In Philip K. Dick’s science fiction classic ‘Do Androids Dream of Electric Sheep?’, Deckard has to wrestle with determining whether others are conscious or not – human or android. Back in reality, I struggle with the idea that a mobile phone such as a Google Android – indeed, anything electronic – could be conscious. The talk looks at modelling brain behaviour in computers, covering a number of well-known ideas in the philosophy of mind in the process.
A Machine Functionalist would maintain that if a computer simulation of a human brain shows the same intelligent behaviour as a human, then it must be conscious.
I argue that, without knowing what it is that causes consciousness, we do not know whether the models used in the simulation are adequate for consciousness to arise, even though they are adequate for intelligence.
However, I would have to accept that such a simulation was possibly conscious, despite how incredulous it seems to me how ‘just a bunch of NAND gates’ could be so. But my bio-chauvinism might just be because of the age in which I’ve grown up and future generations may readily accept the idea of machine consciousness.
One of the foundations of ethics is the identification of what is a moral agent and what is not. If consciousness is a factor in our selection, we then need some other criterion for deciding what is and what is not conscious.
- Could Machines be Conscious?
- Mobile Phones
- Could an Android phone be conscious?
- Scientific Realism vs Anti-Realism
- Physicalism and Functionalism
- Top-Down: Psychology
- Bottom Up: Neurons
- Models of Neuron Behaviour
- Model analogy
- Explanation without Understanding
Below: terse notes to be expanded upon later.
1. Could Machines be Conscious?
- Philip K. Dick
- ‘Do Androids Dream of Electric Sheep?’
- Deckard: How to Distinguish between Human & Android? (Voigt-Kampff test of empathy)
- Human: conscious; empathetic; moral
- Android: unconscious imitation; cold – lacking empathy; expendible
- This is a talk about Consciousness:
- Specifically: Could Non-biological stuff e.g. computers be consciousness?
- In this talk ‘Consciousness’: feeling/something it is like to be/qualia
- Not: intelligence/ cognition
- Science fact, not fiction…
2. Mobile Phones
- Hear ‘Android’ – think of mobile phones
- Google – Android – mobile – iPhone
- Microsoft – Windows – PC – Mac
- So when I say ‘could Androids dream of electric sheep’ I mean…
- Could a mobile phone be conscious?!
- Right software
- 20-30 years’ time
- Technicality: not enough memory
- True for pocket calculator
- Not for a phone…
- Universal Turing Machine
- Android phone – access internet -> Cloud
- 2010: 100billion neurons x 7000 synapses => 1000TB < £30,000
- Access all over internet.
- Very slow!
- IBM supercomputer
- 2019: ‘Blue Brain’ http://www.epfl.ch/bluebrain
- Can download http://www.neuron.yale.edu/neuron
- 2029: futurologist Ray Kurzweil: 10^9 x human brain === $1000
- We can doubt dates
- Point: ubiquitous technology within people’s lifespan
3. Could an Android phone be conscious?
- Show of hands…
- See red, feel pain
- My initial reaction: outrageous idea!
- Design chips – example – run Android – understand completely. Beat me at chess – no mystery. But to see and feel?!
- Before diving into the problem of consciousness – wider picture
- Background digression…
- (consciousness succumbing to scientific approaches)
- Cognitive Science – consistency
- Philosophy of Science – fits in: descriptive
- Philosophical Worldview – metaphysics & ethics
- (personal view – how our thoughts of consciousness fits in with our worldview)
- free will – illusion – not tonight
- what is conscious? – liberalism vs chauvinism
- liberalism extreme: table conscious!
- chauvinism extreme: solipsism
- in the middle: biological/non-biological issue – tonight
- Within philosophy of science…a big issue – which camp…
5. Scientific Realism vs Anti-Realism
- Scientific Realism:
- ‘true description of the world’
- Fermions & Bosons really exist
- Lavoisier – better theory of combustion
- James Burke’s ‘The Day the Universe Changed’
- Scientific theories – Underdetermined
- Rucks and folds in the tablecloth of Space-time
- Fantastic regularity/coherence/consistency
- Model: ‘as if’
- (end of digression)
6. Physicalism and Functionalism
- Get back towards consciousness
- Issue is on one hand…
- Brain: – patently conscious; issue = just material?
- – doctrine = physicalism = materialism = realism
- The only stuff there is is matter
- Computer: – patently just material; issue = conscious?
