mi1_07_SLT

Did you attempt any exercise sheets over the holidays?

• Yes

• No

Which describes the VC dimension for some arbitrary model class?

• # of input dimensions + bias

• maximum # of points that can be shattered by the model.

• min # of dimensions

• how many layers in the model.

"shattering" means...

• to smash and bash data

• to perfectly predict any label configuration of a dataset.

• to find all linear separations.

• minimizing the generalization error.

How many datasets (different point arrangements) do I need to find before I can say that my model shatters p points?

• 1

• 2

• N+1

• p+1

The VC dimension of a perceptron is...

• N+1

• 1

• p

• p+1

The VC dimension of a sinusoid model is...

• 1

• 2

• How am I supposed to even think about this?!

The growth function measures

• how fast my MLP grows

• sample complexity

• degree of polynomials I can fit

• if my VC dimension is positive or negative.

More datapoints will...

• cause the parabolas to become more narrow.

• cause R_emp to move away from R

• cause R_emp to move vertically

• cause R_emp to move closer to R