lecture 3

Question 1

Why are multilayered networks better than shallow networks?

Answer 1

1. There are larger datasets available
2. GPUs became available
   a. With one hidden layer, the network is poor at generalization
   b. With more hidden layers, gradual transformation from input space to output space makes the out-of-set generalization superior to the one-layer hidden network

Question 2

convnets

Answer 2

• Have much fewer units to learn (as opposed to fully connected layers)
• This is because they use a biased local correlation structure
  > they only learn local relations

Question 3

probing cognition with deep neural networks
(DNNs linked to behavior)

Answer 3

 Physical shape of objects matches representation in early layers of deep-convolutional networks

 Perceived shape of objects matches representation in final layers of deep-convolutional networks

Question 4

Masked vs unmasked in the brain

Answer 4

Mask disrupts processing

For complex scenes you need recurrent processing but for simple scenes you do not

Deep network is capable of doing this task

Shallow networks are worse at this in higher masking conditions

Question 5

you can’t present stimuli for short amounts of time with neural networks.
 Single-image models

Answer 5

However, sequential models (i.e., BLnext) improve with time

o This way you can make them more like humans
o With this you can answer how-questions

Question 6

covering law model

Answer 6

what is a good scientific explanation

Quality of quantitative model is judged by
1. Interpretability
2. Predictive power

in these models interpreting == prediction
 If I can understand it, I can use it for prediction
 Simple formulas can describe even the most fundamental phenomena

Question 7

Critique of DNNs as models of behavior

Answer 7

 We want to understand, not predict
 Replacing one box with another
 A clone system might have the same behavior but does not increase understanding

i.e., DNNs are typical predictive models but not a solution for those that are interested in cognition.