mi1_02_backprop

Did you attempt the last exercise sheet?

• Yes

• No

Which logic function f(x₁, x₂) will a single neuron fail to compute?

• NOT

• XOR

• NOR

• AND

Why does the perceptron fail at XOR?

• A perceptron only works with one input.

• It cannot find a linear separation.

• It totally can. With the right bias value.

• A perceptron can only predict one value

What kind of cost function is this?

• quadratic

• linear

• round

• The e stands for exponential

Which cost function is more robust to outliers?

• A linear one

• The quadratic one

Answer explanation

In the linear case the gradient will be constant. This applies to very large error for the outlier or a small error for a "normal" point.

So what's the quadratic cost useful for?

• maximizing the likelihood

• data with no outliers

• it's completely useless

• useful for classification

The purpose of backpropagation is to...

• compute the cost

• make multiplication efficient

• compute gradients

• It is a learning algorithm.

Which cost function would you use for multi-class classification?

• softmax

• cross entropy

• mean squared error

• a linear cost

Which term is backpropagation responsible for?

• blue (first term)

• orange (second term)

• E^T

• The sum

Overfitting means...

• Memorizing the training data

• fitting the noise in the data

• Overshooting the global optimum

• generalizing well on unseen test data