- – doctrine = functionalism
- Implementation of function is not relevant
- computer / beer cans and windmills
- Machine functionalism – computational functionalism
- ‘Computational theory of mind’: thought = computation
- Outrageous! But… [anti-realist position] does a model work? Try it!
7. Top-Down: Psychology
- Two ways of cracking a complex problem: top-down or botton-up
- Behaviour of whole brain; break down
- Daniel Dennett: partway to ‘Consciousness Explained’
- demolish Cartesian dualism & materialism
- less than we think
- Inattentional Blindness: Invisible Gorilla Test
- Self = ‘bundles of perception’
- (Not epiphenomena)
8. Bottom Up: Neurons
- Bottom up – in contrast
- Francis Crick – critical of Dennett – not talking about neurons.
- Neural Correlates of Consciousness
- Thinking of playing tennis
- 40Hz oscillations of firings in a particular region of the brain
- information representation?
- Neural firings: timing not important ‘pulse density modulation’
- No matter…
- not GOFAI
- simulate => get
- Let’s try…
- Imagine – paper instructions on seat – blinking (sometime involuntarily)
- Obvious objection – not enough people here
- Ned Block
- ‘China Brain’
- show red flag – favourite colour!
- anti-functionalist though experiment
- Gerald Edelman
- Not approximations: Artificial Neural Networks /connectionism
- Brain-like behaviour to solve problems
- We need the right connectivity and the right behavioural model of neurons
- Let’s look at…
9. Models of Neuron Behaviour
But a question immediately arises: Is it a model:
- for behaviour?
- for consciousness?
(not necessarily the same)
Time for some science fiction and zombies…
- Star trek teleport Gerry to the pub; scan 1 Angstrom; send (by mobile phone); 3D printer. But don’t destroy the original. Gerry in the pub is a PPD. Conscious?
- Atom-for-atom facsimile.
- Consciousness not logically supervenient on the physical (could be a zombie; walk/talk but light’s on but nobodies home)
- But is naturally supervenient (i.e. not a zombie in this natural world).
- Other type of zombie – functional.
- Scan; simulate (on mobile phone). Functional simulation – conscious, or a functional zombie?
- David Chalmers: Natural and logical supervenience
- Philosophical zombies: physiological (PPD) & functional
- Natural and logical supervenience
- Bring in…
Roger Penrose: ‘The Emperor’s New Mind’
- Consciousness is unknown;
- Whatever underlying quantum mechanics is unknown. Ergo…
- Maybe Consciousness is caused by whatever underlying quantum mechanics!
- Stuart Hameroff
- quantum coherence in microtubules in skeleton of neurons
- speculated: quantum coherence not just within neurons but between them.
- Back to China Brain:
- We got behaviour – Wow! – but just a zombie because we didn’t have the right model.
- Then: extra rules: wait until Z blinks before blinking – loose analogy – timing between neurons IS important.
- Suddenly – right behaviour, and we really have consciousness!
- How do we know we have the right model? (behaviour is the same!)
- Given the ‘right model’, is the explanation satisfactory? (No!)
Labour these 2 points with an analogy…
10. Model analogy
Due to …Paul Churchland … maybe? Explanation:
Physics => Electricity – battery & light bulb
Consciousness => Magnetism – magnet & iron filings
- Simple circuit – no understanding
- Complex circuit(20^e6 transistors – messy loops and coiling of wires)
- Discover magnetism around coils in wire – explanation (correlation)
- Real explanation: Electromagnetism
- Maxwell’s equations – underlying both electricity and magnetism.
Continue with the analogy…
- Electronics simulation models
- Absurdly low level – atomic – never done.
- Other extreme – aggregates of transistors – in practice lots here; fast.
- In between: transistor level models; logic/electrical/with mag effects/ full EM field solver.
High-level – abstraction
Logical – idealized transistor
Electrical capacitative crosstalk
Electrical + inductive crosstalk
Electromagnetic Maxwell: 3-D field solver
- Point: you need to know conscious is in order to know your model is including it!
- Magnetism: have magnetism detector: a compass!
- But a difference: No qualia detector!
- The only thing to calibrate such an instrument against is ourselves!
Going back to the unsatisfactory explanation issue, it seems like the best we can do is have…
11. Explanation without Understanding
But this is just, and only just, what science provides. Scientific explanation:
- Chalmers: structure & dynamics but nothing else.
