COM 1070 Introduction to Artificial Intelligence week 10

COM 1070: Introduction to Artificial Intelligence: week 10 Yorick Wilks Computer Science Department University of Sheffield www. dcs. shef. ac. uk/-yorick

Characteristics of von Neumann architecture n n n A von Neumann machine is a sequential processor; executes instructions in a program one after another. Each string of symbols stored at specific memory location. Access to stored string via numerical address of string’s location. Single seat of control, central processing unit, CPU. But brain may still be a computer, just not a von Neumann machine, or possibly a VNM with enormous redundancy of data.

• Ability of GOFAI and Neural Nets to provide an account of thought and cognition The debate is about whether it is possible to provide a neural computing (NN) account of cognition, or whether we should assume that a symbol system (GOFAI, or Good Old Fashioned AI) is required. Physical symbol system hypothesis: it is possible to construct a universal symbol system that thinks Strong symbol system hypothesis: only universal symbol systems are capable of thinking. I. e. anything which thinks (e. g. human brain) will be a universal symbol system. I. e. all human thinking consists of symbol manipulation. I. e. only computers can think – therefore we are computers. If it’s not a universal symbol system, it can’t think!

Pylyshyn: an advocate of strong symbol system hypothesis Our mental activity consists of manipulation of sentence like symbolic expressions. Human thought = manipulation of sentences of internal mental language. Fodor’s Language of Thought or Mentalese We can show that brain does not have a von Neumann architecture. But this does not disprove Strong Symbol System Hypothesis. Strong symbol system hypothesis says nothing about architecture.

Could have symbol manipulation in parallel system n Could have different memory system: i. e. contentaddressable storage n Instead of CPU could have different parallel control. Brain can form representations of events in world. SSSH: representations in brain take form of sentencelike strings of symbols. I. e. sentence ‘Fido licks Brian’ is symbolic representation of fact that Fido licks Brian. Language of thought, or Mentalese representations: not literally English, but like sentences, in that they have basic vocabulary units, and form in which they are arranged determines their meaning. n

Advantages of mentalese Provides an account of beliefs, intentions and doubts etc. these are also expressed in terms of mentalese sentences. Provides an account of productivity and systematicity. Human language is productive, (no limit to number of sentences we can produce). And systematic (if you can say John loves Mary, you can also say Mary loves John) Human thought: productive and systematic because it relies on Mentalese, which is productive and systematic. Origin of all this in Chomsky’s generative linguistic theories.

Neural Computing: can it form an alternative to SSSH? Is it possible to provide a connectionist account of thought, or is it as SSSH advocates would claim, impossible? Symbols: Subsymbolic Hypothesis versus Strong Symbol system hypothesis Symbol System Hypothesis (Newell and Simon, 1976) ‘. . a physical symbol system has the necessary and sufficient means for general intelligent action. . ’ A symbol designates something. A symbol is atomic (cannot be broken down further). E. g. ‘elephant’ designates an elephant. Or P could designate elephant. Or 01100 could designate elephant, but with no interpretation of the 1 s and 0 s.

Compositional symbol = compound symbol which has meaningful parts, whose overall meaning is determined by meaning of those parts. E. g. sentences of natural language. The kangaroo jumped over the elephant. Distinction between symbol types and symbol tokens E. g. in AMBIGUITY, 9 letter tokens, 8 letter types. Same symbol type can be realised in many different ways. Symbol is both (a) a computational token, and (b) it designates something, i. e. it is a representation.

Connectionism: subsymbolic hypothesis (see Smolensky, 1988). Symbols can be broken down, computation takes place at subsymbolic level. Connectionist representations: distributed representation. Distributed representation = pattern of activity over several nodes. E. g. in a distributed representation of ‘elephant’, ‘elephant’ is represented by distributed representations over several nodes. Thus, there is not an atomic symbol for elephant.

So connectionism rejects Strong Symbol System Hypothesis To complicate matters, one can have localist neural networks, where concepts are represented by single nodes. (e. g. node which represents ‘elephant’). Localist connectionism assumes its symbols given and may be compatible with the SSSH. This discussion applies to distributed (non-localist) representations.

