Can Computer Science Engineer Conscious Machines?

October, 2014 David Gnabasik

Questions to Address & Ground Rules

1) Do we need a new definition of machine consciousness (i.e., sentience) and/or intelligence? What is their relationship?

2) What are the computer science assumptions, components, dependencies required for its possible realization?

What CS tools can we use to analyze this possibility? What’s testable?

3) What computer science framework supports the engineering of intelligent or conscious machines?

Ground Rules:

No more than 24 slides for a 30-minute presentation.

Explicitly state questions/hypotheses. Graphs are useful!

Post the presentation on your Ph.D. website.

Have fun! It doesn’t have to necessarily be about your own work.

Primary Research Resources For This Topic

• Consciousness and Artificial Intelligence: Theoretical foundations and current approaches – AAAI Symposium, Washington DC, 8-11 November 2007*

• Artificial Intelligence and Consciousness, Antonio Chella, Riccardo Manzotti

• IEEE Spectrum – The Rapture of the Geeks, June, 2008 • http://spectrum.ieee.org/static/singularity

• The Coming Technological Singularity, Vernor Vinge, 1993 • https://www-rohan.sdsu.edu/faculty/vinge/misc/singularity.html

What statement best demonstrates that the HAL 9000 computer is conscious or intelligent?

• “It can only be attributable to human error.”

• Dave: “Do you know what happened?” HAL: “I’m sorry, Dave. I don’t have enough information.”

• “Without your space helmet, Dave? You're going to find that rather difficult.”

• “That's a very nice rendering, Dave. I think you've improved a great deal. Can you hold it a bit closer? That's Dr. Hunter, isn't it?”

• “I'm sorry, Frank, I think you missed it. Queen to Bishop 3, Bishop takes Queen, Knight takes Bishop. Mate.”

• “I'm afraid. I'm afraid, Dave. Dave, my mind is going. I can feel it. I can feel it. My mind is going. There is no question about it. I can feel it.”

Artificial Consciousness (AC) - Motivations

Ricardo Sanz (Sanz, 2005) claims there are three reasons for pursuing artificial consciousness:

1) implement and design machines that resemble human beings (cognitive robotics);

2) understand the nature of consciousness (cognitive science);

3) implement and design more efficient control systems.

Owen Holland (Holland, 2003) distinguishes between:

• Weak Artificial Consciousness: design and construction of machine that simulates consciousness;

• Strong Artificial Consciousness: design and construction of conscious machines

Do we need a new definition of (i.e., self-awareness) consciousness or intelligence? • Chella & Manzotti: autonomy, resilience, direct experience,

learning, attention; to feel something. – Why not simply say “provides intelligent responses”?

• Koch & Tononi: Can Machines Be Conscious? • Much brain activity has nothing to do with consciousness. • Being conscious (subjective experience) does not require:

– emotions - dreaming, frontal lobe damage – either explicit or working memory – attention & self-reflection, but coupled to selective attention.

• neocortex-damaged patients have nearly intact perceptual abilities.

– language - animals, infants.

What are the CS assumptions, components, dependencies required for its possible realization?

• Artificial Intelligence (AI), genetic algorithms, neural networks, machine learning in general.

• Analyzing the brain as a circuit diagram:

– Ultrahigh-resolution brain scans computer simulation

• Define machine: an apparatus using or applying mechanical power and having several parts, each with a definite function and together performing a particular task. Now it seems to mean algorithm, a set of deterministic & well-defined rules.

Framework for Artificial Consciousness

• Embodiment

– address the issues of symbol grounding, anchoring, & intentionality

• Simulation and depiction

– develop models of mental imagery, attention, & working memory

• Environmentalism or externalism

– integration between the agent and its environment

• Extended control theory

– study of AC as a kind of extended control loop

Development of Necessary AI Capabilities

• Rodney Brooks (MIT's Humanoid Robotics Group) – “…the brain is a machine, whether this machine is a computer is

another question.“

• Possible development program: 1. Object-recognition capabilities of a 2-year old.

2. Language capabilities of a 4-year old.

3. Manual dexterity of a 6-year old.

4. Social understanding of an 8-year old.

• Most significant factor is continued HW development. – Computational power is the requirement for development of AI.

