Ed Safford III MetroCon 2015 Verification, Validation, and Deployment of Hybrid Neuromorphic Systems...

Post on 15-Apr-2017

186 views 0 download

Transcript of Ed Safford III MetroCon 2015 Verification, Validation, and Deployment of Hybrid Neuromorphic Systems...

1

“Verification, Validation, and

Deployment of Hybrid

Neuromorphic Systems”

E.L. (Ed) Safford III

10-22-2015

2

Some Old Ideas Revisited

Robotics

Drones

Pulse Codes – Spikes

Nonlinear Stochastic Theory

Cybernetics

NOT NEW

3

Past

Limited by our machines, we have obtained operational

capability by:

a) Constrained Inputs (Information Loss)

b) Simplified Computational Models

c) Tasks Assigned to Machine

V&V Has Been Relatively Mature for These Systems:

a) Safety

b) Security

c) Mission Critical

We put “Man-in-the-Loop” for Additional Adaptability

and Assurance

4

Future

New Machines Will Provide Better Performance

Systems Will Become Larger and More Complex

Operational Demands Will Increase:

a) Improved System Adaptability

b) Greater Machine Autonomy

c) Better Coverage of “Must Not”

5

V&V Problem

Assurance Methods Have Not

Adequately Progressed to

Support The Full Potential of

These Systems For

Deployment

Therefore We Have Not Been

Able to Delegate Certain

Desired Levels of Operational

Autonomy

6

Idealized Brain Neural Network

- from m1.behance.net

7

Simplified Brain Neuron

SYNAPSES are where Axon termini

form junctions with Dendrites

8

Neuron Model

Weighted Inputs Outputs

Activation Function

Summation

Neuron

9

Artificial Neural Networks

1. Composed of one or more “layers” of

“neurons”, i.e. neuron models, with nodes and

connections

2. There is an input layer, an output layer, and

one or more additional layers called “hidden”

layers. May have feedback loops.

3. Mathematically can be represented by

composition of functions with known or

unknown parameters. DEs for feedback.

Run Modes: Estimation and Approximation

10

Diagram

y = f(x) = WTσ(VTx) + ε(x)

x0 = 1

x1

x2

xn

yn

y1

.

.

.

.

.

.

.

.

.

11

Neuromorphic (NM) Systems

“ … hardware specifically designed for

machine learning or inspired by biological

neural networks” (1)

“Brain Inspired” or “brain-like” in its

function (2)

1) Appavoo et al, “programmable Smart Machines: A Hybrid Neuromorphic

approach to General Purpose Computation”,

http://people.bu.edu/schuye/files/appavoo-neuroarch-2014.pdf

2) Mohda, D. S., “Introducing A Brain Inspired Computer”, refers to article

published in Science 8 August 2014,

http://www.research.ibm.com/articles/brain-chip.shtml

12

Hybrid Neuromorphic Systems Can be very complex combinations of digital and

analog devices (hardware and software) that

include components which simulate or emulate

neurological function in order to learn and perform

the required tasks.

A broader definition for neuromorphic computing,

covering those systems which include both von

Neumann (VN) and non-VN architectures.

Calimera, Andrea et al, “The Human Brain Project and neuromorphic

computing”,http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3812737/pd

f/191-196.pdf

13

Research

And Technology Advances

14

• Henry Markram EPFL

Lausanne, Switzerland An HBP Pilot (downselect 2012)

Patch clamp study for model data

• Neuroscience Modeling Uses IBM Blue Gene Computer

Michael Hines’s NEURON sw

Blue Brain 2005-2012

15

SyNAPSE is backronym – Neuromorphic

Adaptive Plastic Scalable Electronics

DARPA Program to develop

neuromorphic machine technology that

scales to biological levels

Substantial Funding went to HRL & IBM

SyNAPSE

16

Hewlett-Packard Research Lab

(HRL) Dec 2011 CMOS chip

- Simultaneous memory storage and

logic processing

- Multi-bit fully-addressable memory

with up to 30 Gbits/sq cm

MEMRISTOR Arrays

17

Massively Parallel Cortical

Simulation on Blue Gene/P (Dawn)

Supercomputer at LLNL 2009

- 1.6 Billion Neurons, 8.87 Trillion

Synapses

- Spiking Neurons (STDP) and axon

delay (0.1 ms simulated time step)

IBM “Cat Brain” Sim C2

18

Lab Prototype Digital Neurosynaptic

Core Processor and Integration

Board Announced Aug 2011

- 256 leaky integrate and fire (LIF)

neurons on 45 nm SOI CMOS hw

- deterministic hw with ~1 kHz chip

clock (emulates ~1ms bio time-step)

IBM New uProcessor Core

19

Neuromorphic Architecture Using

Simple LIF Spiking Neurons and

“Binary” Synapses - May 2012

And arranged in layers like brain, BUT

- Synaptic homeostatic renormalization

- Burst spiking time dependent plasticity

(Burst STDP)

New Implementation and Models

25

System Capabilities

• Robots

• - Baxter

• - Brett

• Big-Data-Mining Engines

• Self-Completing Software

• - PLINY, Brain####

• Autonomous Vehicles

• - Google Cars

• - Quadcopters

• Natural Language Processing

• Feature Recognition and Tracking

26

Model-Based V&V “was” Good Enough

Systems were simpler in their complexity

and autonomy

They could be white and black box tested at

points of homeostasis

And …

“A Meta Model could be constructed

that defined acceptable system states and

behaviors”

- Blackburn

27

Then What Happened?

Very Complex Learning Systems

Dramatic Emergent Behavior

28

Autonomous Robot Learning

29

A “Disruptive” System

Technology Announcement

And COMPASS

TN Simulator

June 2014

30

IBM TrueNorth

Video

And/or

Deep Dive

Summer 2015

31

• They Learn From a Teacher

• They Learn From Each Other

• They Learn From Their Environment

• Deep Learning and Reservoir Computing

Emergent Behaviors Must Be Controlled

• New Methods of V&V are Needed to

Achieve Desired Assurance Levels

The Machines Learn

32

Problem Re-Statement

The (Stochastic) Complexity of the Systems

We Are Now (and Will Continue to Be)

Constructing and Deploying (e.g. Reservoir

Computing and Liquid State Machines) Is

Rapidly Exceeding Our Ability To

Adequately and Feasibly Verify, Validate,

and Control Them

33

A neural network can be trained as a test

“oracle” or model of the original system to

determine whether a given test case exposes

a fault or not. - Dale Boren

A firm foundation for the use of neural

networks in feedback control systems has

been developed over the years by many

researchers. - F. L. Lewis

There Are Concepts

34

1. Blackburn, M., Denno, P., Virtual Design and Verification

of Cyber-Physical Systems: Industrial Process Plant

Design, Procedia Computer Science, Volume 28, 2014,

Pages 883-890,

http://www.sciencedirect.com/science/article/pii/S187705

0914000696

There Are Studies

35

There Are Experiments

36

There Are Solicitations

37

Neurogrid

- Stanford

SpiNNaker (Spiking NN Arch.)

- University of Manchester

TrueNorth (Compass, SyNAPSE)

- IBM

There Are Platforms

38

Trainable Testbed Concept

39

Trainable Testbed Benefits

40

IT IS TIME

For the Trainable Testbed