Distributed representations = subsymbolic computation. E. g. representing letter A n Localist scheme, single unit for A, one for B etc. n Or (in a subsymbolic system) letters represented as patterns of activity across 78 units. E. g. A = units 1, 2, and 3, B = units 4, 5, and 6 etc. n Individual units stand for features of letters. Thus letter A will be joint activity of various features it contains. So letter E will share several features with letter F. Thus similarities and differences among items are reflected in similarities and differences among representations.

Symbolic-Subsymbolic distinction. In a symbolic system the computational level coincides with the representational level. In a subsymbolic system, the computational level lies beneath the representational level. Symbol is both (a) a computational token, and (b) it designates something, i. e. it is a representation. But in subsymbolic Connectionism: representations across several units, but computational tokens are units. But does the same issue reappear? Subsymbolic computation may not assume the symbols but what about their features--e. g. horizontal strokes for E and F? ?

Symbolic criticisms of connectionism Arguments in favour of Strong Symbol System Hypothesis Fodor and Pylyshyn (1988) Argument that connectionism is inadequate as a representational system. Focus on issues of compositionality and systematicity Fodor and Pylyshyn (1988) argue n Compositionality and structure sensitivity are necessary to support cognitive processes. n Only classical representations exhibit compositionality and structure sensitivity (systematicity) n Only classical representations can support cognitive processes.

Compositionally concatenative representations: Where complex representations are composed from primitive tokens combined concatenatively Molecular representations can be formed out of constituents, and can be manipulated in accord with syntactic rules. E. g. ‘the kangaroo jumped over the elephant’ If you can understand this, you can also understand a sentence which has the same words in a different order. (Systematicity). E. g. ‘The elephant was jumped over by the kangaroo’ or ‘The elephant jumped over the kangaroo’

Two sentences are composed of same elements. Constituent elements manipulated according to syntactic rules. But, according to Fodor and Pylyshyn, connectionist representations cannot be manipulated like this. F&P: cannot compose simple connectionist representations into more complex representations Much of this argument comes down to the role of tree structures or hierarchies, which are needed to express syntactic relationships --- SSSH people say trees cannot be learned by connectionist systems. . But Jordan Pollack showed in 1986 that they can (up to a point anyway)--this argument is very like a return of the XOR argument at a higher level.

Connectionist Counter-Arguments Distributed representations: nonsymbolic and continuous (developed over hidden units). But van Gelder (1990), can have functional compositionality ‘. . We have functional compositionality when there are general, effective and reliable processes for (a) producing an expression given its constituents, and (b) decomposing the expression back into those constituents…’ Many nets can do (a) and (b) For example, net can be trained to structurally disambiguate sentences. n John saw money with the telescope. n (John (saw money) (with the telescope)).

Complex representations of input sentences developed over hidden units. Then these representations decomposed into required output (ii). Thus, fully distributed representations carry information about syntactic structure of inputs, without being syntactically structured. I. e. they demonstrate a functional compositionality by moving from structures to their components and back again.

. But can distributed representations permit structure sensitive operations? (systematicity) (e. g. changing from active to passive). Fodor and Mc. Laughlin (1990): : to support structure sensitive operations, representations must contain explicit tokens of original constituent parts of complex expression. But can show that connectionist representations can have property of systematicity (permitting structure sensitive operations).

Chalmers (1990), trained a connectionist net to transform representations of active sentences, to passive sentences. n Developed representations for both active and passive sentences. Used RAAM nets (Pollack, 1990) to do this. n Took the fully distributed representations from the RAAM nets, for both active and passive sentences. n Trained a net to translate from active sentences to passive sentences. Training complete: could input active sentences, extract representation, and translate that into the representation for a passive sentence, which could be decoded into the passive sentence.

Also generalised to sentences that were not part of the training set. I. e. Connectionist representations did permit structure sensitive operations, without being decoded into symbols.

Summary of representation argument Claimed by Fodor and Pylyshyn that n Compositionality and structure sensitivity are necessary to support cognitive processes. n Only classical representations exhibit compositionality and structure sensitivity (systematicity) n Only classical representations can support cognitive processes. BUT n Can demonstrate that connectionist nets show functional compositionality. Although representation do not contain tokens, as symbolic representations do, these tokens can be obtained from the representations.

