What is the main difference between a standard (linear) layer in a feedforward neural network and a convolutional layer in a convolutional neural network?
In that context, what does the kernel size of a convolutional layer represent?
a standard layer in a FNN performs a dot product between the input and a weight matrix
a convolution layer in a CNN performs a convolution operation between the input and a set of filter (= kernels).
The kernel size represents the size of the filters sliding over the input, element-wise multiplying the overlapping values and summing them to produce the output
What is the main difference between normal cross validation and nested cross validation?
Name and explain one advantage and one disadvantage of cross-validation over having only a single train, validation and test split?
How do you estimate a standard error for a methods performance using cross-validation?
In CV the model is trained and evaluated using the same set of hyperparameters, k times, each time using a different validation set.
In NCV, an outer loop is used to split the data into training and test set, while an inner loop is used to select the best hyperparameters.
Advantage: more robust / accurate estimate of the models performance on unseen data due to multiple train and test runs
Disadvantage: computationally expensive, since training a model multiple times can be time-consuming, especially with large data sets
Standard error: Record the performances over all k folds -> compute mean and standard deviation (sd) -> standard error = sd divided by square root of k
Given are two methods A and B that predict the same binary classification task (for example, membrane protein or not). Evaluated on the same test set, method A achieves a two-state accuracy (Q2) of 90% and method B has a Q2 of 80%. Given those results, a user might assume that in general method A is better suited for this task. State and explain two different reasons why this assumption might be wrong.
Method A could have been trained on a train set that is more similar to the test set than method B, thus yielding better results
Q2 is a measure of how many correct predictions are made, but ignores the incorrect predictions. If, for example, the test set is very imbalanced and Method A only predicts the majority class, it will have a high Q2 score but never predicts the minority class
Name two performance measurements commonly used in machine learning other than accuracy. Briefly describe what quality of a model those two represent
1) Precision -> ability to avoid false positives
2) Recall -> ability to avoid false negatives
Last changed2 years ago