universal laws and architectures : theory and lessons from brains, bugs, nets, grids,

Universal laws and architectures:Theory and lessons from

brains, bugs, nets, grids, docs, planes, fire, fashion,

art, turbulence, music, buildings, cities, earthquakes, bodies, running, throwing,

Synesthesia, spacecraft, statistical mechanics

Slow

Flexible

Fast

Inflexible

Impossible (law)

Architecture

General Special

1

2

4

8

.1 1.05 .5.2

0l l

0 1l l

Length (meters)

Fragility

ln z pp

p

z

.1s

k

z pz p

10-1 100 101100

101

too fragile

complex

No tradeoff

expensive

fragile

costly

fragile

cheap

robust

Survive

Multiply

thinsmall

thickbig

Laws

Slow

Flexible

vision

eye vision

Act slowdelay

Fast

Inflexible

VOR

fast

Slow

Flexible

Fast

Inflexible

Impossible (law)

Universal laws

Chiang, Low, Calderbank, and Doyle

Starting point

Slow

Flexible

Fast

Inflexible

Impossible (law)

Architecture

General Special

Universal Turing Machine

Fast

Slow

Flexible Inflexible

Apps

OS

HW

AppsOS

HardwareDigital

LumpedDistributed

Tech implications/extensions

DiverseDeconstrained

(Hardware)

Deconstrained(Applications)

Diverse

Layered architectures

ConstrainedBut hidden

Apps

OS

HW

Core Protocols

Fast

Slow

Flexible Inflexible

General Special

Apps

OS

HW

FastCostly

SlowCheap

Flexible Inflexible

General Special

Layered Architecture

AppsOSHW

DigitalLumpedDistrib.

OSHW



LumpedDistrib. Distrib.

Any layer needs all lower layers.

DigitalLumped Lumped

.

AppsOS

HardwareDigital

Lumped

Act

Decide

Sense

Act

Sense

AppsOS

HardwareDigital

Lumped

Decide

Slow

InflexibleFragile

Expensive

vision

Act

delay

Efficiency/instability/layers/feedback

• Money/finance/lobbyists/etc• Society/agriculture/weapons/etc• Bipedalism• Maternal care• Warm blood• Flight• Mitochondria• Translation (ribosomes)• Glycolysis (2011 Science)

• All create new efficiencies but also instabilities• Requires new active/layered/complex/active control

Slow

Inflexible

Fragile

Expensive

vision

Act

delay

+ Neuroscience

Completing the story

Balancing an inverted pendulum

Mechanics+Gravity +Light +

2 20

1 ln

ln

pT j dp

z ppz p

Speed and flexibility tradeoffs Concrete case studies?

Theorems ?

Fast

Slow

Flexible Inflexible

fragile

robust

Slow

Flexible

vision

eye vision

Actslow

delay

Fast

Inflexible

VOR

fast

sense

move Spine

delay=death

sense

move Spine

Reflex

Reflect

sense

move Spine

Reflect

Reflex

Layered?

Physiology

Organs

Neu

rons

Neu

rons

Neu

rons

Cor

tex

Cel

ls

Cor

tex

Cor

tex

Layered architectures?

Cells

sens

e

mov

eSp

ine

Refle

ct

Refle

x

Layered

Prediction

GoalsActions

errors

Actions

Physiology

Organs

Meta-layers

Prediction

GoalsActions

errors

ActionsCor

tex

VisualCortex

VisualThalamus

10x

?

There are 10x feedback neurons

Why?


VisualCortex

VisualThalamus

10x

?


Why?

Whynot?

No feedback

Unfortunately, not intelligent design

Ouch.

Why?

left recurrent laryngeal nerve

Why? Building humans from fish parts.

Fish parts

It could be worse.

VisualCortex

VisualThalamus

?

Physiology

Organs

Prediction

GoalsActions

errors

Actions

Prediction

Goals

Consciousperception

10x

Why?

What are the consequences?

Which blue line is longer?

Which blue line is longer?

With social

pressure, this one.

Standard social psychology experiment.

