S01 s/d S05 - Deep Learning

Because of the mathematics of backpropagation, the earliest layer of Neural Networks might learn almost nothing from the error signal as the number of layer grows. This phenomenon is known as ...

• low accuracy problem

• high loss problem

• vanishing gradient problem

• underfitting problem

Neural Networks before 2010 use sigmoid as the activation function of all hidden layers.

• True

• False

The followings are the factors that make deep learning yields exceptional performance, except ...

• large amount of data

• the existence of tensor

• GPU acceleration

• ability to train deeper Neural Network architecture

How to train deeper Neural Network architecture? (more than 1 answers)

• Use residual connection

• Use batch normalization

• Use Sigmoid

• Use ReLU or its variant

Which one of the following aspects that that is NOT the activities involved in building deep learning model architecture?

• creating learnable parameters

• optimizer

• network model

• data evaluation

1-D tensor is also called ...

• matrix

• scalar

• hash

• vector

Image is an example of 4-D tensor.

• True

• False

Given the following tensor: x = torch.FloatTensor([2.1, 3.0, 5.7, 4.9, 2.8]). Which one is NOT the content of x[:3]?

• 3.0

• 4.9

• 2.1

• 5.7

The followings are the optimizers used in deep learning, except ...

• Linear

• SGD

• Adam

• RMSprop

The function that takes a real-valued number and outputs a number in a range between zero and one is ...

• Linear

• ReLU

• Tanh

• sigmoid