Topic 6 - Connectionism

Question 1

Word Superiority

Answer 1

• People identify letters in words (‘k’ in “work”) more quickly than in non-words (‘p’ in “krmp”)
• If we identify the letters before we identify the word, how does whether or not it’s a word effect our letter detection?

Question 2

Interactive Activation

Answer 2

• Graph of nodes with three levels: feature, letter, and word
• Feature nodes activate/inhibit letter nodes, which activate/inhibit word nodes, which in turn activate/inhibit the letter nodes again
• Tosses serial processing out the window

Question 3

Localist Representations

Answer 3

• Each unit represents just one concept

- If this were the case, losing a single neuron would render us unable to think about a specific concept

Question 4

Distributed Representations

Answer 4

• Lots of different units working in tandem create concepts

- Seeing the forest for the trees (individual neurons don’t have much significance, it’s what they do collectively)

Question 5

Progressive Differentiation

Answer 5

• Children differentiate between different concepts

- Different ideas become more distinct with time as they are learned

Question 6

Canary Network

Answer 6

• Tree of nodes that have different relations (is, can, has, etc.)
• Experience changes the strength of the connections
• Representations from symbolic systems goes out the window
• Big concepts are learned more quickly (plant vs animal) than specific ones (canary vs robin)

Question 7

SRN (Elman net)

Answer 7

• Trying to predict what words will come next
• Intermediate units between input and output refer back to previous “context” units
• End up with similarities/links between different items

Question 8

Third-generation Connectionist Models (general trend)

Answer 8

• Neural networks can grow more complex with time
• Not every unit works the same as every other unit
• In general, we have more experience as opposed to more “lines of code”

Question 9

Connectionism Main Idea

Answer 9

-We need to understand physical implementation so that our “algorithmic” implementation isn’t entirely off-base

Question 10

Parallel Distributed Processing

Answer 10

-Bottom-up and top-down processes at the same time

Question 11

Subsymbolic

Answer 11

-Connections are not between concepts, as individual neurons do not “contain” concepts

Question 12

Dimentia

Answer 12

-Canary network in reverse; specific concepts are slowly lost, then more and more general ones