Neural Networks Flashcards
Beschreib Hebbs Rule
Beschreib den Aufbau eines Perceptrons
Wie lautet die Funktion der inner activation?
Wie funktioniert das Aktualisieren der Gewichte?
Wie minimiert man E?
Über das Gradientenabstiegsverfahren
Beschreib die Rule for the change of weights mathematisch
Wie funktioniert der Perceptron Algorithmus?
Wie funktionert lineare Separierbarkeit mir Perceptrons?
Was sind Sigmoid Neurons?
Haben folgende Aktivierungsfunktion
Why not choose all-same input weights for a layer?
all units will have same activations → information loss
Why not choose all-same output weights for a layer?
learning signals will be same → symmetry-breaking problem
Wie funktioniert Error Backpropagation?
Leite die Sigmoid Activation Function ab
Beschreib Batch update
Beschreib online update
Beschreib Minibatch Updates
Was ist der Momentum Term? Was sind die Vorteile davon?
Which of the statements about convolutional neural networks (CNNs) are true?
1. With CNNs, one can put emphasis on certain objects and/or areas in an image to increase classification accuracy.
2. The kernel size has to be chosen greater than the size of the input image.
3. CNNs can capture dependencies between pixels, which distinguishes them from “classic” feed-forward neural networks.
4. The filters used in CNNs are manually designed for each dataset.
1 und 3
Which of the following statements about the Perceptron algorithm are true?
1. At the start of training a perceptron, all weigths are initialized with 0.
2. Before updating the weights, the errors for all datapoints seen so far have to be calculated.
3. The value of the cost function depends both on the weight parameters of the perceptron and the input.
4. The weights are updated by the gradient of the cost function scaled by the learning rate.
3 und 4
Which of the following statements about Receiver Operating Characteristic (ROC) curves is true?
1. ROC curves are used to measure the harmonic mean of precision and recall.
2. ROC curves are used to visualize the relationship between sensitivity and specificity.
3. ROC curves plot a function dependent on the False Positive rate.
4. The AUC (Area Under Curve) can be used as a measure of model performance.
2 und 4
Which of the functions below are commonly used as activation functions in deep neural networks?
1. Hyperbolic Tangent (tanh)
2. Hyperbolic Secant (sech)
3. Softmax
4. Rectified Linear Unit (ReLU)
5. Sigmoid
alles außer 2
One major disadvantage of linear activation functions is that they often suffer from vanishing gradients, i.e. weight adaptation saturates too quickly during the training process. Stimmt das?
Falsch
Beschreib die Aktivierungsfunktion Hyperbolic Tangent (tanh)
Was sind Vorteile von tanh?
- Outputs are zero-centered, which can help with faster convergence in training
- Stronger gradients compared to the sigmoid function
