : IBM DARPA Synapse Project
Page 25ICRI retreat
The Quest: Specialized Cognitive System
• New HW/SW solutions that are optimized for:– Representation: Massively parallel, somewhat
redundant, semantic rich, info storage– Inference/Reasoning: Massively parallel (>>1000),
probabilistic, hierarchical, pattern matching and abstraction
– Learning: Adding new info/patterns through external source (teaching) or introspective ML.
– Power Efficiency: For effective local and distributed computing
Will they create > 100X efficiency in Cognitive Computing uses?
Page 26ICRI retreat
Closing… and opening for a future
• Community (Academia, Industry, Developers) should create content for intelligence competencies
• Define a framework and platform[s] for intelligence computing:1. Brain inspired2. Collaboration through modularity and openness3. Enable and contain machine learning4. Significantly improved power efficiency5. Structure for ethical choices“The coming 3-5 years are about exquisite sensing; the following decade will be about
making sense of the senses.” Gadi Singer
Thank You !
Platf
orm
Ingr
edie
nts
OS
&
Mid
dlew
are
Apps
&Se
rvic
es
System ArchitectureRobotic platformsEmbedded-to-cloudCrowd/embed. arch.
CommunicationsEmbedded interfacesV2V CommunicationVisible light communicationLow-power GSMBurst RFID
Power harvestingGSM / 802.11 / UHFBarometric / Thermal
SensingFirst Person SensingCamera-ProjectorInertial localizationSensor network
Infra. mediatedPrivacy PreservationCompressive cameras
Resource managementData complexity reductionResource-constrained MLPerpetual sensingEnergy eff. data collection
Machine understandingHuman action/intent understandingObject recognitionspeech recognitionDynamic scene understanding Never-ending learningImitation/reinforcement learning for manipulationPrecision locationParallel ML algorithms
Personalized activity modelsMulti-step task reco.Personalized, joint speech/gesture reco.Multi-sensor / multi-person activity inferenceActivity learning by demonstrationScene understandingStress recognition
HCIHuman robot interactionGoal-driven labelingTask assistanceMobile persuasion
Focus application areasRetailAutomotiveHome
Interactive task assistanceFamily coordinationMobile health and wellness
Intelligent System – Research ingredients
Human Brain CompetenciesVision, Audition, Touch, Proprioception, Cross-Modal
Perception
Physical Skills, Tool Use, Navigation, ProprioceptionActuation
Implicit, Working, Episodic, Semantic, ProceduralMemory
Imitation, Reinforcement, Dialogical, Written, Experimental
Learning
Deduction, Induction, Abduction, Causal, PhysicalReasoning
Tactical, Strategic, Physical, SocialPlanning
Visual, Social, BehavioralAttention
Subgoal Creation, Affect BasedMotivation
Emotional Expression, Understanding Emotions, Perceiving Emotions, Control of Emotions
Emotion
Self Awareness, Theory of Mind, Self Control. Other-Awareness, Empathy
Modeling Self and Other
Appropriate Behavior, Social Communication, Social Inference, Cooperation
Social Interaction
Verbal, Gestural, Pictorial, Language acquisition, Cross-Modal
Communication
Counting Objects, Grounded Small Number Arithmetic, Comparison of Quantitative Properties, Measuring with tools
Quantitative
Physical Construction w/ Objects, Formation of Novel Concepts, Verbal Invention, Social Organization
Building/Creation
Source: Ben Goerzel, AGI 2011