E. g. Can train nets to disambiguate sentences like the following. Intervening hidden unit representation does not contain explicit tokens, but output does. n John saw money with the telescope. n (John (saw money) (with the telescope)) n Can also show that connectionist representations can support structure sensitive operations (systematicity). Chalmers (1990) Translating from active to passive sentences using connectionist distributed representations. So counter-argument against claims that neural nets cannot provide an account of cognition. So Neural Computing can provide an alternative to the Strong Symbol System Hypothesis.

But has yet to be shown if can provide a connectionist account of all aspects of thought. Smolensky, (1988): ‘. . It is far from clear whether connectionist models …have adequate… power to perform high level cognitive tasks. . ’ Connectionism: Does well at accounting for low-level aspects of cognition: e. g. movement, pattern recognition. But Beliefs? Intentions? Can Neural Computing provide an account of ‘Direct Conscious Control’? (Norman, 1986) E. g. consciously planning what to do, introspecting about our thoughts, holding beliefs, making logical inferences, early stages of learning a skill.

Possible that brain best modelled in terms of hybrid system: Connectionist account of lower level processes Symbolic account of higher level processes.

Adaptive Behaviour, and Symbol Grounding Revisited Three different approaches to Artificial Intelligence, two of which we have already encountered many times: Symbolic AI or Traditional AI or GOFAI (Good Old Fashioned AI) n Neural Computing or Connectionism or Parallel Distributed Processing n Adaptive Behaviour or Behaviour-based Robotics Adaptive Behaviour Rodney Brooks, at MIT Artificial Intelligence Lab. (see reference to Brooks on web page for the course) Brooks, R. A. (1991) Intelligence without Reason. MIT AI Lab Memo 1293, April 1991 n

An example: Allen, a reactive robot. (named after Allen Newell? ) Sonar sensors, and odometer to keep track of distance travelled. Controlled by cable from off-board special purpose computer. Lowest level reactive layer; used sonar readings to keep away from moving and static obstacles. - if an obstacle is close, instead of bumping into it, stop. Second level; random wandering. Every 10 seconds, generate a movement in a random direction. Third level: Look for a distant place, and move towards it. Odometry can be used to monitor progress. Three layers made it possible for robot to approach goal, whilst avoiding obstacles.

Goal subsumption: switching control between the modules is driven by the environment, not by a central locus of control. Robot heads for goal until sensors pick up information that there is an obstacle in the way. The obstacle avoidance module cuts in. Once the obstacle has been avoided the goal-finding module takes control again. Robot can move around in the environment although it does not build, or use, any map of that environment, and is only driven by simple environmental cues.

Second example: Herbert (Herbert Simon? ) Wanders about an office environment, picking up coke cans and returning them to start point. Sensors: infrared ports, and laser 3 D data. Actuators: motors driving wheels, and manipulator arm with sensors. Subsumption architecture: several behaviour-generating modules. Modules include obstacle avoidance, wall following, and recognition of coke cans. Control of modules: Only suppression and inhibition between alternative modules - no other internal communication. Each module connected to sensors and to arbitration network which decides which competing action to take.

Description of Herbert in action: When following a wall, Herbert spots a coke can. The robot locates itself in front of the can. The arm motion is then begun - when can is detected with sensors local to the arm, it is picked up. Advantages; naturally opportunistic. If coke can put right in front of Herbert, can collect it and return to start, since no expectations about where coke cans will be found. Can find coke cans in a variety of locations, even if never found there before. But….

Behaviour-based Robotics Idea of building autonomous mobile robots. New approach, where robots operate in the world, and use ‘…highly reactive architectures, with no reasoning systems, no manipulable representations, no symbols, and totally decentralized computation’ (Brooks, 1991) ‘…I wish to build completely autonomous mobile agents that co-exist in the world with humans, and are seen by those humans as intelligent beings in their own right. I will call such agents Creatures. . . ’ (Brooks, 1991) Brooks, R. (1991) Intelligence without Representation Artificial Intelligence, 47, 139 -159.