3D+time

Simulation

Seeing is dreaming

Consciousperception

Consciousperception Zzzzzz…..

Same size?

Same size

Same size

Even when you “know” they are the same, they appear different

Toggle between this slide and the ones before and after

Same size?

Vision: evolved for complex simulation and control, not

2d static pictures

Even when you “know” they are the same, they appear different

3D+time

Simulation+ complex

models(“priors”)

Seeing is dreaming

Consciousperception


3D+time

Simulation+ complex


Seeing is dreaming

Consciousperception

Con

scio

uspe

rcep

tion

Prediction

errors

3D+time

Simulation+ complex


Seeing is dreaming

Consciousperception

Con

scio

uspe

rcep

tion

Prediction

errors

Seeing is believing

3D+time

Simulation+ complex


Seeing is dreaming

Consciousperception

Con

scio

uspe

rcep

tion

Prediction

errors

Seeing is believing

Our most integrated perception

of the world

3D+time

Simulation+ complex


Seeing is dreaming

Consciousperception


3D+time

Simulation+ complex


Seeing is dreaming

Consciousperception

Con

scio

uspe

rcep

tion

Prediction&Control

errors

Unconscious but high level

Effect: rival-schemata ambiguity

SensoryMotor

Prefrontal

Striatum

SlowFlexible

Learning

Ashby & Crossley

Slow

Reflex(Fastest,

LeastFlexible)

Extreme heterogeneity

SensoryMotor

Prefrontal

Striatum

Fast

FastInflexible

SlowFlexible

Learning

Reflex(Fastest,

LeastFlexible)

Ashby & Crossley

Learning can be very slow.

Sense

Universal tradeoffs?

Fast

Slow

Flexible Inflexible

Motor

Prefrontal

Fast

Learn

Reflex

Evolve


Evolution is even slower.

?

Action (fast)?Perception (slow)?

Effect: rival-schemata ambiguity

Slow

Flexible

vision

eye vision

Actslow

delay

Fast

Inflexible

VOR

fast

3D + motion

colorvision

Slowest

See Marge Livingstone

This is pretty good.Stare at the intersection

Slow

Flexible

vision

eye vision

Actslow

delay

Fast

Inflexible

VOR

fast

3D + motion

colorvision

Slowest

Slow

Flexible

vision

Fast

Inflexible

VOR

colorvision

Seeing is dreaming

-3d+motionB&W

slowVOR

fast

Fast

Slow

Flexible Inflexible

slower

Objects

Mixed

Chess experts• can reconstruct entire chessboard with < ~ 5s inspection• can recognize 1e5 distinct patterns• can play multiple games blindfolded and simultaneous• are no better on random boards

(Simon and Gilmartin, de Groot)

www.psywww.com/intropsych/ch07_cognition/expertise_and_domain_specific_knowledge.html

-3d+motionB&W

slowVOR

fast

Fast

Slow

Flexible Inflexible

Faces

slower

Objects

Mixed

Not sure how to draw this…


Fast

Slow

Flexible Inflexible

Reflex

Evolve



When needed, even wasps can do it.

• Polistes fuscatus can differentiate among normal wasp face images more rapidly and accurately than nonface images or manipulated faces. • Polistes metricus is a close relative lacking facial recognition and specialized face learning. • Similar specializations for face learning are found in primates and other mammals, although P. fuscatus represents an independent evolution of specialization. • Convergence toward face specialization in distant taxa as well as divergence among closely related taxa with different recognition behavior suggests that specialized cognition is surprisingly labile and may be adaptively shaped by species-specific selective pressures such as face recognition.

Fig. 1 Images used for training wasps.

