Introduction to Deep Learning

Question 1

Let’s say you want to train a fully connected deep neural network using only
fully connected layers; after training for 1000 epochs, your training accuracy
is 98.7%, but your test accuracy is 67%. Which of the aspect/aspects of the
model should you change to tackle this problem? (Multiple choice)
a) I would increase the batch size.
b) I would add some drop-out layers.
c) I would increase the learning rate ten times and would train only for 100
epochs.
d) I would add batch normalization layers

Answer 1

b) I would add some drop-out layers

d) I would add batch normalization layers

Question 2

What’s the advantage of adding a Maxpooling layer over a fully connected
layer or a convolutional layer? (Multiple choice)
a) It’s computationally less expensive than a convolutional layer or a fully
connected linear layer.
b) It has no trainable parameter.
c) It normalizes the batch data.
d) It has more trainable parameters than a convolutional layer but carries a
lesser computational cost than a linear layer.

Answer 2

a) Its computationally less expensive than a convolutionally layer of a fully connected linear layer

b) It has no trainable parameters