16 apr 27 tucson iit talk

Theoretical characterization and empirical testing of

integrated information theory (IIT) of consciousness

16 Apr 27 @ TucsonNao Tsuchiya 土谷尚嗣

School of Psychological Sciences &Monash Institute of Cognitive & Clinical Neuroscience (MICCN)

Monash University, Australia

"What is it like to be a bat?"

- a pathway to the answer from IIT

16 Apr 27 @ TucsonNao Tsuchiya 土谷尚嗣

School of Psychological Sciences &Monash Institute of Cognitive & Clinical Neuroscience (MICCN)

Monash University, Australia

• Postdoc job (JST-CREST, PI:Kanai, Co-PI: Tsuchiya, 2015-2021)

• Empirical testing of IIT

• If interested, contact me or Ryota via either email or in face.

Masafumi Oizumi

Ralph Adolphs Chris Kovach

Hiroto Kawasaki Matt Howard

Andrew Haun

Talk overview

• “What it is like to be a bat?”

• Any possible approach?

• Overview of IIT

• Empirical measures


• How to approach the bat problem?

What it is like to be a bat?

• Thomas Nagel 1974

• Clarify the difficulty of mind-body problem

• Bat - echolocation system

Several roads to consciousness?

• global workspace

• predictive coding

• quantum brain

• higher order theory

• NCC

• gamma oscillation, coherence, p3b, etc

• Why does vision feels like vision?

• Why any visual experience is more similar to other visual experience than auditory experience?

• Is bat’s echolocation similar to vision, audition, non-conscious or anything else?

bat’s experience

• Impossible stupid question?

• What about the age of universe?

• Direct evidence on evolution?

• Theories that explain all available evidence and make correct predictions can be trusted when making extraordinary predictions

Integrated information theory of consciousness

• Starts from phenomenology, identifies five axioms (1. existence, 2. composition, 3. information, 4. integration, 5. exclusion)

• Tries to seek for potential physical mechanisms that can support the axioms

Tononi 2004, 2008, Oizumi et al 2014 PLoS Comp Bio

Properties of conscious phenomenology

• Intrinsically exists

• Dreams, imagery, hallucinations depends on only neural activity

• Bistable percepts with identical stimuli

Nir & Tononi 2010 TICSLeopold & Logothetis 1999 TICS

Properties of conscious phenomenology

• Highly informative

• Each experience is highly unique and excludes all other possible experiences

Integrated and compositional aspect in conscious phenomenology

• Experience is always composed of various aspects

• Yet, it is always integrated as a whole

• How is integrated information composed?

IIT explains… • why a thalamo-cortical system generates

consciousness while cerebellum, retina (afferent), and motor systems (efferent) do not

IIT predicts …

• waking brains would maximally integrate information (indirectly supported by TMS-EEG experiments)

Massimini et al 2005 Science

Casali et al 2013 Science Trans Medicine

Common complains about IIT• IIT is all about levels of consciousness

• Difficult to understand -> misunderstanding

• Uncomputable and untestable

• No accessible paper (-> writing an essay for Philosophy Compass)

• My talk aims to give:

• 2 intuitive explanations

• 1 empirical tests

Tsuchiya, Haun, Cohen, Oizumi (in press) Return of Consciousness

H=log2(4)=2 I=H-H*=2 Φ=I-I*=2 ΦAB, ΦAC, ΦBC, ΦABC

Integrated information theory in a nut shell

Copy

Copy

Tsuchiya, Haun, Cohen, Oizumi (in press) Return of Consciousness

H=log2(4)=2 I=H-H*=2 Φ=I-I*=2 ΦAB, ΦAC, ΦBC, ΦABC

Note1: H, I, I*,Φ all depend on connectivity & state

Note2: perturbational or observational approaches

Integrated information theory in a nut shell

Continuous variables with Gaussian assumption

Oizumi et al 2016 PLoS Comp

https://figshare.com/articles/phi_toolbox_zip/3203326

1. Go to figshare.com2. Search for “phi_toolbox.zip”, “practical_phi.zip”or “oizumi”

http://figshare.com

Entropy H of X (assuming Gaussian distribution of X):

−500 0 500 1000100

0

100

0

100

0

100

0

100

-500 0 500 1000time from stimulus onset (msec)

bin 200ms

XD

XC

XB

XA


Continuous variables with Gaussian assumptionhttps://figshare.com/articles/phi_toolbox_zip/3203326

Measures variability of responses

Entropy H of X (assuming Gaussian distribution of X):

−500 0 500 1000100

0

100

0

100

0

100

0

100

-500 0 500 1000time from stimulus onset (msec)

bin 200ms

XD

XC

XB

XA


Φ* = I − I *

Continuous variables with Gaussian assumptionhttps://figshare.com/articles/phi_toolbox_zip/3203326

Integrated information=

loss of predictability of past states based on current states,

when system is cut

(Oizumi et al 2016 PLoS Comp)

Intuitive explanation 1

(Oizumi et al 2016 PLoS Comp)

Integrated information=

a distance between the actual and disconnected model

(Oizumi, Tsuchiya, Amari 2015 Arxiv)

Intuitive explanation 2

x1�

x2�

y1�

y2�

Difference�x1�

x2�

y1�

y2�

!me� !me�

Constraints�

Integratedinforma!on�

Generaliza!onoftheoriginalmeasureproposedbyTononiinhisIntegratedInforma!onTheory(Tononi,2008).�

