Transcript 10 6 neuron

Systems of Neuromorphic Adaptive Plastic Scalable Electronics
Bidder’s Workshop and Teaming Meeting
March 4, 2008
Dr. Todd Hylton, Program Manager
DARPA DSO
Approved for Public Release, Distribution Unlimited
1
Introduction and Motivation
Approved for Public Release, Distribution Unlimited
2
Motivation and Objective
Problem
• As compared to biological
systems, today’s intelligent
machines are less efficient
by a factor of a million to a
billion in complex
environments.
• For intelligent machines to
be useful, they must
compete with biological
systems.
Objective
• Develop electronic,
neuromorphic machine
technology that scales to
biological level.
Human Cortex
Simulated Human
Cortex
15 Watts
1010 Watts
I Liter
4x 1010 Liters
von Neumann
Machines
[log]
A trade between
universality and
efficiency
Machine
Complexity
e.g. Gates;
Memory;
Neurons;
Synapses
Power;
Size
Neuromorphic
Machines
• Human level performance
• Dawn of a new age
Dawn of a new
paradigm
“simple”
Program Objective
“complex”
[log]
Environmental Complexity
e.g. Input Combinatorics
Lansner et al
The SyNAPSE program seeks to break the programmable machine paradigm
and define a new path for creating useful, intelligent machines
Approved for Public Release, Distribution Unlimited
3
Vision & Impact
Historical Evolution of Modern Electronics
Transistor
IC
µProcessor
& memory
• End of scaling
• Defect intolerant
• Architectural bottleneck
• Software limited
• No path to biologically
competitive intelligence
Programmable
machines
60 years
DARPA SyNAPSE
Vision for the Future
Electronic
Synapse
“Cortical”
Microcircuit
“Cortex”
Fabric
<<60 years
Intelligent
machines
• Increased component density
• Increased component function
• Defect tolerant
• Neuromorphic information,
learning, cognition,
understanding architecture
• Path to biologically competitive
intelligence
The SyNAPSE program seeks to extend the development of modern electronics
into a new revolutionary new era using a similar paradigm.
Approved for Public Release, Distribution Unlimited
4
Inspiration
Biological-Scale Neuromorphic Electronic Devices
Human NeoCortex
Neuromorphic Electronics
~1010
1010 intersection/cm2 in crossbar
arrays w/ 100 nm pitch
~106
synapses/cm2
~5x108 transistors/cm2 in state of the
art CMOS
Neurons/cm2
~5 x 108 long range axons
@ ~1 Hz
~30 Gbit/sec multiplexed digital
addressing
Conclusion: Gross statistics of biological neural systems might
be realized in modern electronics.
Approved for Public Release, Distribution Unlimited
5
Key Challenges and Goals
Approved for Public Release, Distribution Unlimited
6
Key Goal: Electronic Synapse
Axonic electrode
Dendritic electrode
Crossbar synapse
Soma
The electronic synapse performs computation, memory, and adaptation in a neuromorphic
system. Computation occurs in the electron current (i=v*g) injected through the synapse
conductance g between neurons in response to (spike) voltage v. Memory occurs as a
slowly changing electrophysical property that modifies g. Neuromorphic adaptation (aka
plasticity) occurs as g changes in response to the same voltages used for computation.
Approved for Public Release, Distribution Unlimited
7
Key Goal: Spike Time Dependent Plasticity
Pulse interference
at the synapse
synaptic potential
Δt
Post-synaptic
Neuron
tpre
tpost
time
% change in synaptic conductance
Pre-synaptic
Neuron
t+
0
t-
0
Neurons encode information as “spikes” and
communicate to other neurons in both both
forward (axonic) and backward (dendritic)
directions. The time-relation between forward
and backward spikes arriving at a synapse
determines if the synaptic connection should be
increased or decreased. Connection strength
increases (decreases) whenever forward spikes
are causally (acausually) correlated to
backward spikes.
