Computational Cognitive Modelling Lecture 1 18.11.2018 Computational Cognitive Modelling COGS 511-Lecture 1 General Introduction 18.11.2018 COGS 511 - Bilge Say COGS 511
Related Readings From Course Pack Lecture 1 18.11.2018 Related Readings From Course Pack Cooper, R. Chapter 1: Modelling Cognition McClelland (2009). The Place of Modeling in Cognitive Science. References (extra and optional; given for a complete reference list – not in the course pack) Carpenter and Just, Computational Modeling of High-Level Cognition versus Hypothesis Testing in Sternberg (ed), The Nature of Cognition, 1999. Fernandez, J. Explanation by Computer Simulation in Cognitive Science, Minds and Machines, 13: 269-284, 2003. Steedman, Chap. 5, of Scarborough and Sternberg (eds). Morgan, M.S., & Morrison, M. (1999). Models as mediators (Ed). Cambridge: Cambridge University Press. 18.11.2018 COGS 511 - Bilge Say COGS 511
Models A representation of something that may be used in place of the real thing, abstracting away unimportant features but retaining the essential. (Cooper). A good model is complete (does not abstract out important properties) and faithful (does not introduce features that are not in the original) with respect to its specific purpose. Helpful for understanding a complex system – cognition for the case of cognitive science. Computational cognitive modelling is the development of computer models of cognitive processes and the use of such models to simulate and predict human behaviour. 18.11.2018 COGS 511 - Bilge Say
Models in Philosophy of Science Lecture 1 18.11.2018 Models in Philosophy of Science The task of Philosophy of science is: Generate reflections on the theoretical and methodological issues in scientific practice. Models function in a variety of different ways within sciences. Analog models: Molecules – Billiard-balls, Mechanical model: DNA molecule - Metal-made helix model Scale models: Models in architecture, model airplanes, etc. Treated in relation with theory and phenomena. 18.11.2018 COGS 511 - Bilge Say COGS 511
Models in Philosophy of Science (cont.) Semantic View Models are abstract idealized systems which characterize how the phenomena would have behaved if the idealized conditions were met (Suppe, 1989). Thus, a theory characterizes the model which represents (certain aspects of) phenomena. 18.11.2018 COGS 511 - Bilge Say
Models in Philosophy of Science (cont.) Morrison and Morgan (1999) Models are evaluated in response to four questions: How are models constructed? What do they represent? What role do they have/how do they function in scientific practices? How do we learn from models? 18.11.2018 COGS 511 - Bilge Say
Models in Philosophy of Science (cont.) Their general account based on case studies in physics, chemistry and economy proposes that: Models are autonomous agents, i.e. they are only partially dependent on theories and phenomena Models serve as instruments for investigation in science. 18.11.2018 COGS 511 - Bilge Say
Models in Philosophy of Science (cont.) How are models constructed? Not derived entirely from theory or phenomena Involve both, and also additional “outside” elements (modeling decisions). 18.11.2018 COGS 511 - Bilge Say
Models in Philosophy of Science (cont.) What do they represent? Some aspect of the phenomena or some aspect of theories 18.11.2018 COGS 511 - Bilge Say
Models in Philosophy of Science (cont.) What role do they have/how do they function in scientific practices? Function as tools or instruments. 18.11.2018 COGS 511 - Bilge Say
Models in Philosophy of Science (cont.) How do we learn from models? Not by looking at a model, but by building and manipulating it. 18.11.2018 COGS 511 - Bilge Say
Computer Science vs Cognitive Science Program: data structures + algorithms= running programs Representation: Implied by the architecture, mathematical definition of the problem, design specification of the task, the software paradigm used Algorithms: Simplicity, efficiency and complexity trade-offs. Mind = mental representation + computational procedures = cognition Representation: Cognitive Architecture or the ontology of human mental process is not given. Hope: algorithms and representations posited will clarify the architecture, too. Algorithms: Performance on realistic data, simplicity in terms of plausibility 18.11.2018 COGS 511 - Bilge Say
Artificial Intelligence vs Cognitive Science Lecture 1 18.11.2018 Artificial Intelligence vs Cognitive Science The study and automation of intelligent behaviour (Luger & Stubblefield) Success: Commercial/Performance – as described by proposals such as Turing test (?) or in a limited domain aI: the study of human intelligence with computer as a tool (Yeap, 97) vs Ai: the study of machine intelligence as artificial intelligence Theoretical, experimental or applied (Rumelhart) Failures (?): Frame problem, syntax vs semantics/intentionality The study of cognition, mental activity involving acquisition, storage transformation and use of knowledge; study of mental processes such as memory, language, thought, perception, consciousness .... Success: “Competence” -explanatory power of a cognitive theory: pyschological and neurological plausibility, computational and representational power, practical applicability to education, design