Exam, neural diagnosis and monitoring

Question 1

what does partition clustering method applied to feature-based data representation provide with?

Answer 1

The partition clustering methods provide a splitting of the feature space into a fixed number of clusters

Question 2

What does the Naive Bayers classifiers assume under the supervised learning paradigm?

Answer 2

Class conditional independence between features

Question 3

What can be used as a performance index regarding the evaluation of the performance of a classifier system?

Answer 3

An estimate of the probability of classification error can be used as a performance index

Question 4

What is visual evoked potentials (VEP)?

Answer 4

• A visual evoked potential is a potential caused by a visual stimulus
• (example an alternating checkerboard pattern on a computer screen)
• Responses are recorded from electrodes that are placed on the back of your head and are observed as a reading on an (EEG)

Question 5

Regarding the classification of VEPs?

Answer 5

A possible data representation that serves as input to the decision module for this problem is a feature vector containing latencies and amplitudes of particular peaks of the VEP signal

Question 6

The miniskowski distance Equation of classification of VEPs?

Answer 6

d=[Number of features corresponding to amplitude]
r=2 vid Euclidean distance
r=1 using Manhattan distance

Question 7

What is K-means?

Answer 7

• A Clustering algorithm
• Uses the least squared Euclidean distance which is the intuitive nearest mean
• partitional
• attempt to minimize the distance between points labeled to be in a cluster and a point designated as the center of the cluster

Question 8

1. How should classification based on FDG-PET images of alzheimer disease vs normal population be made?
2. What is a possible data prespresentation?

3-5. How can the feature space be reduced?

Answer 8

1. Computed aided diagnosis systems for the classification of AD vs N may be based on support vector machines
2. One possible feature-space data representation for the FDG-PET images is using the Voxel intensities as features
3. Feature selection methods can be used to reduce the dimensionality of the feature space
4. Mapping of the feature space into another space of lower dimensionality such as principal component analysis (PCA) is a form of dimensionality reduction
5. Clustering methods can be used as a pre-processing step for the feature selection procedure
6. The lower number of features used, the better classification result

Question 9

How can clustering algorithms be obtained?

Answer 9

Different clustering algorithms can be obtained by adopting different (di)similarity measures between objects or between clusters

Question 10

why is clustering algoritms useful?

Answer 10

For exploratory data analysis, devising structure from unlabelled data

Question 11

what does MAP classifiers require?

Answer 11

the estimation/defenition of a priori class probabilities

Question 12

What does the Naive Bayes classifiers assume?

Answer 12

Class conditional independence between features

Question 13

What is Bayesian classifiers?

Answer 13

A sub-class of statistical classification methods

Question 14

what data is used in semi-supervised learning methods?

Answer 14

Semi-supervised learning methods use both labelled and unlabelled data for training, and labelled data for testing

Question 15

Name one pre-processing step for the classification of VEPs!

Answer 15

One possible pre-processing step consists of averaging over multiple EEG segments synchronized with the visual stimuli onset, in order to recover the VEP signal from the background EEG activity

Question 16

What is K-NN?

Answer 16

• k-nearest neighbors algoritm
• Method used for classification and regression
• k= closest training examples in the feature space
• output: class membership
• a type of instance-based learning
• among the simplest of all machine learning algoritms
• NOT used for clustering

Question 17

what is K-medoids?

Answer 17

A clustering algoritm related to the k-means algoritm

• partitional
• attempt to minimize the distance between points labeled to be in a cluster and a point designated as the center of the cluster
• k-medoids chooses datapoints as centers and uses a generalization of the Manhattan Norm(r=1) to define distance between datapoints instead of the neuclidig distance.

Question 18

what is DBSCAN?

Answer 18

• Density-based spatial clustering of applications with noice
• given a set of points in some space, it groups together points that are closely packed together, marking out outliers in low-density regions
• one of the most common clustering algoritms

Question 19

what is normalized cuts algoritm?

Answer 19

• A spectral clustering technique
• commonly used for image segmentation
• It partitions points into two sets, based on the eigenvector corresponding to the second-smallest eigenvalue of the symmetric normalized Laplacian

Question 20

How is the classification error probability (PE) often estimated?

Answer 20

often estimated experimentally based on training and test sets

Question 21

Name one of the possible indices for performance evaluation of a classification/decision system!

Answer 21

The classification error probability (PE)

Question 22

What happens when the size of the training set increases, when estimating the error probability?

Answer 22

The k-fold cross validation method converges to the true Pe