Δt = (tpre – tpost)
Approved for Public Release, Distribution Unlimited
8
Key Goal: Neuromorphic Architecture
• Possible approaches
– “Bottom-up” based on neuro-psychophysical models of biological systems
– “Top-down” based on large scale neuroinformatics / connectomics
– Artificial Neural Networks
– First principles design
– “Evolutionary” optimization of model
structures
Approved for Public Release, Distribution Unlimited
9
Key Goal: Electronic Implementation
• Chip fabrication
– Novel materials and structures on CMOS
• Spike processing
– Spike time encoding
– Spike time dependent plasticity
• Connectivity
– Hardwired
– Addressed / programmable
– On-chip / off chip
• Power
• Size
• Supports Neuromorphic Architecture
Approved for Public Release, Distribution Unlimited
10
Key Goal: Large Scale Simulation
• Using programmable machines to design and test
intelligent machines
– Architectural design, validation, development
– Chip design / validation
– Mammalian scale simulations of systems and components
– Functional performance testing in environments
• Large scale digital hardware
– “Supercomputer” scale
– Specialized hardware development may be appropriate
– Rebuilding the current computer architecture “from scratch”
is outside the scope of this solicitation
Approved for Public Release, Distribution Unlimited
11
Key Goal: Training & Evaluation Environments
(Image removed)
• Train and evaluate machine intelligence across capabilities found in mammalian
species (106 range of brain size)
• Virtual environment for the evolution of intelligent machines
• Fill long-standing need for authoritative machine intelligence evaluation
Approved for Public Release, Distribution Unlimited
12
Approach: Training & Evaluation Environments
Task Area
Sensory Perception
(Image removed)
Decision & Planning
(Image removed)
Navigation & Survival
(Image removed)
Features
Cognitive Area
• Identification/classification of spatiotemporal objects in animation or
video
• Multi-dimensional complexity
variability
• Core task of all cognitive systems
• Quantitative measures of complexity
• Objective measures of performance
• Easily scaled
• Human interaction
• “Abstract” cognition
• Interaction in complex, dynamic
environments.
• Comparison to small animal studies
• Exercises all levels of cognition
• Most difficult to score and scale
Approved for Public Release, Distribution Unlimited
13
Disciplinary Integration Challenge
Materials & Physics
• Crossbars
• Electronic Synapses
• CMOS Integration
Theory
• Information
• Computation
• Communication
• Cognition
• Learning
Computer Science
& Electrical Engineering
• Large Scale Computation
• CAD Tools
• Design Validation
• Electronic Architecture
Disciplinary Gap
Neuroscience
• Neuroinformatics
• Neurophysiology
• Neuroanatomy
• Neural models
• Neural simulation
• Animal models
(Image removed)
VLSI CMOS
• Device Design
• Analog-Digital
• Asynchronous
• Sub-threshold
neuromorphic
• Fabrication
• Test
• Packaging
SyNAPSE must bridge the disciplinary gap
Approved for Public Release, Distribution Unlimited
14
Program Plan and Milestones
Approved for Public Release, Distribution Unlimited
15
Program Approach
Model
System (SyNAPSE)
Modules
(e.g. visual cortex)
Top-down
(simulation)
Make
Networks
(e.g. cortical column)
Biological Scale
Machine Intelligence
Measure
Employ theoretical and empirical approaches
constrained by practicality.
Circuits
(e.g. center-surround)
Architecture
Components
(e.g. synapse / neuron)
Bottom-up
(devices)
Simulation
Hardware
Materials
(e.g. memristors)
Environment
Attack the problem “bottom-up” and “topdown” and force disciplinary integration with a
common set of objectives.
Sponsor a suite of complementary capabilities
to build, train, and evaluate devices.
Approved for Public Release, Distribution Unlimited
16
Program Components
• Hardware will likely include CMOS devices, novel synaptic
components, and combinations of hard-wired and
programmable/virtual connectivity and will support critical information
processing techniques like spike time encoding and spike time
dependent plasticity.
• Architectures will support critical structures and functions observed in
biological systems such as connectivity, hierarchical organization, core
component circuitry, competitive self-organization, and
modulatory/reinforcement systems.
• Large scale digital simulations of circuits and systems will be used to
prove component and whole system functionality and to inform overall
system development in advance of neuromorphic hardware
implementation.