etc. Define AI. Alternative Definitions: building intelligent machines (rational, goal-directed agents?) Study of how to make computers do things, which at the moment humans do better (?) – Rich and Knight Intelligence (?) Rationality: do your best accordance with your goals, given what you know Turing Test: Computer cannot be distinguished from a human (originally a gender game) by an human interrogator Criticisims of Turing test: Can rocks imitate? Behaviorist, limited, syntactic (Searle’s Chinese Room argument) Variations: Total Turing Test, Inverted Turing Test List: cognitive processes you know Theoretical AI: algorithms, procedures, etc inspired by human or biological intelligence Experimental AI: testing these by writing programs and running behavioural experimentation Applied AI: Using them for some practical purpose Frame problem: Failure to come into terms with the nonrepresentational Character of background knowledge of a typical human being. 18.11.2018 COGS 511 - Bilge Say COGS 511
Lecture 1 18.11.2018 An Example from Chess Human experts use relatively shallow searches, averaging only three or four moves deep; perceptual patterns and their recognition play an important part in guiding the search. Chess programs rely on extensive search and optimization of search techniques. Deep Blue evaluated 200 million moves per second in 1997. Other models: planes vs birds, Galileo’s balls for gravity, a mechanical heart... 18.11.2018 COGS 511 - Bilge Say COGS 511
Computational Models in Cognitive Science Lecture 1 18.11.2018 Computational Models in Cognitive Science A computer program which implements a theory of some aspect of cognition (Green) Representations and processes of some cognitive theory made precise by analogy with data structures and algorithms (Thagard) Do computational models have to subscribe to strong AI view (aim: building machines that duplicate minds) to be useful as research tools in cognitive science? Not necessarily! Weak AI: Can machines be made to act as if they were intelligent? Is it necessary to believe either in strong or weak AI to do cognitive modelling? 18.11.2018 COGS 511 - Bilge Say COGS 511
Some Philosophical Background Functionalism: Most general features of cognition must be independent of neurology- the physical system – and the embodiment of mind. Mental states are abstract functions that get us from a given input to a given output. Cognitivism: All there is to cognition is in mental states and thought. Computational Theory of Mind ~Computational Representational Understanding of Mind: Human cognition can be best understood in terms of representational structures in the mind and computational procedures that operate on them. 18.11.2018 COGS 511 - Bilge Say
Computational Theory of Mind Lecture 1 18.11.2018 Computational Theory of Mind Thought processes are computations on representations. The mind can be realized/implemented outside of the brain eg. in a digital computer. Is the mind a digital computer? Church-Turing Thesis: The Universal Turing Machine can perform any calculation that can be described by an effective procedure. How do we relate all these cognitive modelling? Do you have to be a follower of some kind of functionalism to be able think that computational modelling will contribute smt to cognitive science? Effective procedure?: Can be Done by a human clerk working with pen and paper 18.11.2018 COGS 511 - Bilge Say COGS 511
Misconceptions of Church-Turing Thesis It doesn’t say that given a standard computer, you can compute any rule-governed input-output function.It doesn’t rule out machines (or brains) that compute non-Turing computable functions. Thus, it does not entail that brains can be simulated by a Universal Turing Machine. 18.11.2018 COGS 511 - Bilge Say
Questions Can a certain approach contradict with Computational Theory of Mind (mental representation + computational processes = cognition) and still involve computational modelling ? (Yes –see dynamical approaches) Do you have to ascribe to a functionalist view (mental states are abstract functions – can be described independent of brain states) to do computational modelling ? (No – see computational neuroscience) 18.11.2018 COGS 511 - Bilge Say
The Status of a Computational Model “It is not the computer program that is the theory, at best they inspire the construction of a theory.” (Scheutz) “Simulation is not a reasonable goal for cognitive science.” (Fodor) “AI is to psychology as Disneyland is to physics.” (Green) “Artificial Intelligence is to cognitive science as mathematics is to physics.” (Rumelhart) 18.11.2018 COGS 511 - Bilge Say
Marr’s Levels of Analysis Computational: What information processing is being solved, and why? Algorithmic: Representation and Programming. How is the problem being solved? Implementational: What physical properties are required to build such a system? Hardware (e.g. brainstates) 18.11.2018 COGS 511 - Bilge Say
Subject Model = = = = Computational Computational Algorithmic Lecture 1 18.11.2018 Subject Model Computational = Computational One-to-many One-to-many Algorithmic = Algorithmic Against functionalism: It is only by understanding the nature of brain mechanisms that we can make good hypotheses about the workings of the mind. Do you agree? Architectural = Architectural One-to-many One-to-many Implementational = Implementational (Dawson, 98) COGS 511
