Transcript New AI 7
A New Artificial Intelligence 7 Kevin Warwick Embodiment & Questions Issues of Modern AI • We will look here at some of the important questions facing AI today • We will open up some of the directions being taken • We will attempt to move away from the restrictions imposed by Classical AI Brains • A brain has different neuronal structures each with a specialised role – sensory, motor, inter • Neurons communicate through BINARY (not analogue) codes • We know something about the physical chemical aspects of the brain • We know almost nothing about how memories are encoded or faces are recognised Innate Knowledge • Can learning occur on a blank slate? • Must there be some prior bias? • Are memories inherited? • Meaningful convergence of ANNs depends on number of neurons + topology + learning • Is this also true of a brain? • Are there hard wired cognitive biases? Genetics/emergence • Darwinian (natural) selection – shapes individual behaviours • AND/OR • Lamarckian evolution – offspring inherit acquired characteristics (e.g. giraffe) • LEAD TO • Strengthening of particular circuits in the brain & weakening of others Plato – unsupervised learning? • How can you enquire, Socrates, into that which you do not already know? • What will you put forth as the subject of the enquiry? • And if you find out what you want, how will you ever know that this is what you did not know? • i.e. how can we know we are someplace when we do not know where we are going? Questions • Perceptions depend on distributed neural codes – how are these combined? • What we perceive is highly dependent on how our brain attempts to interpret a situation/scene – how? • How does an individual acquire language? • How does a brain index temporally related information? Agents + Emergence • Idea - The mind is organised into sets of specialised functional units (Minsky) • Modular theories good for agents • Emergent globally intelligent behaviour arises from the cooperation of large numbers of agents • Supported by fMRI scans Piaget • Humans assimilate external phenomena according to our present understanding • We accommodate our understanding to the demands of the phenomena Kant • Schemata – apriori structure used to organise experience of the external world • Observation is not passive and neutral but active and interpretive Perception • Perceived information never fits precisely into our schemata • Depends on I/O devices – in humans and robots • With different I/O the real world will be perceived differently • Each entity has a different concept of reality • There is NO absolute reality! (Berkeley) Embodiment in cognition • Classical AI – instantiation of a physical symbol system is irrelevant to its performance – structure is important (Brain in a vat) • New AI - Intelligent action requires a physical embodiment that allows the entity to be integrated in the world • Present day robot I/O limited – requires more complexity in interfacing Culture • Classical AI – Individual mind is the sole source of intelligence • But knowledge is a social construct – an understanding of the social context of knowledge and behaviour is also important (memes!) Interpretations - Communication • Symbols are used in context – a domain has different interpretations, depending on the goals • Sign interpretation – coding system • The meaning of a symbol is understood in the context of its role as an interpretor Falsifiable Computation • Any number of confirming experiments are not sufficient for confirmation of a theory • Scientific theories must be falsifiable • There must exist circumstances under which a model is a poor approximant • Many computational models are not falsifiable – universal machines! • Need computation that is falsifiable Let’s Move On • Classical AI – (Hobbes/Locke/Aristotle) – • • • intelligent processes conform to universal laws and are understandable/modelable Converse (Winograd/Penrose/Weisenbaum) – important aspects of intelligence cannot be modelled A model/simulation is not the real thing The only ‘exact’ simulation of a human brain would be that specific human brain and no other – even then it would need to be in its place/time Differences • Just because something is different does not make it worse • A simulation of a human brain could be more/less intelligent/conscious/selfaware/understanding • Models/simulations are used to explore, explain & predict – if a model is proven to be accurate for this then that’s just fine Comments on Intelligence • As long as we understand the basics of what intelligence is, that is sufficient • We should not be bogged down by trying to copy exactly the functioning of the human brain, interesting though that might be • More interesting is to create entities that are intelligent in their own right Next • Growing Brains – Biological AI Contact Information • Web site: www.kevinwarwick.com • Email: [email protected] • Tel: (44)-1189-318210 • Fax: (44)-1189-318220 • Professor Kevin Warwick, Department of Cybernetics, University of Reading, Whiteknights, Reading, RG6 6AY,UK