True/False

Question 1

The mean-field theory for Hopfield network yields the exact value for the critical storage capacity.

Answer 1

False

Question 2

That the energy cannot increase under the deterministic Hopfield dynamics is a consequence of the fact that the weights are symmetric.

Answer 2

True

Question 3

The stochastic update rule for the Hopfield network is different from the Metropolis algorithm.

Answer 3

True

Question 4

All stored patterns are local minima of the energy function.

Answer 4

False

Question 5

The detailed balance condition is a necessary condition for the Markov-Chan Monte-Carlo algorithm to converge.

Answer 5

False

Question 6

That the energy cannot increase under the deterministic Hopfield dynamics is valid only if the thresholds are put to zero.

Answer 6

False

Question 7

For a given 𝛼, the one-step error probability for the deterministic Hopfield network is lower when the diagonal weights are set to zero.

Answer 7

False

Question 8

In the limit of N → ∞ the order parameter m𝜇 can have more than one component of order unity, the other components are small.

Answer 8

True

Question 9

The stochastic update rule for the Hopfield network is identical to the Metropolis algorithm.

Answer 9

False

Question 10

The detailed balance condition is a necessary condition for the Markov-Chain Monte-Carlo algorithm to converge.

Answer 10

False

Question 11

That the energy cannot increase under the deterministic Hopfield dynamics is a consequence of the fact that the weights are symmetric.

Answer 11

True

Question 12

The mean-field theory for the Hopfield network yields the exact value for the critical storage capacity.

Answer 12

False

Question 13

All stored patterns are local minima of the energy function.

Answer 13

False

Question 14

Not all local minima of the energy function of the Hopfield network correspond to stored patterns.

Answer 14

True

Question 15

The stochastic update rule of the Hopfield network is different from the Metropolis algorithm.

Answer 15

True

Question 16

That the energy cannot increase under the deterministic Hopfield dynamics is a consequence of the fact that the diagonal weights are set to zero.

Answer 16

False

Question 17

That the energy cannot increase under the deterministic Hopfield dynamics holds also when the thresholds are zero.

Answer 17

True

Question 18

The detailed condition is a necessary condition for the Markov-Chain Monte-Carlo algorithm to converge.

Answer 18

False

Question 19

A perceptron that solves the parity problem with N inputs contains at least N^2 hidden neurons.

Answer 19

False

Question 20

Increasing the number of hidden neurons in the network increases the risk of overfitting.

Answer 20

True

Question 21

Two hidden layers are necessary to approximate any real valued-function with N inputs and one output in terms of a perceptron.

Answer 21

False

Question 22

Using stochastic gradient decent in backpropagation assures that the energy either decreases or stays constant.

Answer 22

False

Question 23

In minimisation with a Lagrange multiplier, the function multiplying the Lagrange multiplier can also assume negative values.

Answer 23

False

Question 24

Some of the functions with 5 Boolean valued inputs and one Boolean valued output are linearly separable.

Answer 24

True

Question 25

Different layers of a deep network learn at different speeds because their effects on the output are different.

Answer 25

True

Question 26

The weights in a perceptron are symmetric.

Answer 26

False

Question 27

L1-regularisation reduces small weights more than L2-regularisation.

Answer 27

True

Question 28

Weight decay helps against overfitting.

Answer 28

True

Question 29

Increasing the number of hidden neurons in the network increases the risk of overfitting.

Answer 29

True

Question 30

Two hidden layers are necessary to approximate any real valued-function with N inputs and one output in terms of a perceptron.

Answer 30

False

Question 31

Pruning increases the risk of overfitting.

Answer 31

False

Question 32

Using stochastic gradient decent in backpropagation assures that the energy either decreases or stays constant.

Answer 32

False

Question 33

In minimisation with Lagrange multiplier, the function the Lagrange multiplier must be equal to or larger than zero.

Answer 33

True

Question 34

Back-propagation is a form of unsupervised learning.

Answer 34

False

Question 35

To make use of back-propagation, it is necessary to know how the target outputs of input patterns in the training set.

Answer 35

True

Question 36

"Early stopping" in back-propagation helps to avoid being stuck in local minima of energy.

Answer 36

False

Question 37

"Early stopping" in back-propagation is a way to avoid overfitting.

Answer 37

True

Question 38

Using stochastic path through weight space in back-propagation helps to avoid being stuck in local minima of energy.

Answer 38

True

Question 39

Using stochastic path through weight space in back-propagation prevents overfitting.

Answer 39

False

Question 40

Using stochastic path through weight space in back-propagation assures that the energy either decreases or stays constant.

Answer 40

False

Question 41

There are 2^(2^n) functions with n Boolean valued inputs and one Boolean valued output.

Answer 41

True

Question 42

None of the functions with 5 Boolean valued inputs and one Boolean valued output are linearly separable.

Answer 42

False

Question 43

There are precisely 24 functions with 3 Boolean valued inputs and one Boolean valued output (equal to zero ore one) where exactly three of the possible inputs maps to zero.

Answer 43

False

Question 44

Oja's learning is a form of unsupervised learning.

Answer 44

False

Question 45

Then dimension of the output space of a Kohonen network must be equal to the dimension of the input space.

Answer 45

True

Question 46

The number of neurons in the input layer of a perceptron is equal to the number of input patterns.

Answer 46

True

Question 47

You need access to the state of all neurons in a multilayer perceptron when updating all weights through backpropagation.

Answer 47

True

Question 48

Consider the Hopfield network. If a pattern is stable it must be an eigenvector of the weight matrix.

Answer 48

False

Question 49

If you store two orthogonal patterns in a Hopfield network, they must always turn out unstable.

Answer 49

False

Question 50

Kohonen algorithm learns convex distributions better than concave ones.

Answer 50

True

Question 51

The number of N-dimensional Boolean functions is 2^N.

Answer 51

False. it is (the number of choices)^(the number of inputs)

Question 52

The weight matrices in a perceptron are symmetric.

Answer 52

False

Question 53

Using g(b)=b as activation function and putting all thresholds to zero in a multilayer perceptron allows you to solve some linearly inseparable problems.

Answer 53

False

Question 54

You need at least four radial basis functions for the XOR-problem to be linearly separable in the space spanned by the radial spaces.

Answer 54

False

Question 55

Consider p>2 patterns uniformly distributed on a circle. None of the eigenvalues of the covariance matrix of the patterns is zero.

Answer 55

True

Question 56

Assume that the weight vector in Oja's rule corresponds to a stable steady state after a given iteration. The weight vector may change in the next iteration.

Answer 56

True

Question 57

If your Kohonen network is supposed to learn the distribution P(xi), it is important to generate the patterns xi^(my) before you start training the network.

Answer 57

False

Question 58

All one-dimensional Boolean problems are linearly separable.

Answer 58

True

Question 59

In Kohonen's algorithm. the neurons have fixed positions in the output space.

Answer 59

True

Question 60

Some elements of the covariance matrix are variances.

Answer 60

True