• Extrapolate the idea of exponential technological growth several versions of the singularity. – AI is the main tool for the paradigm shift required by the

singularity.

Gordon E. Moore's law: the number of transistors in a dense integrated circuit doubles approximately every two years. (1965)

Kurzweil believes that the exponential growth of Moore's law will continue beyond the use of integrated circuits into technologies that will lead to the technological singularity.

Dennard Scaling = Constant Electric Field Scaling • Dennard scaling theory deals with switching speeds and physical

characteristics of transistors heat dissipation, max clock speeds.

• MOSFETs continue to function as voltage-controlled switches while all key figures of merit such as layout density, operating speed, and energy efficiency improve – provided geometric dimensions, voltages, and doping concentrations are consistently scaled to maintain a constant electric field.

• Dennard scaling made Moore’s law work and it stopped working in 2004!

• Leakage, etc. Power becomes the limiting problem. Scaling variability!!

Application of Moore's Law

• In 2003, Intel predicted the end would come between 2013 and 2018 with 16 nanometer manufacturing and 5 nanometer gates, due to quantum tunneling.

• Joel Hruska, Intel’s former chief architect: Moore’s law will be dead within a decade, Aug. 30, 2013 [1]: – “It is a common (but mistaken) belief that Moore's law makes predictions

regarding all forms of technology, when it was originally intended to apply only to semiconductor circuits. Kurzweil has hypothesized that Moore's law will apply – at least by inference – to any problem that can be attacked by digital computers as is in its essence also a digital problem. But nothing else we’ve ever discovered has scaled like semiconductor design. A 2011 study in the journal Science showed that the peak of the rate of change of the world's capacity to compute information was in the year 1998, when the world's technological capacity to compute information on general-purpose computers grew at 88% per year.[2] Since then, technological change has clearly slowed.”

• Continued, exponential HW growth extrapolates an extrapolation (Moore's law) to another field.

Extending Moore’s Law - Continued Scaling • Molecular integrated circuits; nanotechnology; quantum computing. • Use spin state of electron spintronics, tunnel junctions, advanced

confinement of channel materials via nano-wire geometry. • Microarchitecture techniques exploiting the growth of available

transistor count. – Out-of-order execution and on-chip caching and prefetching reduce the memory

latency bottleneck at the expense of using more transistors and increasing processor complexity.

– Pollack's Rule: performance increases due to microarchitecture techniques are square root of the number of transistors or the area of a processor.

• IBM SyNAPSE Chip (Aug, 2014) [^] – A neurosynaptic computer chip with 1 million programmable neurons, 256 million

programmable synapses and 46 billion synaptic operations per second per watt, operating at a minuscule 70mW.

– On August 7 2014, IBM announced their second generation SyNAPTIC chip, which contains the most transistors in a Neurosynaptic chip to date at 5.4 billion.[3]

• Stanford’s Neurocore chip: Neurogrid [4,5] (April, 2014) – Stanford bioengineers developed a circuit board modeled on the human brain: – 16 custom designed Neurocore chips simulate 1 million neurons and billions of

synaptic connections. – Neurogrid is claimed to be 9,000 times faster and more energy efficient than a PC.

How would we as Computer Scientists engineer intelligent or conscious machines?

• Options from Vernor Vinge’s 1993 essay:

– Create super-human AI in computers (hyper-intelligence)

– Human intelligence amplification through novel UI.

– Biomedical: enhance neurological operation of our brains.

– The Internet becomes self-aware.

– Assumption: “…algorithms are of central importance to the existence of minds.”

• What is the best way to build a conscious machine?

– Copy the mammalian brain (harder)

– Evolve a machine (easier).

• Brooks’ development program.

Revisit: Do we need a new definition of consciousness or intelligence?

• Koch & Tononi: Can Machines Be Conscious? • Essential properties of consciousness:

– related to the amount of integrated information generated. – information is defined as the reduction of uncertainty that occurs

when one among many possible outcomes is chosen. – A photodiode has just 2 states (1 bit each), but we have an

uncountable number of things that are not seen.

