Automating Science: Panel – Stephen Ames, John Wilkins, Greg Restall, Kevin Korb
A discussion among philosophers, mathematicians and AI experts on whether science can be automated, what it means to automate science, and the implications of automating science – including discussion on the technological singularity.
– implementing science in a computer – Bayesian methods – most promising normative standard for doing inductive inference
– vehicle : causal Bayesian networks – probability distributions over random variables showing causal relationships
– probabilifying relationships – tests whose evidence can raise the probability
– 05:23 does Bayesianism misrepresent the majority of what people do in science?
– 07:05 How to automate the generation of new hypotheses?
– Is there a clean dividing line between discovery and justification? (Popper’s view on the difference between the context of discovery and context of justification) Sure we discuss the difference between the concepts – but what is the difference between the implementation?
– 08:42 Automation of Science from beginning to end: concept formation, discovery of hypotheses, developing experiments, testing hypotheses, making inferences … hypotheses testing has been done – through concept formation is an interestingly difficult problem
– 9:38 – does everyone on the panel agree that automation of science is possible? Stephen Ames: not yet, but the goal is imminent, until it’s done it’s an open question – Kevin/John: logically possible, question is will we do it – Greg Restall: Don’t know, can there be one formal system that can generate anything classed as science? A degree of open-endedness may be required, the system will need to represent itself etc (Godel!=mysticism, automation!=representing something in a formal deductive theory)
– 13:04 There is a Godel theorem that applies to a formal representation for automating science – that means that the formal representation can’t do everything – therefore what’s the scope of a formal system that can automate science? What will the formal representation and automated science implementation look like?
– 14:20 Going beyond formal representations to automate science (John Searle objects to AI on the basis of formal representations not being universal problem solvers)
– 15:45 Abductive inference (inference to the best explanation) – & Popper’s pessimism about a logic of discovery has no foundation – where does it come from? Calling it logic (if logic means deduction) is misleading perhaps – abduction is not deductive, but it can be formalised.
– 17:10 Some classified systems fall out of neural networks or clustering programs – Google’s concept of a cat is not deductive (IFAIK)
– 19:29 Map & territory – Turing Test – ‘if you can’t tell the difference between the model and the real system – then in practice there is no difference’ – the behavioural test is probably a pretty good one for intelligence
– 22:03 Discussion on IBM Watson on Jeopardy – a lot of natural language processing but not natural language generation
– 24:09 Bayesianism – in mathematics and in humans reasoning probabilistically – it introduced the concept of not seeing everything in black and white. People get statistical problems wrong often when they are asked to answer intuitively. Is the technology likely to have a broad impact?
– 26:26 Human thinking, subjective statistical reasoning – and the mismatch between the public communicative act often sounding like Boolean logic – a mismatch between our internal representation and the tools we have for externally representing likelihoods
29:08 Low hanging fruit in human communication probabilistic reasoning – Bayesian nets and argument maps (Bayesian nets strengths between premises and conclusions)
– 29:41 Human inquiry, wondering and asking questions – how do we automate asking questions (as distinct from making statements)? Scientific abduction is connected to asking questions – there is no reason why asking questions can’t be automated – there is contrasted explanations and conceptual space theory where you can characterise a question – causal explanation using causal Bayesian networks (and when proposing an explanation it must be supported some explanatory context)
– 32:29 Automating Philosophy – if you can automate science you can automate philosophy –
– 34:02 Stanford Computational Metaphysics project (colleagues with Greg Restall) – Stanford Computational Metaphysics project – formalization of representations of relationships between concepts – going back to Leibniz – complex notions can be boiled down to simpler primitive notions and grinding out these primitive notions computationally – they are making genuine discoveries
Weak Reading: can some philosophy be automated – yes
Strong Reading of q: can All of philosophy be automated? – there seem to be some things that count as philosophy that don’t look like they will be automated in the next 10 years
– 35:41 If what we’re is interested in is to represent and automate the production of reasoning formally (not only to evaluate), as long as the domain is such that we are making claims and we are interested in the inferential connections between the claims, then a lot of the properties of reasoning are subject matter agnostic.