M J Sheehan, E A Tibbetts Science 2011;334:1272-1275Published by AAAS

vision

eye vision

Actslow

delay

VOR

Slow

Flexible

Fast

Inflexible

Illusion

Highly evolved (hidden)

architecture

Seeing is dreaming

3D+time

Simulation+ complex


Seeing is dreaming

Consciousperception

Con

scio

uspe

rcep

tion

errors

Prediction&Control

-3d+motionB&W

slowVOR

fast

Fast

Slow

Flexible Inflexible

Faces

slower

Objects

Mixed

Not sure how to draw this…

SensoryMotor

Prefrontal

Striatum

SlowFlexible

Learning

Ashby & Crossley

Slow

Reflex(Fastest,

LeastFlexible)

Extreme heterogeneity

SensoryMotor

Prefrontal

Striatum

Fast

FastInflexible

SlowFlexible

Learning

Reflex(Fastest,

LeastFlexible)

Ashby & Crossley


Sense


Fast

Slow

Flexible Inflexible

Motor

Prefrontal

Fast

Learn

Reflex

Evolve



Sense

Fast

Slow

Flexible Inflexible

Motor

Prefrontal

Fast

Learn

Reflex

Evolve

Apps

OS

HW

SenseMotor

Prefrontal

Fast

Learn

Reflex

Evolve

vision

VORInternal delays

.01

1

delaysec/m

Aα

C

.1

Aβ

AγAδ

area

.1 1 10diam(m)

.01 1 100

Axon size and

speed

Fast

Small

Slow

Large

delaysec/m

.01

1

.1

area

.1 1 10diam(m)

.01 1 100

Retinal ganglion axons?

Other myelinated

Why such extreme diversity in delay and size?

VOR

Why no diversity in VOR?

control feedback

RNA

mRNA

RNApTranscription

Other Control

Gene

AminoAcids

Proteins

Ribosomes

Other Control

Translation

Metabolism Products

Signal transductionATP

DNANew gene

SensoryMotor

Prefrontal

Striatum

Reflex

Learning

Software

Hardware

DigitalAnalog

Universal layered

architectures

FastCostly

SlowCheap

Flexible Inflexible

General Special


AppsOSHW


OSHW





Slow

Flexible

Fast

Inflexible

Impossible (law)

Architecture

Architecture (constraints that

deconstrain)

General Special

Universal laws and architectures (Turing)

ideal

0. HGT1. DNA repair2. Mutation 3. DNA replication4. Transcription5. Translation6. Metabolism7. Signal transduction8. …

AppsOS

HardwareDigital

LumpedDistributed


control feedback

RNA

mRNA

RNApTranscription

Other Control

Gene

AminoAcids

Proteins

Ribosomes

Other Control

Translation

Metabolism Products


DNANew gene

Horizontal Gene

Transfer

Sequence ~100 E Coli (not chosen randomly)• ~ 4K genes per cell• ~20K different genes in total (pangenome)• ~ 1K universally shared genes • ~ 300 essential (minimal) genes

control feedback

RNA

mRNA

RNApTranscription

Other Control

Gene

AminoAcids

Proteins

Ribosomes

Other Control

Translation

Metabolism Products


DNANew gene

FastCostly

SlowCheap

Flexible Inflexible

General Special

HGT DNA repair

Mutation DNA replication

Transcription Translation

MetabolismSignal…


Slow

Fast

Flexible InflexibleGeneral Special

AppsOSHWDig.Lump.Distrib.

OSHWDig.Lump.Distrib.

DigitalLump.Distrib.


HGT DNA repair Mutation DNA replication Transcription

Translation Metabolism

Signal

SenseMotor

Prefrontal

Fast

Learn

Reflex

Evolve

vision

VOR

Slow

Fast







Translation

Metabolism Signal

SenseMotor

Prefrontal

Fast

Learn

Reflex

EvolveVOR

Many systems/organisms only use lower layers


Metabolism

Signal

SenseMotor Fast

Reflex

EvolveVOR

Many systems/organisms only use lower layers

Analog

Red blood cellsDenucleated neurons

Most metazoans

200 microns

Denucleated neurons

Megaphragma mymaripenne

5 day lifespan7,400 neurons vshousefly 340,000 honeybee 850,000.