• Environments will be evolving virtual platforms for the training,
evaluation and benchmarking of intelligent machines
Approved for Public Release, Distribution Unlimited
17
Phase 1
Hardware
Component
synapse (and
neuron)
development
CMOS process
and core
circuit
development
Microcircuit
architecture
development
Preparatory
studies only
Environment
Emulation
& Simulation
Phase 0
Architecture
& Tools
Program Outline
Preparatory
studies only
Phase 3
Phase 4
CMOS process
integration
~106 neuron
single chip
implementation
“Mouse” level
~108 neuron
multi-chip robot
at “Cat” level
System level
architecture
development
~106 neuron
design for
simulation and
hardware layout
~108 neuron
design for
simulation and
hardware layout
Simulate large
neural
subsystem
dynamics
“Mouse” level
benchmark
(~ 106 neuron)
“Cat” level
benchmark
(~ 108 neuron)
Build Sensory,
Planning and
Navigation
environments
Add Audition,
Proprioception
and Survival
“All mammal”
complexity
Add Touch and
Symbolic
environments
“Small mammal”
complexity
Phase 2
Comprehensive
design
capability
Sustain
Program Phases 1-4 may be combined per the BAA instructions
Approved for Public Release, Distribution Unlimited
18
Phase 0 Go No-Go Metrics
Hardware
• Synaptic density scalable to > 1010/cm2
• Operating speed >10 Hz
• Consumes < 10-12 Joules per synaptic operation (at scale)
• Dynamic range of synaptic conductance > 10 with >3 bit resolution
• Synaptic conductance increase >1%/pulse for presynaptic spike applied
somewhere within 80-1 msec before a postsynaptic spike
• Synaptic conductance decrease >1%/pulse for presynaptic spike applied
somewhere within 80-1 msec after postsynaptic spike.
• 0%-0.02% conductance decrease if presynaptic spike applied > 100 msec
before or after postsynaptic spike
• Maintains performance over 3 x 108 synaptic operations
Architecture
• Specify and validate by simulation the function of core microcircuit
assemblies using measured synaptic properties.
• The microcircuits must support the larger system architecture and support
spike time encoding, spike time dependent plasticity, and competitive neural
dynamics.
Approved for Public Release, Distribution Unlimited
~ 9 months
19
Go/No-Go Milestones Set 1
Hardware
• Demonstrate all core micro-circuit functions in hardware
• Specify a chip fabrication process supporting the architecture with >1010 synapse/cm2
and >106 neurons/cm2
Architecture
• Demonstrate a complete neuromorphic design methodology that can specify all the
components, subsystems, and connectivity of a complete system.
• Specify a corresponding electronic implementation of the neuromorphic design
methodology supporting > 1014 synapses, > 1010 neurons, mammalian connectivity, < 1
kW, < 2L
Simulation
• Demonstrate dynamic neural activity, network stability, synaptic plasticity and selforganization in response to sensory stimulation and system-level
modulation/reinforcement in a system of ~ 106 neurons modeled on mammalian cortex
Environment
• Demonstrate virtual Visual Perception, Decision and Planning, and Navigation
Environments with a selectable range of complexity corresponding roughly to the
capabilities demonstrated across a ~104 range in brain size in small-to-medium
mammalian species
Approved for Public Release, Distribution Unlimited
20
Go/No-Go Milestones Set 2
Hardware
• Demonstrate chip fabrication of >1010 synapse/cm2, >106 neurons/cm2
Architecture
• Design a neural system of ~106 neurons and ~1010 synapses for simulation
testing
• Design a corresponding single chip neural system of ~106 neurons and ~1010
synapses
Simulation
• Demonstrate a simulated neural system of ~106 neurons performing at
“mouse” level in the virtual environment
Environment
• Expand the Sensory Environment to include training and evaluation of
Auditory Perception and Proprioception
• Expand the Navigation Environment to include features stressing Competition
for Resources and Survival
• Demonstrate a selectable range of complexity corresponding roughly to the
capabilities demonstrated across a ~106 range in brain size mammalian
species
Approved for Public Release, Distribution Unlimited
21
Go/No-Go Milestones Set 3
Hardware
• Fabricate a single chip neural system of ~106 neurons and package into a
fully functioning assembly. Show “mouse” level performance in the virtual
environment.
Architecture
• Design a neural system of ~108 neurons and ~1012 synapses for simulation
testing
• Design a corresponding single chip neural system of ~108 neurons and ~1012
synapses
Simulation
• Demonstrate a simulated neural system of ~108 neurons performing at “cat”
level
Environment
• Add Touch to the Sensory Environment
• Add Symbolic Environment
Approved for Public Release, Distribution Unlimited
22
Final Metric – Milestone Set 4
Hardware
• Fabricate a multi-chip neural system of ~108 neurons and instantiate
into a robotic platform performing at “cat” level
Approved for Public Release, Distribution Unlimited
23
Proposal Technical Requirements
Approved for Public Release, Distribution Unlimited
24
Proposal Requirements (1)
• Describe an approach to developing an integrated neuromorphic
architecture serving as a foundation for the development of intelligent
machines.
