Consciousness and Intelligence

Dr. Christof Koch

Allen Institute for Brain Science

Seattle, USA Cartesian Certainty

As a radical skeptic, the only certainty Rene Descartes had

Je pense, donc je suis

translated later on as

Cogito, ergo sum.

or, in modern language,

I am conscious, therefore I am

Rene Descartes (1637) Intelligence and Consciousness

• Intelligence - the ability to understand new ideas, to adapt to new environments, to learn from experience, to think abstractly, to plan and to reason • It can be decomposed into crystalline and fluid intelligence and can be measured (IQ, g-factor) • Consciousness - the ability to experience something, to see, hear, feel angry, or explicitly recall an event • Many animals besides humans experience the sights and sounds of the world Lilac Chaser What do we know about Consciousness?

• Consciousness is associated with some complex, adaptive, biological networks (not immune system nor enteric nervous system) • Consciousness does not require behavior • Consciousness can be dissociated from emotion, selective attention, long-term memory and language • Self-consciousness is one of many aspects of consciousness, highly developed in adult neuro-typical humans, less so in infants, certain patients and non-human animals Many Zombie Behaviors

Many - if not most - behaviors occur in the absence of conscious sensations, or consciousness occurs after the fact:

• Over-trained routines - shaving, dressing, tennis, video games, keyboard typing, driving, rock-climbing, dancing • Reaching and grabbing, posture adjustments • Generating and understanding speech • Eye-movement control • High-level decision making (e.g. choice blindness, dissociations) Neuronal Correlates of Consciousness (NCC)

Search for the minimal neuronal mechanisms jointly sufficient for any one conscious perception, the NCC. For every conscious percept, there will be a NCC. Inducing the NCC will induce the perception; inactivating the NCC will eliminate it. Crick & Koch (Nature 1995) Cerebellum

Yu et al (Brain 2014)

• 69 out of 86 billion neurons are in the cerebellum • Main deficit following cerebellar lesions are ataxia, slurred speech and unsteady gait • The cerebellum is not a significant part of the NCC NCC in Parietal-Tempero-Occipital Cortex

CONTENT STATE TMS fMRI ANATOMY

Frässle et al 2014 Bagattini et al 2015 Brickner 1952 MRI EEG MEG

Koivisto and Revonsuo 2010 Andersen et al 2015 Matarò et al 2001 fMRI EEG ECoG seizure post seizure Siclari et al 2014 Noy et al 2015 Bai et al 2010

conscious motor unconscious intention movements EBS EBS EEG

Rangarajan et al 2014 Desmurget et al 2009 Siclari et al 2014

Koch et al (Nature Rev. Neurosci. 2016) The Integrated Information Theory of Consciousness

Giulio Tononi

Tononi’s Phi: A Voyage from the Brain to the Soul (2013) Tononi & Koch (Phil. Trans. Roy. Soc. Long 2015) cause (past) effect (future) Intrinsic existence

A E OR

AND XOR B C A D Composition Higher order mechanism [ABC] On Off B C Indeterminate blue book [AB] [AC] [BC] book A A A effect (future) blue cause (past) left B[A] C B[B] C [C]B C A A A Elementary mechanisms B C B C B C

Information Cause-effect structure Cause Effect Axioms - repertoires repertoires Postulates - Properties that Essential mechanisms in a property of AA OR BB CC state must have to every experience AND XOR Integration account for experience Partitioned cause-effect structure

AA MIP OR BB CC AND XOR

Exclusion Φmax Conceptual structure Cause Effect repertoires repertoires time ϕmax0.5 of concepts

0.33 E A OR 0.25

0.25 AND B C XOR

0.17 Φ=1.0 D A 0.17 Φ=0.22 Complex Φmax = 1.92 An experience is identical to a conceptual structure supported by a complex of elements in a state

EXPERIENCE CONCEPTUAL STRUCTURE (QUALE)

Cause repertoires Effect repertoires Cause-effect space ϕmax of concepts 000 0.5 001 110 100 101 0.33 BC C 110 101 B 0.25 AB 111 A 101 ABC 0.25 001 future past states 011 001 0.17 states 100 011 010 010 0.17 Probability of state Quale (ΦMax = 1.92) Concept (ϕmax = 0.17)

A E Quantity: Irreducibility of the OR conceptual structure (Φmax) AND XOR Quality: “Shape” of the quale in B C qualia space (“constellation”) D Predictions from IIT - Breakdown of Cortical Integration During Unconsciousness

Massimini et al (Science 2005; Brain, 2012) Casali et al (Science Transl Med. 2013) Feed-Forward Systems

• Purely feed-forward systems, such as deep convolutional neural networks, have Φmax = 0 • They have no intrinsic cause-effect power • While they can be smart, it does not feel like anything to be such a network

GoogLeNet, a 22 layers deep NN Two Functional Equivalent Systems - A Conscious one and a Zombie Difference between the Real and the Simulated

• Just like a computer simulating a rainstorm will not get wet, a computer simulating conscious behavior will not be conscious

• For a computer to have human-level consciousness, it will have to have the causal powers of the human brain (e.g. using neuromorphic architecture) Consciousness versus Intelligence Consciousness

0, 0 Intelligence So What?

• Does it matter whether or not it feels-like-something to be a synth? • It makes all the difference in the world to the synth itself. • Having the ability to experience pain and pleasure, i.e. to be a subject, implies a set of minimal rights and ethical obligations on the side of its creator and/or owner. • If a synth is not conscious, then it is an object like any other object and can be treated as such. • Answering the question of whether computer software and robots can be conscious will be a critical challenge in the years to come.