DeepLearning_GPT3_questions

Softmax

ReLU

Tanh

Sigmoid

To introduce non-linearity into the network

To shift the activation function to the left or right

To reduce the risk of overfitting

To ensure that the output is always positive

Data augmentation

Dropout

All of the above

Early stopping

Faster convergence to a good solution

Reduction of overfitting

Improved generalization to new data

All of the above

Naive Bayes

RMSProp

Adam

Stochastic Gradient Descent (SGD)

The sum of the absolute differences between predicted and target values.

The negative sum of the dot product between weights and inputs for all misclassified examples.

The entropy of the predicted probabilities compared to the true labels.

The mean squared error between predicted and target values.

The squared difference between the predicted output and the target output of a perceptron.

The number of misclassified examples by a perceptron.

The number of iterations required for a perceptron to converge.

The average of the distances between the decision boundary and the training examples.

They require less training data than other types of neural networks

They can handle variable-sized inputs

They can learn spatial hierarchies of features

They are more interpretable than other types of neural networks