control feedback

RNA

mRNA

RNApTranscription

Other Control

Gene

AminoAcids

Proteins

Ribosomes

Other Control

Translation

Metabolism Products


DNANew gene

SensoryMotor

Prefrontal

Striatum

Reflex

LearningSoftware

Hardware

DigitalAnalog

Horizontal Gene

Transfer

Horizontal App

Transfer

Horizontal Meme

Transfer

Accelerating evolution

Requires shared “OS”

control feedback

RNA

mRNA

RNApTranscription

Other Control

Gene

AminoAcids

Proteins

Ribosomes

Other Control

Translation

Metabolism Products


DNANew gene

SensoryMotor

Prefrontal

Striatum

Reflex

LearningSoftware

Hardware

DigitalAnalog

Horizontal Gene

Transfer

Horizontal App

Transfer

Horizontal Meme

Transfer

What can go wrong?

Horizontal Bad Gene Transfer

Horizontal Bad App Transfer

Horizontal Bad Meme

Transfer

Parasites &

Hijacking

Fragility?

Exploiting layered

architecture

Virus

Virus

Meme

FastCostly

SlowCheap

Flexible Inflexible

General Special

Layered immunity

Adaptive

OS?

Innate

Vaccination ≈ horizontal transfer?

Slow

Fast







Translation Metabolism

Signal

SenseMotor

Prefrontal

Fast

Learn

Reflex

Evolve

vision

VORAdapt

Adapt

Slow

Fast


Complex modelsRobust controllers

Efficiency/instability/layers/feedback

Working backwards• Money/finance/lobbyists/etc• Society/agriculture/weapons/etc• Bipedalism• Maternal care• Warm blood• Flight• Mitochondria• Translation (ribosomes)• Glycolysis (see 2011 Science paper)

• All create new efficiencies but also instabilities• Requires new active/layered/complex/active control

Slow

Fast

Flexible Inflexible

Undecidable

Universal Turing

MachineArchitecture

(constraints that deconstrain)

General Special

Laws and architectures(Turing, 1936)

Turing 1912-1954

Turing’s 3 step research:0. Virtual (TM) machines1. hard limits, (un)decidability

using standard model (TM)2. Universal architecture

achieving hard limits (UTM)3. Practical implementation in

digital electronics (biology?)

Essentials:0. Model1. Universal laws2. Universal architecture3. Practical implementation

Software

Hardware

DigitalAnalog

Turing as “new”

starting point?

Fast

Slow

Flexible/General

Inflexible/Specific

Undecidable NP P

hard limits

Really slow

Computational complexity

Decidable

Fast

Slow

NP P

hard limits

Really slow

Computational complexity

Decidable

Space?

Flexible/General

Inflexible/Specific

NP P

Decidable

Computational complexityPSPACENPPNL

PSPACE ≠ NL

NLPSPACE

Space is powerful and/or cheap.

Conjectures, biology• Memory potential • Examples

– Insects– Scrub jays– Autistic Savants

Gallistel and King

• But why so rare and/or accidental?• Large memory, computation of limited value?• Selection favors fast robust action? • Brains are distributed (not studied by Gallistel)

Essentials:0. Model1. Universal laws2. Universal architecture3. Practical implementation

….0. Model

0. Model0. Model1. Universal laws2. Universal architecture3. Practical implementation

But there are always hidden assumptions…

0. Model1. Universal laws2. Universal architecture3. Practical implementation

Software

Hardware

DigitalAnalog

0. Analog components 1. 2. 3.

0. Digital components + algorithms

1. Undecidability2. Universal TM

0. Software1. Undecidability2. OS + Apps

Control, OR

CommsCompute

Physics

Shannon

Bode

Turing

Gödel

EinsteinHeisenberg

Carnot

Boltzmann

Theory?Deep, but fragmented, incoherent, incomplete

Nash

Von Neumann

Control, OR

Compute

Bode

Turing

Delay and risk are most important

• Worst-case (“risk”)• Time complexity (delay)

• Worst-case (“risk”)• Delay severely degrades

robust performance

Computation for control• Off-line design• On-line implementation• Learning and adaptation

Control, OR

CommunicateCompute

Physics

Shannon

Bode

Turing

Einstein

Heisenberg

Carnot

Boltzmann

Delay and risk are

most important

Delay and risk are

least important

Communicate

Physics

Shannon

Einstein

Heisenberg

Carnot

Boltzmann

Dominates “high

impactscience” literature

• Average case (risk neutral)• Random ensembles• Asymptotic (infinite delay)

• Space complexity

Control, OR

CommunicateCompute

Physics

Shannon

Bode

Turing

Delay and risk are

most important

Delay and risk are

least important

New progress!