• To be conscious you need to be a single integrated entity with a large repertoire of available states. – Conscious experience implies differentiating among many states. – Photodiode states are completely independent, but our states

are very interdependent. – The repertoire of available states cannot be subdivided.

Yes, we need a new definition of consciousness or intelligence.

• Theoretical framework: Integrated Information Theory of Consciousness [*]

• IITC suggests a test of assessing machine consciousness à la Turing Test: – Describe a scene that efficiently differentiates the

scene's key features from the immense range of other possible scenes.

• IITC does not explain why consciousness exists, but it should uphold “common sense” on common test cases. Given the experimental inaccessibility of consciousness, this is the only test available to us.

Integrated Information Theory of Consciousness

• (1) Φ quantitatively measures the amount of “integrated information” in a physical system (i.e. information that can’t be localized in the system’s individual parts) by minimizing, over all subdivisions of a physical system into two parts A and B, some measure of the mutual information between A’s outputs & B’s inputs & vice versa.

• (2) A physical system is “conscious” if and only if it has a large value of Φ—and indeed, that a system is more conscious the larger its Φ value.

• (3) Significant challenges: – We don’t know the detailed interconnection network of the

human brain. – Not even clear what we should define that network to be. – Computing Φ(f,x) probably computationally intractable.

• Consider a discrete system in a state x=(x1,…,xn)∈Sn, where S is a finite alphabet (the simplest case is S={0,1}). Imagine that the system evolves via an updating function f:Sn→Sn. Can the xi‘s be partitioned into two sets A and B, of roughly comparable size, such that the updates to the variables in A don’t depend very much on the variables in B and vice versa. If such a partition exists, then we say that the computation of f does not involve global integration of information, which on Tononi’s theory is a defining aspect of consciousness.

• Formally, given a partition (A,B) of {1,…,n}, write an input y=(y1,…,yn)∈Sn to f in the form (yA,yB), where yA consists of the y variables in A and yB consists of the y variables in B. Think of f as mapping an input pair (yA,yB) to an output pair (zA,zB). Define the “effective information” EI(A→B) as H(zB | A random, yB=xB) EI(A→B) is the Shannon entropy of the output variables in B, if the input variables in A are drawn uniformly at random, while the input variables in B are fixed to their values in x. It’s a measure of the dependence of B on A in the computation of f(x). Similarly, define EI(B→A) := H(zA | B random, yA=xA).

• Consider the sum: Φ(A,B) := EI(A→B) + EI(B→A).

• Intuitively, we want the integrated information Φ=Φ(f,x) be the minimum of Φ(A,B), over all 2n-2 possible partitions of {1,…,n} into nonempty sets A and B. The idea is that Φ should be large, if and only if it’s not possible to partition the variables into two sets A and B, in such a way that not much information flows from A to B or vice versa when f(x) is computed.

Conceptual Summary

• Clearly define measureable & testable terms: – 2 major obstacles: direct experience & intentionality. – Consciousness, intelligence: still difficult measure & test.

• Recognize fundamental physical (hardware) laws: – Moore’s law. Dennard’s scaling law.

• Identify and examine assumptions & extrapolations – Exponential growth rates.

• Recognize fundamental engineering tradeoffs – Complexity, reliability.

• Subtle relation between intelligence & consciousness – Promising relation between simulation & consciousness. – Ability to mentally simulate the world to optimize behavior.

• This is the best time to be in computer science! – Innovations needed at all levels of abstraction. – Lots of exciting research across full HW / SW stack.

Slide Notes http://en.wikipedia.org/wiki/Moore's_law

Location of this pdf: http://cse.ucdenver.edu/~dgnabasik/Food%20for%20Thought%20II%20-%20Singularity.pdf

12 - [*] http://www.biomedcentral.com/1471-2202/5/42/

13 - http://www.imdb.com/character/ch0002900/quotes

16 - Moore's second law, also called Rock's law, which is that the capital cost of a semiconductor fab also increases exponentially over time.

