Localist vs distributed coding

Question 1

Localist (grandmother cell)

Answer 1

each word represented by single unit

Question 2

distributed

Answer 2

words are coded across a set pattern of activation (each unit involved in multiple codes)

Question 3

Neural network models

Answer 3

reject localist word codes learn to name words via distributed codes
hidden layer of units so can never find full word units activate pattern of activation
knowledge distributed

Question 4

Back-propagation

Answer 4

train models to change connection weights to reduce to reduce output error.

Question 5

localist Macaque monkey study

Answer 5

high selectivity for complex info in cortex
recorded 850 neurons in IT
one cell responded selectively to 1 of 27 faces

Question 6

OJ Simpson cell

Answer 6

cell from IT in Monkey responded to selectively

Question 7

limitations of localist studies

Answer 7

• only present a set number of images
  -representations highly selective but not grandmother cells

Question 8

superposition catastrophe

Answer 8

difficult to coactivate multiple things in STM so cortical systems that support STM need to be highly selective (local) as they are unambiguous

Question 9

Example of superposition catastrophe

Answer 9

if want to code for two names at same time then add the codes together unable to tell what mean as could indicate multiple names
so cant unambiguously code for more then one thing at same time in a set of distributed units

Question 10

neural system of STM (Botvinick and Plaut)

Answer 10

successfully co-activates multiple items
input takes a pattern of activation and reproduces it at the output
units in the hidden layer activate themselves so activation persists
distributed layer hidden allows for names to be recalled in order (blend pattern)

Question 11

evidence of hidden unit response

Answer 11

plot activation for each hidden unit response to word.
unit responds to a range of inputs with no selectivity - cannot infer what was presented by looking at activation for one unit.

Question 12

simulations of Botvinick and Plaut model

Answer 12

trained on 300 words. when trained 1 syllable at a time no superposition catastrophe so distributed as no single unit codes
when trained on multiple bonding patterns occur - unit responds so localist and superposition catastrophe