– Describe the base components of your architecture and their function.
These base components may be the analogs of biological neurons,
synapses and/or small assemblies of such elements. Describe the
computational, communication and learning functions of these base
components.
– Describe one or more core micro-assemblies of the base components and
their corresponding function.
– Describe your approach for developing functional assemblies from the core
assemblies. These assemblies should provide core cognitive functions such
as sensory perception, motor control, executive control and others.
– Describe your approach to integrate functional assemblies into complete
cognitive systems including sensory perception, declarative learning and
memory, procedural learning and memory, executive control, and motor
function.
– Describe any plan to incorporate neuro-anatomical/physiological data into
the architecture.
Approved for Public Release, Distribution Unlimited
25
Proposal Requirements (2)
• Describe a high-level, conceptual electronics
implementation capable of supporting the neuromorphic
architecture of (1) having
– 1010 neurons
– 1014 synapses
– operating with temporal dynamics comparable to biological
systems
– total power <1kW
– total volume <2L
– interfaces for sensory inputs and motor outputs
Approved for Public Release, Distribution Unlimited
26
Proposal Requirements (3)
• Describe an approach to developing nanometer-scale, plastic
synaptic components consistent with (1) and (2). Multiple
approaches are encouraged for this task.
• Describe an approach to developing electronic neuronal
processing units (neurons) consistent with (1), (2) and (3).
• Describe an electronic coding, communication and synaptic
update scheme consistent with (1), (2), and (3).
• Describe a plan of computer simulation/emulation to enable the
near real-time simulation of neuromorphic systems up to 108
neurons and 1012 synapses.
• Describe a plan to obtain and import descriptions of neural
systems from neuro-biological databases (as appropriate).
• Describe key technical challenges and approaches to achieving
these goals and any other items in the critical path.
Approved for Public Release, Distribution Unlimited
27
Proposal Requirements (4)
• Describe an approach for developing a virtual training and evaluation
environment comprised of the following tasks.
– A Planning and Decision (Game) Task that provides quantitative measures
of complexity and objective and comparative measures of performance;
– A Sensory Perception Task that provides quantitative measures of
performance of identification/classification of spatio-temporal objects in
animation or video;
– A Navigation Task that captures the challenges confronted in navigating in
complex, dynamic environments. The purpose of this task is to evaluate a
collection of cognitive capabilities and to provide a point of comparison to
animal studies.
• Describe a means to scale the complexity of these tasks over the entire
range of mammalian intelligence (~106 range in brain size).
• Describe a capability for hosting the environment including hardware,
software and system support.
• Describe an interface for interacting with the environment.
Approved for Public Release, Distribution Unlimited
28
Proposal Requirements (5)
Environmental tasks will require
• Adaptation in dynamic, uncertain, probabilistic environments that include
partial, erroneous and sometimes contradictory information
• Response times that force speed-accuracy tradeoffs
• Knowledge Integration over
– Different sources and times of knowledge acquisition; and
– Multiple levels of perception, planning and reasoning.
• Interaction with other (human or machine) agents.
• Feedback based on
– Reinforcement of generic, high-level goals
– Supervision using a tutor (learning mode)
• Scalability to match system complexity and support incremental learning
• Scoring to provide quantitative measures of performance
• Benchmarking to provide comparative measures of performance.
Approved for Public Release, Distribution Unlimited
29
Proposal Evaluation
Approved for Public Release, Distribution Unlimited
30
Evaluation Criteria
1)
2)
3)
4)
Ability to Meet Go/No-Go Metrics
Scientific and Technical Merit
Value to Defense
Management Approach and
Proposer’s Capabilities and Related
Experience
5) Cost and Schedule Realism.
Approved for Public Release, Distribution Unlimited
31
Ability to Meet Go/No-Go Metrics
• The proposal establishes clear and well defined research
go/no-go metrics to be used as exit and entry criteria for
Government approval to progress through phases of the
proposed effort.
• The feasibility and likelihood of the proposed approach
for satisfying the program go/no-go metrics are explicitly
described and clearly substantiated.