17 - As of 2014, the highest transistor count in a commercially available CPU is over 4.3 billion transistors, in Intel's 15-core Xeon IvyBridge-EX. http://en.wikipedia.org/wiki/Transistor_count

20 - "PPTExponentialGrowthof Computing" by Coutesy of Ray Kurzweil and Kurzweil Technologies, Inc. - en:PPTExponentialGrowthof_Computing.jpg. Licensed under Creative Commons Attribution 1.0 via Wikimedia Commons - http://commons.wikimedia.org/wiki/File:PPTExponentialGrowthof_Computing.jpg#mediaviewer/File:PPTExponentialGrowthof_Computing.jpg

21 - "PPTMooresLawai" by Courtesy of Ray Kurzweil and Kurzweil Technologies, Inc. - en:Image:PPTMooresLawai.jpg. Licensed under Creative Commons Attribution 1.0 via Wikimedia Commons - http://commons.wikimedia.org/wiki/File:PPTMooresLawai.jpg#mediaviewer/File:PPTMooresLawai.jpg

22 - [1] http://www.extremetech.com/computing/165331-intels-former-chief-architect-moores-law-will-be-dead-within-a-decade

22 - [2] Hilbert, Martin; López, Priscila (2011). "The World’s Technological Capacity to Store, Communicate, and Compute Information". Science 332 (6025): 60–65. Bibcode:2011Sci...332...60H. doi:10.1126/science.1200970. PMID 21310967. Free access to the study through www.martinhilbert.net/WorldInfoCapacity.html and video animation ideas.economist.com/video/giant-sifting-sound-0

23 - [^] The chip started out as a software model running in one of the "Blue-Gene" series of super computers. It was then moved into FPGAs a few years back, and now into its own ASIC implemented using a 28nm process.

23 - [3] http://www-03.ibm.com/press/us/en/pressrelease/44529.wss

23 - [4] 'Neurogrid: A Mixed-Analog-Digital Multichip System for Large-Scale Neural Simulations,'Proceedings of the IEEE (doi: 10.1109/JPROC.2014.2313565)

23 - [5] http://news.stanford.edu/pr/2014/pr-neurogrid-boahen-engineering-042814.html

27 - [6] Journal of Consciousness Exploration & Research | July 2011 | Vol. 2 | Issue 5 | pp. 691-705

Additional References Adami, C. (2006). “What Do Robots Dreams Of?” Science 314 (5802): 1093-1094.

Aleksander, I. (2000). How to Build a Mind. London, Weidenfeld & Nicolson.

Aleksander, I. (2001). “The Self 'out there'.” Nature 413: 23.

Aleksander, I. and H. Morton (2005). “Enacted Theories of Visual Awareness, A Neuromodelling Analysis”. in BVAI 2005, LNCS 3704.

Atkinson, A. P., M. S. C. Thomas, et al. (2000). “Consciousness: mapping the theoretical landscape.” Trends in Cognitive Sciences 4 (10): 372-382.

Baars, B. J. (1988). A Cognitive Theory of Consciousness. Cambridge, Cambridge University Press.

Baars, B. J. (2002). “The Conscious Access Hypothesis: origins and recent evidence.” Trends in Cognitive Sciences 6 (1): 47-52.

Bongard, J., v. Zykov, et al. (2006). “Resilient Machines Through Continuous Self-Modeling.” Science 314 (5802): 1118-1121.

Chella, A., S. Gaglio, et al. (2001). “Conceptual representations of actions for autonomous robots.” Robotics and Autonomous Systems 34 (4): 251-264.

Chella, A. and R. Manzotti (2007). Artificial Consciousness. Exeter (UK), Imprint Academic.

Chrisley, R. (1995). “Non-conceptual Content and Robotics: Taking Embodiment Seriously”. in Android Epistemology. F. K, G. C and H. P., Cambridge, AAAI/MIT Press: 141-166.

Chrisley, R. (2003). “Embodied artificial intelligence.” Artificial Intelligence 149: 131-150.

Cotterill, R. M. J. (1995). “On the unity of conscious experience.” Journal of Consciousness Studies 2: 290-311.

Dennett, D. C. (1991). Consciousness explained. Boston, Little Brown and Co.

Drestke, F. (2000). Perception, Knowledge and Belief. Cambridge, Cambridge University Press.

Edelman, G. M. and G. Tononi (2000). A Universe of Consciousness. How Matter Becomes Imagination. London, Allen Lane.

