04_Week 4 notes (Cognitive Architectures).pdf

Cognitive Architectures Cognitive architecture: how particular mental functions are realized in the brain, what kinds of knowledge they rely on, the steps of the processes, the kinds of principles they follow, etc. ways of framing explanations about regularities in the mind/brain what are the...

Cognitive Architectures Cognitive architecture: how particular mental functions are realized in the brain, what kinds of knowledge they rely on, the steps of the processes, the kinds of principles they follow, etc. ways of framing explanations about regularities in the mind/brain what are the wired properties of our cognitive capacities how the information is stored and how it flows from one component to the other in order to realize its input-output functions how the units of mental representation (concepts) are put together to form thoughts (plans, decisions, language comprehension, etc.) Different representationalist schools: Symbolic (Turing-like, computations) Connectionist (brain-like, associations) Symbolic Architecture characterize mental representations as symbols and mental processes as computations over symbols Symbols = mental representations or units of information Cognitive processes = “mechanical” devices that operate over representations by following rules (the rules encoded in the “machine table”)  Since the machine can be implemented physically as a sequence of “on” and “off” switches, it can be said to represent and process information (symbols in sequences) like instructions for electrical circuits, like the neurons the rules of the machine cause the machine to change states and operate over other representations following other rules Determined by input-output operations with a limited number of representations and a limited number of rules in its table of instructions, the machine can perform computations ad infinitum FEATURE-BASED APPROACH Features as the basic elements carrying meaning, one is committed to a system that computes sets of features to yield a concept: Rules for computing compilations of features Concepts need to be bound; composition LIMITATIONS rule-governed behaviour lacks the characteristic malleability (or plasticity) of human cognitive functions inadequate to model higher cognitive systems (thoughts) Connectionist Architecture conceive of cognitive systems as embodied in highly interconnected units (or “nodes”) in a network that exhibit the typical behaviour of neurons in a network such as the human cortex nodes = abstract mental representations distributed across a network any given node will activate (or “fire” output and carry on the activation) as a function of: The values (weights) of input connections the strength of the total input it receives from other nodes the threshold level (i.e., the level necessary for it to respond to input) Processes: patterns of activation of nodes Interconnected nodes can be setup as simple associations yielding a certain state (associative networks( can be feed-forward: input from one layer is passed forward to another layer can be recurrent: with feedback provided from later layers of nodes to earlier ones, thus adjusting the input to the state of the network Different types of connectionism: Parallel Distributed Processing (PDP): Multiple nodes corresponding to features (e.g., semantic features) Local: Each Node corresponds to a major category (e.g., “word”, DOG) Feature-based approach Activation: concepts are obtained by the state of the organism during a pattern of activation No rules but (quasi-) unconstrained activation No composition LIMITATIONS 1. representations are arbitrary labels assigned to nodes assigned by the modeller 2. we don’t know exactly how to conceive of the weights or thresholds that make a given representation fire (or produce an output to another node) Assumptions of Cognitive Architectures PRODUCTIVITY: infinite complex Mental Representations with a finite number of simplex ones Infinite propositions that a system can encode achieved within a finite system a need to postulate MRs that have combinatorial structure in which some elementary MRs combine to form complex ones SYSTEMATICITY: Complex Mental Representations of a given form => infinite complex Mental Representations of similar form capacity to entertain certain MRs is intrinsically connected to our ability to entertain certain others with similar form because the two expressions have the same syntactic structure COMPOSITIONALITY: The meaning/content of a complex Mental Representation is a function of of the meaning of its simplex ones and how they are structured That’s possible because simplex Mental Representations contribute the same content across complex Mental Representations ***Just activation won’t do, structure matters Contrasts PRODUCTIVITY Symbolic: finite number of symbols can form an infinite number of symbolic expressions because MR have constituent structure Connectionist: each node stands for a representation (simple or complex) adding units changes the connectivity, changes the structure the machine/brain SYSTEMATICITY Symbolic: since MR have constituent structure, if one can think P&Q one can also think Q&P Connectionist: MR do not have constituent structure Being in a state S1 (John loves Mary) is fundamentally different than being in a state S2 (Mary loves John) because they require the activation of different nodes Have to create a new node for each new concept COMPOSITIONALITY Symbolic: the thought P&Q actually contains the symbols P and Q, as the thought Q&P contains the same thoughts P and Q End product contains properties of each symbol following structure Connectionist: the thought P&Q is a function of the activation of the nodes P and Q which in turn activate the node P&Q the node P&Q does not actually contain P and Q Compromise Symbolic architectures can serve as models for cognitive processes, at least for higher-level cognitive processes thoughts, language comprehension, and conceptual knowledge Connectionist architectures seem to be well suited to model the level of physical realization of cognitive processes how neurons are connected into networks that are vehicles of mental processes to model the implementation level Symbols & Symbol Systems HISTORICAL DEFINITIONS Symbols in the mind standing for things in the world (Descartes) Symbols, in his conception are tokens to which we assign meaning (Whitehead) a physical pattern that has a semantic life (i.e., it stands for something) and that enters into complex expressions (Newell & Simon) similar to a data pointer which when accessed, retrieves data structures Computational processes rely on the continuous patterning of symbols and the access to data structures governed not by what the symbols “mean” but by how sequences of symbols are structured