10 – Memory 5 – representation of knowledge Flashcards
In Collins and Loftus’s Hierarchical Network model (1972), time to make semantic decisions is a function of what?
Distance between nodes in the hierarchy.
e.g. canary is a bird < a canary is an animal
A canary has wings < a canary has skin
What evidence against Collins and Loftus’s Hierarchical Network is there?
Time to make decisions is not influenced only by distance between node, but also by typicality and ambiguity.
e.g. a sparrow is a bird< a lark is a bird - sparrows and larks should be at the same level in the hierarchy
a bat is not a bird> a dog is not a bird - bats and birds are unconnected in this model
In what way does Collins and Loftus’s Hierarchical Network model contradict what we know about how children learn words?
Children don’t learn about animals, then birds, then types of birds. Other way round –there’s a duck! Doesn’t fit with the nature of children’s experience.
What network did Collins and Loftus suggest in response to criticism of the Hierarchical Network model? And how does this new model work?
Associationist Network. Nodes represent concepts, and proximity reflects strength of semantic relationship. These structures emerge from the co-occurence (contiguity) of concepts –a child is more likely to hear orange, green, blue in the same context.
This allows for strong associations to develop between concepts that share some conceptual relationship (red and fire engine, red and roses) even though they’re not in the same taxonomical category.
How does the Parallel Distribution Processing model work?
When you perceive an item - e.g. canary – various nodes are activated which represent the various properties and relations of the item. A structure corresponding to the concept ‘canary’ develops by these nodes being activated or inhibited. Also, the connection weights between nodes vary with repeated exposure, so the system gradually comes to represent reality more accurately.
How does the weighting system result in error correction in the PDP model?
Every node in this network receives activation from nodes projecting to it and sums it according to weighting equation. Then propagates that activation to other nodes it’s connected to. Network initially starts off with these weights being random. Weightings occur as inputs are combined.
Discrepancy between actual pattern of activation for ‘canary’ and the correct pattern is BACK-PROPAGATED through the network. It’s used to adjust the weights throughout the connections so that next time the output will match the input more closely.
In the PDP model, knowledge is learned by gradually adjusting ______ _______ until the network can reproduce the correct ______ for all trained inputs. Learning MUST be _________.
Knowledge learned by gradually adjusting connection weights until the network can reproduce the correct output for all trained inputs
What are 2 explanatory advantages of the PDP metaphor?
- Spontaneous generalisation - e.g. category learning. If the network has learned the concept for robin. If it learns properties for robin, then learns properties for canary, which activates similar connections. When sees kookaburra, easy to identify as sharing properties. Also accounts for interference from atypical examples –a penguin is a bird even though it doesn’t fly/ bat flies but not bird.
- Graceful degradation – e.g. semantic dementia.Pps don’t lose particular bit of knowledge.
In the PDP model, the the general abstract semantic knowledge of the concept corresponds to what?
“Average activity” across patterns for separate instances of the item.
What can PDP systems NOT do easily?
- Learn exceptions (e.g. penguin, bat)
– One-trial learning
– AB-AC learning: catastrophic interference
These problems all arise from the need for very gradual learning of ‘semantic memory’
What role might the hippocampus play in consolidating learning?
Activated information is held in the hippocampus, then incorporated in the neocortex. The hippocampus and MTL might bind together activated information to form new ‘episodic’ memories.
How are episodes constructed in memory?
Episodes consist of concept nodes and nodes for the context in which we see a stimulus. These concepts get bound together into episode nodes in the hippocampus.
Recognition tasks of implicit memory activate the ______ node, while recollection in explicit memory tasks activates the _______ node.
Recognition tasks of implicit memory activate the concept node, while recollection in explicit memory tasks activates the episode node.
How long does it take for memories to be consolidated in the neocortex? What evidence is there for this?
Around two years. When there is hippocampal damage and anterograde amnesia, not only can people not store new memories, but also showed retrograde amnesia for two years prior. Memories of last two years may not be fully consolidated.