See “Elephants don’t play chess”, (1990 paper by Brooks) Brooks, Rodney A. (1990) “Elephants don’t play chess”. In Pattie Maes (Ed) Designing autonomous Agents, Cambridge, Mass: MIT Press. Because elephants don’t play chess, no reason to assume they are not intelligent. Emphasis on kind of behaviour exemplified by elephants, rather than on more abstract human behaviours (e. g. games, speech recognition, problem solving).

A Creature must cope appropriately and in a timely fashion with changes in its dynamic environment. n A Creature should be robust with respect to its environment: minor changes in the properties of the world should not lead to total collapse of the Creature’s behaviour; rather one should only expect a gradual change in the capabilities of the Creature as the environment changes more and more. n A Creature should be able to maintain multiple goals and, depending on the circumstances it finds itself in, change which particular goals it is actively pursuing; thus it can both adapt to surroundings and capitalize on fortuitous circumstances. n A Creature should do something in the world: it should have some purpose in being. n

Set of principles (Brooks, 1991) n The goal is to study complete integrated intelligent autonomous agents. n The agents should be embodied as mobile robots situated in unmodified worlds found round laboratory. (embodiment). n Robots should operate under different environmental conditions - e. g. in different lighting conditions, when sensors and actuators drift in calibration (situatedness). n Robots should operate on timescales commensurate with timescales used by humans (situatedness).

Key Topics of Behaviour-based Approach Situatedness n Embodiment n (animal or insect) Intelligence n Emergence n

Situatedness A situated automation is a finite-state machine whose inputs are provided by sensors connected to the environment, and whose outputs are connected to effectors. The world is its own best model Traditional AI, working in symbolic abstracted domain. Problem solvers which are not participating in the world as agents. Dealing with model world - no real connection to external world.

Alternative approach, to use a mobile robot which uses the world as its own model, referring to information from sensors rather than internal world model. Representations are developed which capture relationships of entities to robot. Situated agent must respond in timely fashion to inputs; but much information from the world.

Embodiment The world grounds symbolic regress Embodiment: Physical grounding of robot in real world. According to Brooks (1991), embodiment is critical for 2 reasons. n Only an embodied agent is validated as one that can deal with real world. n Only through a physical grounding can any internal symbolic system be given meaning.

Brooksian view of Intelligence is determined by the dynamics of interaction with the world Some activities we think of as intelligent have only been taking place for a small fraction of our evolutionary lineage. ‘Simple’ behaviours to do with perception and mobility took much longer to evolve. Would make sense to begin by looking at simpler animals. - looking at dynamics of interaction of robot with its environment.

Emergence Intelligence is in the eye of the observer Intelligence emerges from interaction of components of the system. Behaviour-based approach - intelligence emerges from interaction of simple modules. e. g. Obstacle avoidance, goal finding, wall following modules.

Main ideas n No central model maintained of world n No central locus of control n No separation into perceptual system, central system and actuation system n Behavioural competence improved by adding one more behaviour specific network to existing network. Crude analogy to evolutionary development n No hierarchical development n Layers or behaviours run in parallel

Criticisms? This approach won’t necessarily lead to system capable of more complex behaviours. A new controller is needed for each task. The experimenter is deciding on what modules to add, and what environment and task the robot should be exposed to. - not the same as evolution. But in terms of evolution, new behaviours and new mental structures are learnt in response to the environment, not added by an experimenter. Similarly, in the development of an individual, new representational structures are developed in response to the environment, not added by an experimenter.

It would be more impressive if the robot learnt new behaviour modules in response to the environment. This possibility is discussed by Brooks (1991), but has not yet been successfully tackled. Emphasis in this approach on reacting to the environment. And it is the case that apparently quite sophisticated behaviours can result from simple reaction to the environment. But representations are needed for more complex tasks. e. g. ‘Find an empty can and bring it back to the starting point’ requires the formation of an internal representation corresponding to a map. Need to provide an account of the development of representations.

Symbol Grounding revisited Traditional view: the language of thought (Fodor, 1975), that The mind is a symbol system and cognition is symbol manipulation. Advocates of symbolic model of mind (e. g. Fodor, and Pylysyn) argue that symbol strings capture what mental phenomena such as thoughts and beliefs are. Symbol system: symbols (arbitrary physical tokens) manipulated on the basis of explicit rules. Rule-governed symbol manipulation is based on syntax of symbol tokens (not their meaning).