Franklin, S. (2003). “IDA: A Conscious Artefact?” in Machine Consciousness. O. Holland. Exeter (UK), Imprint Academic.

Haikonen, P. O. (2003). The Cognitive Approach to Conscious Machine. London, Imprint Academic.

Harnad, S. (1990). “The Symbol Grounding Problem.” Physica D (42): 335-346.

Harnad, S. (1995). “Grounding symbolic capacity in robotic capacity”. in "Artificial Route" to "Artificial Intelligence": Building Situated Embodied Agents. L. Steels and R. A. Brooks. New York, Erlbaum.

Hesslow, G. (2002). “Conscious thought as simulation of behaviour and perception.” Trends in Cognitive Sciences 6 (6): 242-247.

Hesslow, G. (2003). Can the simulation theory explain the inner world? Lund (Sweden), Department of Physiological Sciences.

Holland, O., (2003). Machine consciousness. New York, Imprint Academic.

Jennings, C. (2000). “In Search of Consciousness.” Nature Neuroscience 3 (8): 1.

Kuipers, B. (2005). “Consciousness: drinking from the firehose of experience”. in National Conference on Artificial Intelligence (AAAI-05).

Manzotti, R. (2005). “The What Problem: Can a Theory of Consciousness be Useful?” in Yearbook of the Artificial. P. Lang. Berna.

Manzotti, R. (2006). “An alternative process view of conscious perception.” Journal of Consciousness Studies 13 (6): 45-79.

Manzotti, R. (2007). “From Artificial Intelligence to Artificial Consciousness”. in Artificial Consciousness. A. Chella and R. Manzotti. London, Imprint Academic.

McCarthy, J. (1995). “Making Robot Conscious of their Mental States”. in Machine Intelligence. S. Muggleton. Oxford, Oxford University Press.

McDermott, D. (2001). Mind and Mechanism. Cambridge (Mass), MIT Press.

Minsky, M. (1991). “Conscious Machines”. in Machinery of Consciousness, National Research Council of Canada.

Minsky, M. (2006). The Emotion Machine: Commonsense Thinking, Artificial Intelligence, and the Future of the Human Mind. New York, Simon & Schuster.

Noë, A. (2004). Action in Perception. Cambridge (Mass), MIT Press.

O' Regan, K. and A. Noe (2001). “A sensorimotor account of visual perception and consciousness.” Behavioral and Brain Sciences 24 (5).

Revonsuo, A. (1995). “Consciousness, dreams, and virtual realities.” Philosophical Psychology 8: 35-58.

Rockwell, T. (2005). Neither ghost nor brain. Cambridge (Mass), MIT Press.

Sanz, R. (2005). “Design and Implementation of an Artificial Conscious Machine”. in IWAC2005, Agrigento.

Shanahan, M. P. (2005). “Global Access, Embodiment, and the Conscious Subject.” Journal of Consciousness Studies 12 (12): 46-66.

Sloman, A. (2003). “Virtual Machines and Consciousness.” Journal of Consciousness Studies 10 (4-5).

Stein, B. E. and M. A. Meredith (1999). The merging of the senses. Cambridge (Mass), MIT Press.

Taylor, J. G. (2002). “Paying attention to consciousness.” Trends in Cognitive Sciences 6 (5): 206-210.

Ziemke, T. (2001). “The Construction of 'Reality' in the Robot: Constructivist Perpectives on Situated Artificial Intelligence and Adaptive Robotics.” Foundations of Science 6 (1-3): 163-233.

Nanotechnology

• Richard Jones: Rupturing the Nanotech Rapture – Molecular nanotechnology reduces all material things to the

status of software. – Molecular machines can do parallel processing (duroquinone). – Complex nano-assembly is certainly possible, but the watery

nanoscale environment of cell biology is hostile to engineering! – Biology: lack of rigidity, excessive stickiness/friction, constant

random motion, strong surface forces: propels the self-assembly of sophisticated structures.

– The range of environments in which rigid nano-machines could operate is quite limited.

– DNA itself can be used as a construction (self-assembly) machine to make programmed structures.

– Brain-interface systems have already been built (256 neurons) – New polymer, nano-particle based photovoltaics solar cells.