Introduction Last time: Why are people smarter than machines? - different hardware (or "wetware") lots of slow, dumb, noisy components that cooperate to solve problems: a massively parallel machine What are the outstanding features of human computation? -constraint satisfaction -noise tolerance -interaction of many types of knowledge What are the features of connectionist models? -simple processing units -connected via weighted links -spread activation and inhibition -relax to a solution More models: The skilled typist People type fast! Interstroke interval is shorter than the time it takes for a signal to go from hand to brain and back. How do they do it? Must be a parallel, overlapping process Model: More models: Interactive activation model People are better at perceiving letters when they are in words: Paradigm: Time *A=& 0 #### 70 ms later A later: forced choice R In a word: WAKE #### A R In a "pseudoword": MAVE More models: Interactive activation model Model: More models: Interactive activation model Behavior: More models: Interactive activation model This model demonstrated: context effects: top-down activation helps perception Rule-like behavior: acts as if it knows the orthographic rules of English (pseudoword effect) Noise tolerance: cleans up noisy inputs even for novel inputs! More models: The Jets & the Sharks Motivation Humans have a content-addressable memory I'm thinking of someone. They are: a former actor intelligent a former president old You probably retrieved the right person, even though part of the description is incorrect! Computers access memories differently: To get an item out of memory, you need: an address (where it is in memory) the CPU gets the contents of that address puts it in a working memory (register) More models: The Jets and the Sharks The domain Basically, the interactive activation model applied to memory Suppose we have two gangs: More models: The Jets and the Sharks The Model More models: The Jets and the Sharks Properties 1. Content addressable We can retrieve all of someone's name from their properties 2. Graceful degradation Works even with some misleading information: "best match search" 3. Default assignment Missing features can be "filled in" 4. Spontaneous generalization Even though only specific properties of individuals are encoded prototype formation