Introduction to Deep Learning

What if we would like to have prediction output (binary classification) represented by probability, which activation function is the best choice?

• tanh

• ReLu

• Sigmoid

• Linear

In the context of deep learning, computer determines derivative of a given function using mathematical formula

• True

• False

Given a deep neural network shown in the figure, what is the dimension of weight at the 3rd layer?

• 5 x 4

• 4 x 5

• 4 x 3

• 3 x 4

We can directly pass the matrix representation of a color image as an input of the deep neural network

• True

• False

Given a color image of size 120 x 80 pixels, what is the dimension of its feature vector?

• 9,600

• 28,800

• 8,400

• 25,200

What is CORRECT about gradient descent?

• It returns the optimum parameters that minimize the cost function

• It returns the optimum parameters that maximize the cost function

• It returns the optimum parameters that minimize the loss function

• It returns the optimum parameters that maximize the loss function

The minimum point is the point that slope changes from negative to positive

• True

• False

What is NOT the property of sigmoid function?

• Input is any real number

• Output is a real number in range between 0 to 1

• Nearly linear when input is close to zero

• Slope is close to zero when input is close to zero

What is the limitation of perceptron?

• Cannot solve non-linearly separable problem

• Cannot handle very large data

• Assume normal distribution

• Assume independent of features

What if the learning rate is too high?

• Model is difficult to converge

• Model consumes more computational resources

• Model takes longer time to converge

• Model is difficult to generalize