Symbols can be rulefully combined; primitive atomic symbol tokens can be combined to form composite symbol-token strings. Resulting symbol-token strings can be given a meaning - i. e. they are semantically interpretable. BUT approach of assuming that mind is symbol system can be criticised - in terms of symbol grounding.

A criticism of symbol systems is that symbol system capable of passing the Turing Test will not be a mind, because the symbols have no semantics (meaning) (remember the Chinese Room) From Searle (1997) The Mystery of Consciousness n Programs are entirely syntactical n Minds have a semantics n Syntax is not the same as, not be itself sufficient for, semantics Therefore programs are not minds. QED

‘…It does not matter how well the system can imitate the behaviour of someone who really does understand, nor how complex the symbol manipulations are; you cannot milk semantics out of syntactical processes alone…’ (Searle, 1997).

Symbol grounding, as discussed by Stevan Harnard, S (1990) The Symbol Grounding Problem. Physical D 42, 335 -346. Copy of paper can be obtained from: http: //www. cogsci. soton. ac. uk/harnad/genpub. html N. B. see the relevant websites for this course at http: //www. dcs. shef. ac. uk/~yorick/ai_course/aicourse. html

Computation consists of manipulation of meaningless symbols. For them to have meaning they must be grounded in non-symbolic base. Like the idea of trying to learn Chinese from a Chinese dictionary. Standard reply of symbolist (e. g. Fodor, 1980) is that the meaning of the symbols comes from connecting the symbol system to the world “in the right way”. But how could this be done? Harnard provides one possible solution: Symbols need to have some intrinsic semantics or real meaning.

For Harnard, symbols are grounded in iconic representations of the world. e. g. consider the symbol “horse” iconic representation of a horse, is a representation of the shapes that horses cast on our retinas (i. e. sensory surface of the eye). From these iconic representations (many from individual views of horses), we form a categorical representation - that captures the features we need to identify a horse. Thus the name “horse” is grounded in iconic and categorical representations, learned from experience.

Similarly, “stripes” is grounded in iconic and categorical representations, learned from experience. These symbols can be combined: “zebra” = “horse” & “stripes”. New symbol of zebra, is built up from the grounded representations of “horse” and “stripes”, which gives it meaning. Harnard is proposing a hybrid system, in which thought is assumed to be symbol manipulation, but the symbols are grounded in iconic and categorical representations of the world. The problem with all this is WHICH symbols are so grounded (peace, courage, Hamlet? )

Other solutions to symbol grounding problem have been proposed. Essential idea is that symbols need to be given some meaning - need for grounding in meaningful representations, to escape from circularity of defining symbols in terms of symbols. 2 other (partial) solutions n Adaptive Behaviour and embodiment n Connectionist (neural computing) n Both are robot-prosthetic arguments, the first without representations and the second with implicit ones.

Symbol Grounding and Adaptive Behaviour Would make sense to have symbols physically grounded in real world. Embodiment: n Only an embodied agent is validated as one that can deal with real world n Only through a physical grounding can any internal symbolic system be given meaning. n But adaptive behaviour people don’t want to have symbols, grounded or not.

Suggests a new approach to grounding symbolic representations - but as yet no clear account of how symbols might emerge due to such interactions with the real world. Emphasis in work on behaviour-based robotics has been on behaviour without representation. New approach, where robots operate in the world, and use ‘…highly reactive architectures, with no reasoning systems, no manipulable representations, no symbols, and totally decentralized computation. ’ (Brooks, 1991)

Symbol Grounding and Neural Nets Alternative idea: symbols are grounded in connectionist representations. Connectionist symbols = distributed representations = patterns of activation across several units. Connectionist symbols have internal structure. They are not meaningless in the same way that atomic symbols are. This is a persuasive argument made by Chalmers (1992) Subsymbolic computation and the Chinese Room. In J. Dinsmore (Ed) The symbolic and connectionist paradigms: closing the gap, Lawrence Erlbaum: Hillsdale, New Jersey. pp 25 -49.

Topics to think about n Mycin n SHRDLU n PARRY n Expert systems n Chinese Room n Turing Test n Weak & Strong AI

n Neural networks n Adaptive behaviour n Symbolic AI n Symbol grounding