- Explanation = correlation
- Anti-realist: models of reality: understanding is not what science does
- ‘Aha!’: unification of 2 separate phenomena to new underlying explanation
- e.g. Newtonian revolution: paradigm shift: action at a distance: Moon – string
- No problem for us now but difficult for people at the time to ‘get’
- Joseph Levine & the Explanatory Gap: Epistemological
- Paul Churchland: Undergraduates no problem with brains as ‘just a bunch of neurons’ (Crick)
- I think… when born – computers only in institutions. Lego, Meccano – mechanical
- Children’s generation growing up where ‘intelligent’ electronic gadgets ubiquitous. They will accept: computers could be conscious.
- ‘Inference to Best Explanation’: Current neural models ARE adequate. No need for quantum mechanics
- Implementation is irrelevant.
- Consciousness: fundamental property
- Psychophysical laws: Methodological Gap
- Computer simulations of brains
- Point: Naïve functionalism: good simulation => get phenomenon: false
- Cannot tell what model is adequate
- No real help on machine consciousness!
- Help understanding of biological consciousness: pathology
- Could Androids dream? ‘No!’ => ‘Quite Possibly’!
- Science: Explanation without understanding: correlation
- Psychophysical explanation: difficult!
- Depends on worldview!
- Future generations may readily accept machine consciousness
- Tide is against chauvinists
- [Result of show of hands poll?]
- Maybe problem:
- like androids in Dick’s novel, I lack empathy towards fellow conscious beings; unable to identify with them –
- not chained to my phone like some.
“If one hasn’t felt a kind of vertigo of astonishment, when facing the thought that consciousness is a wholly physical phenomenon in every respect, then one hasn’t begun to be a thoughtful materialist. One hasn’t got to the starting line.”
The title of this talk is obviously derived from the famous Philip K. Dick science-fiction novel ‘Do Androids Dream of Electric Sheep?’ which was popularized by the cult film ‘Bladerunner’.
In both the book and the film, the main protagonist, Deckard, is a bounty hunter whose job is to track down androids, verify they are indeed androids (for these androids are physically identical to humans) and destroy them. For humans are conscious, empathetic and moral whereas androids are deemed to be unconscious imitations, cold and expendible. Deckard must administer the so-called ‘Voigt-Kampff test’, a psychological empathy test, to distinguish between the conscious and the unconscious. Yet, the androids can be remarkably human (Rutger Hauer’s ‘tears in the rain’ soliloquy in the film) and Deckard wonders whether he himself is an android.
2. Science Fact
But in this talk, we are interested in science fact rather than science fiction. When I ask ‘Could androids dream of electric sheep?’ I am asking ‘could non-biological stuff like an Android mobile phone be conscious?’ By ‘conscious’, I mean actually having the same sort of thing to the awareness, sentience, ‘something it is like to be’, ‘qualia’ or whatever else you want to call it. And by ‘could?’, I mean not now but in say 20 or 30 years’ time – a human generation span. So in around the year 2040, when we might be further ahead in our understanding of consciousness than we are at present, could we take say a 2013 ‘Jelly Bean’ Android smartphone, download an app and run it and for it then to be conscious?(!)
3. No Imitations!
I need to make it clear I am not talking about imitating consciousness. Chatbots have developed significantly since Joseph Weizenbaum’s ELIZA program famously fooled students briefly in the 1960s. And the human-computer interface has improved too. The combination of the two can give us a Siri/Google Now voice-recognizing speaking ‘digital assistant’ on our phones now. But imagine if we had a button on the smartphone’s screen and every time we pressed the button the ‘digital assistant’ let out a howling shriek of pain. We would not for one moment think there was anything related to pain actually going on inside the phone. This is clearly just imitation and not the real deal. We might seek to prove it by getting hold of the source code to find where it went something like ‘if button-press then make-shrieking-sound’.
4. Whole Brain Simulation
One approach to get the real deal of conscious on a phone (or any other computer), would be to do a ‘whole brain simulation’. We could:
- take a computer model of a neuron (for example, from here),
- work out how thousands of neurons connect together to form cortical columns (see picture of 5 columns below)
- simulate a complete sheet of cortical columns – the huge cortical sheet that gets crumpled up to fit inside your skull.
