Environment and Goals Jointly Direct Category Acquisition / Wikibook - Semantic Networks

Love Article

• Why and how do people categorize?
• why- to develop a schema that corresponds well to their real life. We need that so we aren't routinely mislead.
• how- from either exemplars or prototypes (maybe both? we're unsure)
• What issue is important but ignored by prototype and exemplar?
• The author argues that its not so clear, but we have to take into account flexibility of structures as well as people's goals. Instead we should use clusters.
• clusters: bundles of experiences that group together - a more flexible way of representing info.
• Encompass both prototypes and exemplars in that a category represented by one cluster is a prototype and one category with many clusters is an exemplar.
• exemplar prototype continuum (a spectrum)
• Knowledge gets more complex the more we learn, and this model follows that.
• The author says the world comes in natural chunks, so it makes sense that we would make clusters out of those things that naturally go together
• from this perspective, clustering is driven by the environment, but it may also be influenced by goals. -ex. you may look at a cow differently if you are a vet or a butcher.

Wikibook - Semantic Networks

• A node represents a concept, also stores various characteristics of that concept.
• a link represents what ties different concepts together.
• They argue that it is stored heirachically in nodes.
• demonstrated that the time it takes to respond correlates with the distances in the networks.
• Collins and quillian were key in the move to semantic networks.
• Semantic network: evidence that things are stored by specificity. Specifics are stored, and the rest of the process is inferred, (but its not always this way).
• ex. you can say very quickly that a sports car is fast and needs fuel.
• Spreading activation: adjacent concepts are activated, and then are more easily retrieved fro memory, they are joined. Thinking of one thing makes you think of another.
• This was proved with the lexical decision task - is this a word or isnt it?
• Typicality effect: reaction times are faster for more typical members of a category
• Connectionistic Approach (vs. semantic approach, saying a node is not just one concept): A concept is distributed across other networks - Multiple elements working together.
• Neural network:
• input layer - sensory perception/from environment
• hidden layer- inter-neurons between the two
• output layer- what we do in response to those things
• Connectivist approach may be more likely as they are more complex networks, not just a single model.
• Neural networks work well in practice
• Parallel distributed processing- processing takes place in parallel lines, output is distributed across many units (which leads to neural networks), not a single node in a network.
• so, if you destroy one unit, the whole thing won't break down.
• Computational Knowledge Representation
• knowledge is full of useful information, rather than as supporting a model of cognitive activity. 
• it doesn't come before the actual cognitive or neural processes, but it is something we draw on during those.
• Applications of knowledge Representation: computational knowledge representation provides tools to make knowledge accessible.

Love article continued:

• He uses a neural network approach to understand concepts: SUSTAIN model about categorization
• Clusters: like the hidden layer (represented by circles on diagram) formed mental averages... refine themselves.
• 2 kinds of learning: supervised and unsupervised.
• new clusters are created when old clusters fail for goals (supervised)
• SUSTAI N changes when its surprised by experiences, and creates new clusters.
• Inference: related to concepts - you know that something is in a category but you don't know a lot about it... but you can infer. A missing factor inferred from others.
• classification categorization - focuses on information that distinguishes categories.
• the SUSTAIN model, he claims, can handle both types of learning.
• he found inference learning leads them to more complete knowledge - best for classroom learning.
• Two groups with different goals (like fisherman) and can draw different inferences from the same information - the SUSTAIN model can account for differences in goals, like this.
• SUSTAIN is good for the relationship between clusters and rules: clusters with selective attention are mimicking rules/ maybe they then are rules.
• SUSTAIN takes into account rules.
• conclusion: categorization - searching for regularities in the world (taking into account what's in the environment (and differing goals). It starts simple, then experience leads to added clusters as experiences and surprises are encountered.