• The proposal reflects a mature and quantitative
understanding of the proposed go/no-go metrics, the
statistical confidence with which they may be measured,
and their relationship to the concept of operations that
will result from successful performance
Approved for Public Release, Distribution Unlimited
32
Scientific and Technical Merit
• Proposers must demonstrate that their proposal
is innovative and unique, that the technical
approach is sound, that they have an
understanding of critical technical issues and
risk, and that they have a plan for mitigation of
those risks.
• A significant improvement in capability or
understanding above the state of the art must be
demonstrated.
• All milestones must be clearly and quantitatively
described.
Approved for Public Release, Distribution Unlimited
33
Value to Defense
• Proposers must demonstrate the longterm potential of successful research to
radically change military capability or
improve national security with a clear
statement of the goals of their program,
and a quantitative comparison with
existing technology as appropriate.
Approved for Public Release, Distribution Unlimited
34
Management Approach and Proposer’s Capabilities
and Related Experience
• The appropriateness, effectiveness, and reliability of the
management structure are appropriate to the diversity of tasks,
technologies and partnering strategy.
• The qualifications of Principal Investigator and key Task Leaders are
appropriate and support the overall management plan.
• The qualifications of the proposer’s key personnel are of adequate
range, depth, and mix of expertise to address all technical and
programmatic aspects of the proposal.
• The proposer's prior experience in similar efforts must clearly
demonstrate an ability to deliver products that meet the proposed
technical performance within the proposed budget and schedule.
• The proposed team has the expertise to manage the cost and
schedule.
• Similar efforts completed/ongoing by the proposer in this area are
fully described including identification of other Government sponsors.
Approved for Public Release, Distribution Unlimited
35
Cost and Schedule Realism
•
The objective of this criterion is to establish that the proposed costs are
realistic for the technical and management approach offered, as well as
to determine the proposer’s practical understanding of the effort. This
will be principally measured by cost per labor-hour and number of
labor-hours proposed.
•
The evaluation criterion recognizes that undue emphasis on cost may
motivate proposers to offer low-risk ideas with minimum uncertainty
and to staff the effort with junior personnel in order to be in a more
competitive posture. DARPA discourages such cost strategies.
•
Cost reduction approaches that will be received favorably include
innovative management concepts that maximize direct funding for
technology and limit diversion of funds into overhead.
•
The proposer’s abilities to aggressively pursue performance metrics in
the shortest timeframe and to accurately account for that timeframe will
be evaluated, as well as proposer’s ability to understand, identify, and
mitigate any potential risk in schedule
Approved for Public Release, Distribution Unlimited
36
Administrative Items
Approved for Public Release, Distribution Unlimited
37
Approved for public release, distribution unlimited
BAA Solicitation Schedule
• BAA 08-28
– Estimated posting date – March 17, 2008
• Proposal Due Date
– May 2, 2008, no later than 4:00PM EST
– BAA will remain open for 1 year
• Anticipated Contract Award
– August 2008
Approved for Public Release, Distribution Unlimited
38
Approved for public release, distribution unlimited
Proposal Format
• Proposals must consist of two volumes-technical and
cost.
• Technical- Maximum of 55 pages including references,
tables, and charts. Please do not include separate
articles or CDs as these will not be used in the review
process.
• Cost-contains a cover sheet, detailed cost break down,
and supporting cost and pricing information.
• For detailed description of proposal format see the BAA at
http://www.darpa.mil/baa/BAA08-28.html
Approved for Public Release, Distribution Unlimited
39
Approved for public release, distribution unlimited
Other Comments on the Proposal
• DARPA requests proposals for the full scope of
development
– All proposals must address all of the technical areas listed in
the BAA
– Proposals addressing only individual components of the
overall program will be considered non-responsive
• Coherent integration and management of
multidisciplinary research organizations is required.
• Structure proposals to reduce risk early and to give
the government flexibility in task/phase funding
Approved for Public Release, Distribution Unlimited
40
Approved for public release, distribution unlimited
Teaming Website
• http://www.sainc.com/SyNAPSETeaming/index.asp
A teaming website has been created to facilitate the organization of teams to
address all program component areas.
Approved for Public Release, Distribution Unlimited
41
Approved for public release, distribution unlimited
Discussion
Discussions are
strongly encouraged during
teaming and proposal formulation.
Please submit questions by noon so that they may
be answered during the FAQ segment of the
workshop.
Approved for Public Release, Distribution Unlimited
42