Method Theory Computational Algorithmic Computer Simulations Lecture 1 18.11.2018 Method Theory Computational Behavioural Experiments One-to-many Algorithmic Computer Simulations What kind of differences do you observe between theories and models in your home discipline vs. that of cognitive science? Architectural One-to-many Implementational Cognitive and Computational Neuroscience Adapted from (Brent, 96) 18.11.2018 COGS 511 - Bilge Say COGS 511
The Function of Computational Models Computational Cognitive Model Simulates Generates Implements Behaviour Cognitive Process Describes Explains Theory Cooper (2002) – Ch.1 18.11.2018 COGS 511 - Bilge Say
Explanation by Computer Simulation (Fernandez, 2003) Lecture 1 18.11.2018 Explanation by Computer Simulation (Fernandez, 2003) Causal Explanation: The system uses a program in order to compute a certain input-output mapping. Explaining how you cooked a tasty dish Do you have enough justification for that? Functional Analysis The system executes a program which amounts to computing a certain mapping. Explaining how an car manufacturing assembly line works Multiple realizability? Analogy with an assembly line vs recipe following Searle’s “the wall implementing the Wordstar program” analogy.. 18.11.2018 COGS 511 - Bilge Say COGS 511
Advantages of Computational Modelling Clarify, formally and unambiguously specify a certain cognitive theory Create experimental participants that are durable, flexible etc. – in silico Allow detailed evaluation and exploration of cognitive theories by means of raising new hypotheses Enable interaction between studies in different disciplines Not THE method, but a complementary method 18.11.2018 COGS 511 - Bilge Say
Strategies Develop a model of some task or behaviour in order to learn more about it: “a fishing trip” Implement a pre-existing, verbally specified highly complex theory to see if its theoretical assumptions are sufficient/necessary to account for the target behaviour. Generate predictions/hypotheses to be then tested by behavioural experiments. Platform: Cognitive models of individual processes vs “unified” approach – cognitive architectures 18.11.2018 COGS 511 - Bilge Say
Evaluation of Models Behavioural Outcome Modelling: Roughly showing similar behaviours as human beings Qualitative Modelling: Same qualitative behaviours that characterize human behaviour, e.g. similar improvement, deteoriation Quantitative Modelling: Similar quantitative behaviour as exhibited by humans, indicated by quantitative performance measures A combination of the above (Sun, 98) 18.11.2018 COGS 511 - Bilge Say
Practical Problems with Cognitive Modelling Lecture 1 18.11.2018 Practical Problems with Cognitive Modelling Goodness-of-fit problems Individual Differences Incidental Details Problems- scalability and sensitivity analysis needed Problematic Predictive Power Statistical interpretation varies as compared to hypothesis-testing statistics usage in psychology Theory-model amalgamation Complexity and understandability trade-offs Isolated modelling – not enough interaction with different levels of theorizing and methods. “Practical” meaning leaving the philosophical objections aside... Incidental Details: which aspects of the model are necessary, sufficient or irrelevant for the qualitative behaviour of the model. Compare how hypothesis testing proceeds... 18.11.2018 COGS 511 - Bilge Say COGS 511
Method Theory Computational Algorithmic Computer Simulations Lecture 1 18.11.2018 Method Theory Computational Behavioural Experiments One-to-many Algorithmic Computer Simulations Architectural One-to-many Implementational Cognitive and Computational Neuroscience Adapted from (Brent, 96) 18.11.2018 COGS 511 - Bilge Say COGS 511
Paradigms in Computational Modelling Symbolic systems – best for accounting for rationality, systematicity etc. of symbol systems? Connectionism – biologically plausible ? Dynamicisim – best for exploring embodied, situated, temporal cognition? Hybrid approaches Similar Division in AI: GOFAI – Good, Old Fashioned AI vs NFAI – New Fangled AI 18.11.2018 COGS 511 - Bilge Say
Achievements for Cognitive Modelling Shaping theories for various cognitive domains: language and skill acquisition, individual differences in working memory, cognitive lesioning simulations and neuropsychology. Applied areas: Human-computer interaction, intelligent-tutoring systems 18.11.2018 COGS 511 - Bilge Say
Future for Cognitive Modelling Integration of Computational Neuroscience and more abstract forms of cognitive modelling – e.g. Blue Brain project More interaction between Artificial Intelligence and Cognitive Modelling – esp in Cognitive Architectures More emphasis in hybrid models – symbolic, dynamic, connectionist, bayesian etc. 18.11.2018 COGS 511 - Bilge Say
Lecture 2 Unified Theories of Cognition Cognitive Architectures 18.11.2018 Lecture 2 Unified Theories of Cognition Cognitive Architectures Sample Architectures vs Frameworks Reading: Langley, Laird and Rogers (2009) Cognitive Architectures Start Readings for the project and think about your project groups. Check Forum for online activity. 18.11.2018 COGS 511 - Bilge Say COGS 511