Modern Turing tradeoffs: PC, smartphone, router, etc

AppsOSHW


OSHW




Fast

Slow

Flexible Inflexible

Accident or necessity?

General Special

AppsOSHW


OSHW




Fast

Slow

Flexible Inflexible

General Special

What does this cartoon mean?

Fast

Slow

Flexible Inflexible

General Special

There are tradeoffs

AppsOSHW


Fast

Slow Software

Hardware

DigitalAnalog

Apps

OS

Speed depends on implementation

AppsOSHW

DigitalAnalog

…

Software

Hardware

DigitalAnalog

Apps

OS

HWDigitalAnalog

…Lumped analogDistrib. analog

Every layer needs all layers below

AppsOSHW

DigitalAnalog

…

HWDigitalAnalog

…Lumped analogDistrib. analog

Networks permit heterogeneity in # and kind of layers

Device 1 Device 2

Device 3

We completely understand

this but don’t explain it well.

Software

Hardware

DigitalAnalog

Apps

OS

AppsOSHW


OSHW





Flexibility depends on implementation

AppsOSHW


OSHW




Fast

Slow

Flexible Inflexible

General Special

Simplify/compress the cartoon (there is no standard notation)

AppsOSHW


OSHW




Fast

Slow

Flexible Inflexible

General Special

Software

Hardware

DigitalAnalog

Simplify/compress the cartoon (there is no standard notation)

Software

Hardware

DigitalAnalog

Flexible/ Adaptable/Evolvable

Horizontal App

Transfer

Depends crucially on

layered architecture

AppsOS

HW

Digital

Lumped

Distrib.

OSHW

Digital

Lumped

Distrib.

DigitalLumped

Distrib.

LumpedDistrib. Distrib.Fast

Slow

Flexible Inflexible

SensoryMotor

Prefrontal

Striatum

Reflex

Learning

Catabolism

AA

RNA

transl.Proteins

xRNA transc.Precursors

Nucl.AA

DNARepl.Gene

ATP

ATPRibosome

RNAp

DNAp

Software

Hardware

DigitalAnalog

Flexible/ Adaptable/Evolvable

Horizontal Gene

Transfer

Horizontal App

Transfer

Horizontal Meme

Transfer

Depends crucially on

layered architecture

.01

1

delaysec/m

Aα

C

20 m

.2 m

2 s/m

.008 s/m

.1

Axon size and

speed

area

.1 1 10diam(m)

.01 1 100

.01

1

delaysec/m

Aα

C

.1

Aβ

AγAδ

area

.1 1 10diam(m)

.01 1 100

delaysec/m

.01

1

.1 myelinated

unmyelinated

area

.1 1 10diam(m)

.01 1 100

Axon size and speed

2marea

delaysec/m

.01

1

.1 1 10diam(m)

.110 m

.1 m

1 m

myelinated

.01 1 100

myelin

delaysec/m

.01

1

.1

area

.1 1 10diam(m)

.01 1 100

Retinal ganglion axons?

Other myelinated

Why such extreme diversity in delay and size?

VOR

Why no diversity in VOR?

What about noise?