Distance

Full model Disconnected model


Stochas(cinterac(on�

Mutualinforma(on�

Integratedinforma(on�

Transferentropy�

x1�

x2�

y1�

y2�

x1�

x2�

y1�

y2�

x1�

x2�

y1�

y2�

x1�

x2�

y1�

y2�

Disconnectedmodels:q(X,Y)�

ConstraintsintheGaussiancase�

Coordinates�

Stochas(cinterac(on�

Mutualinforma(on�

Integratedinforma(on�

Transferentropy�

x1�

x2�

y1�

y2�

x1�

x2�

y1�

y2�

x1�

x2�

y1�

y2�

x1�

x2�

y1�

y2�

Disconnectedmodels:q(X,Y)�

ConstraintsintheGaussiancase�

Coordinates�

A unified framework for causal interactions based on information geometry


• Empirical tests of IIT on contents of consciousness

• IIT’s prediction

• Patterns of integrated information should be isomorphic to contents of consciousness

Empirical testing of isomorphism

• Test the correspondence between the two constructs (i.g., phenomenology vs. mathematical constructs derived by a theory)

• Repeat the tests with a huge number of situations, including conscious/non-conscious perception, ambiguous perception, etc — marriage with NCC & no-report paradigm

Tsuchiya, Taguchi, Saigo 2016 Neurosci ResearchTsuchiya, Frassle, Wilke, Lamme 2015 TICS

Mutual information between present and past (tau=6-10 ms)

−400 −200 0 200 400 600 8005

5.2

5.4

5.6

H/c

han

−400 −200 0 200 400 600 8004.2

4.4

4.6

4.8

Hco

nd/c

han

−400 −200 0 200 400 600 8000.6

0.8

1

MI/c

han

−400 −200 0 200 400 600 800

0.6

0.8

1

I*/ch

an

−400 −200 0 200 400 600 8000

0.2

0.4

phi*

time from I onset

H ≈ log|Σ|

I = H - HCOND

φ* = I - I*

HCOND

I*

−400 −200 0 200 400 600 8005

5.2

5.4

5.6

H/c

han

−400 −200 0 200 400 600 8004.2

4.4

4.6

4.8

Hco

nd/c

han

−400 −200 0 200 400 600 8000.6

0.8

1

MI/c

han

−400 −200 0 200 400 600 800

0.6

0.8

1

I*/ch

an

−400 −200 0 200 400 600 8000

0.2

0.4ph

i*

time from I onset

H ≈ log|Σ|

I = H - HCOND

φ* = I - I*

HCOND

I*

Gaussian assumption

Randomness of the system - entropy: H(Xt)

Haun et al 2016 bioRxiv

−400 −200 0 200 400 600 8005

5.2

5.4

5.6

H/c

han

−400 −200 0 200 400 600 8004.2

4.4

4.6

4.8

Hco

nd/c

han

−400 −200 0 200 400 600 8000.6

0.8

1

MI/c

han

−400 −200 0 200 400 600 800

0.6

0.8

1

I*/ch

an

−400 −200 0 200 400 600 8000

0.2

0.4ph

i*

time from I onset

H ≈ log|Σ|

I = H - HCOND

φ* = I - I*

HCOND

I*

−400 −200 0 200 400 600 8005

5.2

5.4

5.6

H/c

han

−400 −200 0 200 400 600 8004.2

4.4

4.6

4.8

Hco

nd/c

han

−400 −200 0 200 400 600 8000.6

0.8

1

MI/c

han

−400 −200 0 200 400 600 800

0.6

0.8

1

I*/ch

an

−400 −200 0 200 400 600 8000

0.2

0.4

phi*

time from I onset

H ≈ log|Σ|

I = H - HCOND

φ* = I - I*

HCOND

I*

−400 −200 0 200 400 600 8005

5.2

5.4

5.6

H/c

han

−400 −200 0 200 400 600 8004.2

4.4

4.6

4.8

Hco

nd/c

han

−400 −200 0 200 400 600 8000.6

0.8

1

MI/c

han

−400 −200 0 200 400 600 800

0.6

0.8

1

I*/ch

an

−400 −200 0 200 400 600 8000

0.2

0.4ph

i*

time from I onset

H ≈ log|Σ|

I = H - HCOND

φ* = I - I*

HCOND

I*

Mutual information between present and past (tau=3 ms)


Unmasked conditions - 3 of 5 categories shown


Backward masking - 17 ms face in one of 4 quadrants


Visibility=4 Visibility=3 Visibility=2 Visibility=1

Highcontrast

face

Middlecontrast

face

Lowcontrast

face

Exemplar isomorphism across different tasks


x1�

x2�

y1�

y2�

x1�

x2�

y1�

y2�

x1�

x2�

y1�

y2�

Comparing isomorphisms across candidates

Equal-time interactions(synchrony, correlations, entropy)

All lagged interactions(mutual information, predictive information)

Isolated, total causal interactions (integrated information)

Oizumi et al 2015 arXiv

0.1

1

Φ

CFSBM UNM

Faces(CFS,BM,UNM) Masked Faces & Mondrians (CFS,BM,UNM)Houses (UNM) Tools (UNM)

0.1

1

I

0.1

1

H


Quantifying isomorphism across different tasks and subjects as classification accuracy


Summary of this talk• Overview of IIT

• Integrated information as:

• loss of decoding accuracy (Φ*)

• distance between full and disconnected models (ΦG)


• Phi patterns: hierarchy of causal interactions better correlates with conscious perception than entropy or mutual information.

A path to bats’ experience…

• Can be compared between sensory modalities

• Across individuals

• Across species

• Common topological properties of phi* pattern for vision vs audition?

• What is it like to be a bat?

• Dissolution of the Hard Problem?

Acknowledgment

• Thanks for your attention & consciousness!

• Support

• ARC Future Fellowship

• ARC Discovery Project

• JST PRESTO grant