Now, we could simulate all 85 billion-odd neurons on a smartphone. We wouldn’t have enough memory on the phone to store all the information but we could use cloud storage accessible via the mobile network for that. It’s true this would be incredibly slow; it would take years (real-time) to simulate a fraction of a second (simulation-time). So it might not be a practicable proposition. But that’s not the point. This is a thought experiment. We could do it. And if we did, would it be conscious?
It might be better to use a supercomputer to actually do the simulation, whilst we could still use our phone to interact with it. So, now with an Android phone communicating across the internet to the supercomputer, we could have the Siri-like interface as before. But this time, pressing the ‘pain’ button on the screen would cause the supercomputer to fire the sense neurons corresponding to the big toe of the whole brain simulation. Those signals would go up the brain stem and might provoke an ‘ow’ response. Now, do we think there would be any pain this time?
5. It’s Happening Now
This might all sound a bit like the science fiction I said I wasn’t going to talk about. But work is already underway towards this. For example, In 2005 the ‘Blue Brain Project’ was started at EPFL, Lausanne. Using an IBM ‘Blue Gene’ supercomputer, they are working towards a whole brain simulation. Milestones for the project are:
- 2006: simulation of a rat cortical column (10000 neurons, 10^8 synapses).
- 2011: ‘mesocircuit’ of 100 rat cortical columns (10^6 neurons).
- 2014: whole rat brain: 10,000 cortical columns, 10^8 neurons.
- 2023: whole human brain: 10^11 neurons.
Then, in 2013, the EU allocated €1B funds to the new ‘Human Brain Project’ which effectively subsumes the Blue Brain Project. To get an idea of the overall scope of the project, here is an overview of the various Subprojects:
- Brain Simulation platform: software, based on the Blue Brain Project. For simulating mouse and human brains.
- High-Performance Computing platform: IBM JUQUEEN Blue Gene supercomputer at Jülich Forschungszentrum, near Cologne.
- The Mouse brain atlas: U. Edinburgh/U. Madrid. We can experiment with mice . And it is a springboard towards…
- The Human brain atlas: Jülich Forschungszentrum, near Cologne.
- Cognitive architectures: top-down approach mapping understood cognitive tasks to simulation stimulation/response CEA, Paris (French atomic energy authority).
- Mathematical/theoretical abstraction: establishing a new http://www.eitn.eu/ European Institute for Theoretical Neuroscience (EITN), Paris.
- Neuroinformatics: organization and accessing of brain data. EPFL / Karolinska Unstitutet, Stockholm.
- Medical informatics: managing clinical data to get practically useful results from simulations. CHUV (hospital), Lausanne.
- Neurorobotics: connecting the brain simulation to a virtual environment. TUMünchen/EPFL.
- Ethics and Society issues.
- Neuromorphic Physical Models: analogue/mixed-signal, U. Heidelberg.
- Neuromorphic Multi-core: digital, building on the U. Manchester SpiNNaker project.
And there are other big projects going on, these being two significant ones Stateside:
- The ‘Human Connectome Project’/ (2009-2014). fMRI and diffusion MRI mapping of major connectivity in 1200 individuals. NIH / Washington U., St. Louis / U. Minnesota / Harvard.
- The new ‘BRAIN Initiative’ (2013-2023). NIH/DARPA/NSF.
Noone Gets Wet in a Simulation of a Tropical Storm
Briefly back to the question of whether a whole brain simulation could really feel pain, you might ask: why would we expect a simulation to have consciousness? After all, it’s just a simulation! It’s not the real thing. Noone gets wet in a simulation of a tropical storm!
simulating the weather in a region, such as over Japan, below. Imagine. Divide the volume into 1km3 cubes and model wind speed, direction, pressure, water vapour etc. Can see predictions of rain but obviously no wetness. Divide further into 1m^3 cubes (now have 20×10^15 instead of 20 million cubes). Now model the terrain underneath, including height. Now model surface water and run-off into rivers, out to sea (and evaporation back into the air). Seeing the water movement in the simulation, could point and say – there is liquid flow. Divide further to 1mm^3 (there would now be 20×10^24 cubes – astronomically large!). Now, individual raindrops falling out of the sky would be modelled. As the raindrops hit the ground, or any other modelled object, it would run off it just like in the fluid simulation, below. Now I think you can say there would be wetness in the simulation. So whether you get a particular phenomenon in a physics simulation depends upon the level of the modelling in the simulation.