Practically everythingdelay

low

high

errorlow high

delay

low

high

errorlow high

?2

delay

?2

noise

.01

1

delaysec/m

20 m

.2 m

2 s/m

.008 s/m

.1

area

.1 1 10diam(m)

.01 1 100

120 m/s

.5 m/s 2.03 m

2300 m

250 delay

410 area

speed costs

40 x bandwidth

Need to check x

.01

1

delaysec/m

.1

area

.1 1 10diam(m)

.01 1 100

250 delay

410 area

speed costs

40 x bandwidth

Need to check x

.01

1

delaysec/m

.1

area

.1 1 10diam(m)

.01 1 100

250 delay

410 area

speed costs

40 x bandwidth

Need to check x

Idea seems right, but details might not be…

.01 1 100area

2

/16 m sm

2

/.4 m sm

2 2

bw speedm m

.01

1

.1 1 10diam(m)

.1delayaxon

Max Mutual Info

Communicate

Physics

delay and

perfect parts

B&Wvision

e=d-u

slow

delay

brains

1e5-1e8 kT J/bit(Laughlin)

1 kT J/bit(Landauer)

delayaxon

.01

1

.1

» 1 kT J/bit

» 1e5-1e8 kT J/bit

Theory

Biology

Better theory?

Budd JML et al. (2010)Neocortical Axon Arbors Trade-off Material and Conduction Delay Conservation. PLoS Comput Biol 6(3): e1000711. doi:10.1371/journal.pcbi.1000711

.01

1

.1

.01 1 1002area d

1delay d

periphery, reflex

centrally, reflect

AppsOS

HW

Digital

Lumped

Distrib.

OSHW

Digital

Lumped

Distrib.

DigitalLumped

Distrib.

LumpedDistrib.

HGT DNA repair

Mutation DNA replication

TranscriptionMetabolism

Signal…

VOR

Slow

Fast

Flexible Inflexible

General Special

visionSystem

tradeoffs

component tradeoffs

network tradeoffs?

Yorie Nakahira

Fast

Cheap

Random

Axon geometry?delay

wire costmin

min

(cartoon version)

Fast

Cheap

Random

actual

Axon geometry?delay

wire costmin

min

Fast

Cheap

Random

actual

Axon geometry?

delay

wire costmin

min

synchro

ideally

delayperfect synchro

wiring

wiring

delaysynchro

sweet spot

delaysynchro

wiring

wiring

delaysynchro

sweet spot

Universal laws (constraints)

Well known from physics and chemistry:• Classical: gravity, energy, carbon,…• Modern: speed of light, “Heisenberg”, …

Not so much:• Robustness

– Critical to study of complex systems– Unknown outside narrow technical disciplines– Theorems, not mere metaphors

Control, OR

CommsCompute

Physics

Shannon

Bode

Turing

Gödel

EinsteinHeisenberg

Carnot

Boltzmann

Theory?Deep, but fragmented, incoherent, incomplete

Nash

Von Neumann

Control, OR

Compute

Bode

Turing

Delay and risk are most important

• Worst-case (“risk”)• Time complexity (delay)

• Worst-case (“risk”)• Delay severely degrades

robust performance

Computation for control• Off-line design• On-line implementation• Learning and adaptation

Control, OR

CommunicateCompute

Physics

Shannon

Bode

Turing

Einstein

Heisenberg

Carnot

Boltzmann

Delay and risk are

most important

Delay and risk are

least important

Communicate

Physics

Shannon

Einstein

Heisenberg

Carnot

Boltzmann

Dominates “high

impactscience” literature

• Average case (risk neutral)• Random ensembles• Asymptotic (infinite delay)

• Space complexity

Control, OR

CommunicateCompute

Physics

Shannon

Bode

Turing

Delay and risk are

most important

Delay and risk are

least important

New progress!

Control, OR

CommunicateCompute

Physics

Delay and risk are

most important

Delay and risk are

least important

New progress!

Martins (UMD)Rotkowitz (UMD)

Sarma (JHU)

Slow

Flexible

vision

eye vision

Actslow

delay

Fast

Inflexible

VOR

fast

B&W (luminence only):3D, motion, and action

3D + motion

universal laws and architectures : theory and lessons from brains, bugs, nets, grids,

Documents

blue line

new efficiencies

fish parts28

social pressure

oshwdigitallumpeddistri

length meters fragility

lower layers

d timesimulationseeing