– 36:46 (Rohan McLeod) Regarding Creationism is it better to think of it as a poor hypothesis or non-science? – not an exclusive disjunct, can start as a poor hypothesis and later become not-science or science – it depends on the stage at the time – science rules things out of contention – and at some point creationism had not been ruled out
– 38:16 (Rohan McLeod) Is economics a science or does it have the potential to be (or is it intrinsically not possible for it to be a science) and why?
Are there value judgements in science? And if there are how do you falsify a hypothesis that conveys a value judgement? physicists make value judgements on hypothesis “h1 is good, h2 is bad” – economics may have reducible normative components but physics doesn’t (electrons aren’t the kinds of things that economies are) – Michael ??? paper on value judgements – “there is no such thing as a factual judgement that does not involve value” – while there are normative components to economics, it is studied from at least one remove – problem is economists try to make normative judgements like “a good economy/market/corporation will do X”
– 42:22 Problems with economics – incredibly complex, it’s hard to model, without a model exists a vacuum that gets filled with ideology – (are ideologies normative?)
– 42:56 One of the problems with economics is it gets treated like a natural system (in physics or chemistry) which hides all the values which are getting smuggled in – commitments and values which are operative and contribute to the configuration of the system – a contention is whether economics should be a science (Kevin: Yes, Stephen: No) – perhaps economics could be called a nascent science (in the process of being born)
– 44:28 (James Fodor) Well known scientists have thought that their theories were implicit in nature before they found them – what’s the role of intuition in automating science & philosophy? – need intuitions to drive things forward – intuition in the abduction area – to drive inspiration for generating hypothesis – though a lot of what get’s called intuition is really the unconscious processing of a trained mind (an experienced driver doesn’t have to process how to drive a car) – Louis Pasteur’s prepared mind – trained prior probabilities
– 46:55 The Singularity – disagreement? John Wilkins suspects it’s not physically possible – Where does Moore’s Law (or its equivalents in other hardware paradigms) peter out? The software problem could be solved near or far. Kevin agrees with I.J. Good – recursively improving abilities without (obvious) end (within thermodynamic limits). Kevin Korb explains the intelligence explosion.
– 50:31 Stephen Ames discusses his view of the singularity – but disagrees with uploading on the grounds of needing to commit to philosophical naturalism
– 51:52 Greg Restall mistrusts IT corporations to get uploading right – Kevin expresses concerns about using star-trek transporters – the lack of physical continuity. Greg discusses theories of intelligence – planes fly as do birds, but planes are not birds – they are differing
– 54:07 John Wilkins – way too much emphasis is put on propositional knowledge and communication in describing intelligence – each human has roughly the same amount of processing power – too much rests on academic pretense and conceit.
– 54:57 The Harvard Rule – under conditions of consistent lighting, feeding etc – the organism will do as it damn well pleases. But biology will defeat simple models.. Also Hulls rule – no matter what the law in biology is there is an exception (inc Hull’s law) – so simulated biology may be difficult. We won’t simulate an entire organism – we can’t simulate a cell. Kevin objects
– 58:30 Greg R. says simulations and models do give us useful information – even if we isolate certain properties in simulation that are not isolated in the real world – John Wilkins suggests that there will be a point where it works until it doesn’t
– 1:00:08 One of the biggest differences between humans and mice is 40 million years of evolution in both directions – the problem is in evo biol is your inductive projectability – we’ve observed it in these cases, therefore we expect it in this – it fades out relatively rapidly in direct disproportion to the degree of relatedness
– 1:01:35 Colin Kline – PSYCHE – and other AI programs making discoveries – David Chalmers have proposed the Hard Problem of Consciousness – pZombies – but we are all pZombies, so we will develop systems that are conscious because there is to such thing as consciousness. Kevin is with Dennet – info processing functioning is what consciousness supervenes upon
Greg – concept formation in systems like PSYCHE – but this milestone might be very early in the development of what we think of as agency – if the machine is worried about being turned off or complains about getting board, then we are onto something