340 likes | 500 Views
Copycat Architecture & IDA’s Perception. By Uma Ramamurthy Cognitive Science Seminar January 29, 2003. Introduction. Analogy-making is a process of high-level perception.
E N D
Copycat Architecture & IDA’s Perception By Uma Ramamurthy Cognitive Science Seminar January 29, 2003
Introduction • Analogy-making is a process of high-level perception. • Copycat models analogy-making where conceptual fluidity emerges from complex, subconscious interaction between perception and concepts • Distilling the essence of one situation • Adapting that to fit another situation
Copycat in detail… Makes analogies between situations in an idealized microworld of letter-string analogy problems abc abd ijk ? xyz ?
Copycat’s Task • Use concepts it possesses to build perceptual structures – • Description of objects • Bonds between objects of the same string • Groups of objects in a string • Correspondences between objects in different strings • Represent its understanding of the problem • Formulate a solution to the given problem
Copycat’s task Given: abc abd Solve: ijk ? OR: xyz ?
Copycat’s Architecture • Four main elements: • Slipnet • Workspace • Coderack (with codelets) • Temperature
Copycat’s Architecture (contd.) Slipnet: A B X Y Z Links rightmost successor opposite predecessor leftmost last first Label nodes
Copycat’s Architecture (contd.) Slipnet: • Concepts residing in a network of nodes and links • A concept’s central region is a node • A concept’s associative halo may include any of the nodes linked to the central node • A node is activated when instances of it are perceived by codelets • Activation spreads from a node to nearby nodes as a function of their conceptual proximity
Copycat’s Architecture (contd.) Slipnet: • Rate of activation decay is a function of the node’s conceptual depth • Conceptual depth – a static, pre-assigned number – captures the generality and abstractness of the concept • “opposite” is deeper than “successor” • Length of a link between nodes indicates degree of association conceptual proximity • shorter the link, greater the degree of association
Copycat’s Architecture (contd.) Slipnet: • Link lengths vary in response to the system’s perception • Label node active all links labeled by the node shrink, indicating close relationships, hence more slippable • Slippage from one node to to a neighbor is probabilistically made function of the conceptual proximity between the two nodes
Copycat’s Architecture (contd.) Workspace: • Place where perceptual structures are build heirarchically on top of the given input (3 strings of letters) • Six types of structures - • Descriptions of objects • Bonds between objects in the same string • Groups of objects in the same string • Correspondences between objects in different strings • Rule describing the change from initial to modified string • Translated Rule - how the target string should be modified to get the answer string • Structures have time-varying strength value - a function of several factors including conceptual depth and activation
Copycat’s Architecture (contd.) Coderack: • “A codelet is a piece of code that carries out some small, local task that is part of the process of building a structure” • Any run of the system starts with a standard, initial population of codelets • Codelets wait on the rack to be chosen the choice is probabilistic • Codelets that build perceptual structures send activation to Slipnet nodes, representing the concepts associated with that structure
Copycat’s Architecture (contd.) Types of Codelets: • Scouting Codelets for all structures • Strength (of a structure) Evaluating Codelets • Building Codelets • Breaking Structure Codelets • Translate Rule Codelets
Copycat’s Architecture (contd.) Temperature: • Measures degree of perceptual organization in the system • Controls degree of randomness used in making decisions • Higher temperature little information to base decisions (more random) • Lower temperature greater certainty about basis of decisions • Stopping more likely at lower temperature • Final temperature indicator of program’s satisfaction with the answer it created
Copycat System Program run: Until a rule has been built and translated, do: Choose a codelet and remove it from the Coderack; Run the chosen codelet; If N codelets have run, then: Update the Slipnet; Update the Coderack with relevant codelets; Build the answer according to the translated rule
Example Run a b c a b d i j k
Perceptual Structures A B C A B D a b c a b d I J K i j k 100 rightmostletter middleletter rightmostletter leftmostletter middleletter leftmostletter leftmostletter middleletter rightmostletter
Perceptual Structures (contd.) A B C A B D a b c a b d I J K i j k 81 rightmostletter middleletter rightmostletter leftmostletter middleletter leftmostletter leftmostletter middleletter rightmostletter Lmost Lmost Mid Mid Rmost Rmost
Perceptual Structures (contd.) Replace letter-category of rmost letter by successor A B C A B D a b c a b d I J K i j k 48 rightmostletter middleletter rightmostletter leftmostletter middleletter leftmostletter leftmostletter middleletter rightmostletter Lmost Lmost Mid Mid Rmost Rmost
Perceptual Structures (contd.) Replace letter-category of rmost letter by successor A B C A B D a b c a b d I J K i j k 31 rightmostletter middleletter rightmostletter leftmostletter middleletter leftmostletter leftmostletter middleletter rightmostletter Lmost Lmost Mid Mid Rmost Rmost
Perceptual Structures (contd.) Replace letter-category of rmost letter by successor A B C A B D a b c a b d I J K i j k 15 rightmostletter middleletter rightmostletter leftmostletter middleletter leftmostletter leftmostletter middleletter rightmostletter Lmost Lmost Mid Mid Rmost Rmost
Perceptual Structures (contd.) Replace letter-category of rmost letter by successor A B C A B D a b c a b d I J K i j k i j l 12 rightmostletter middleletter rightmostletter leftmostletter middleletter leftmostletter leftmostletter middleletter rightmostletter Lmost Lmost Mid Mid Rmost Rmost Replace letter-category of rmost letter by successor
Copycat’s Shortcomings • No Learning • No new Slipnet nodes can be created • No new Codelets can be generated • Works only in a fixed, predefined domain
IDA’s Perception Module • Slipnet representing concepts such as location, job-types, message-types • Workspace to process incoming email message – feature extraction from the message by Perceptual Codelets to activate Slipnet nodes • Filling up a template in the Workspace with mandatory fields of a message-type • No Temperature
IDA’s Slipnet • Nodes • Links • Features • Message types
Piece of IDA’s Slipnet NRFK norfolk nor Norfolk . . . Jacksonville Norfolk San Diego Miami location information request preference acceptance
IDA’s Perception Module • Output from the Perception Module goes to the Attention Registers of the Focus/Working Memory • “Internal” perception via access to Navy databases and other information stores • Input/Output piece handles incoming and outgoing email messages of IDA
In the computational model… • Fixed, pre-defined set of message-types are “perceived” • Everything “perceived” goes to long-term associative memory • No learning in the current version of IDA
In the Conceptual Model… • Everything perceived will not go to long-term memory • Perceptual Learning occurs through “curiosity” and “fringe consciousness” • A conscious interaction with declarative memory (IDA’s long-term/autobiographical memory) may result in generation of • New Slipnet nodes • Possibly new Perceptual Codelets • Slipnet nodes may decay and “disappear” from the system over time if they remain unused for LONG times • … and much more