Notes 20110208 CIS 6050 Neural Networks
From SnOwy - Ed's Wiki Notebook
Speaker: Dr. David Calvert
Contents |
Literature Review
- Due 18th -- find ≥ 3 papers to review for uses of ANNs
Techniques Overview
- Sample data ...
- Records = { 1, 2, 3 ... n }
- Feature1 = { a11, a12, a13 ... a1n }
- Feature2 = { a21, a22, a23 ... a2n }
- Featurem = { am1, am2, am3 ... amn }
- n = number of samples
- m = number of features (measured per sample)
- can be discrete or part of a series
- dimensionality of the data is m
Data Reduction
- possible to have too many features
- network learns the wrong thing
- network hones in on an incorrect pattern
- unnecessary, meaningless complexity (noise)
- transformations of other inputs
- example: linear transformations -- F1; F2 = F1+1; F3 = F1/2;
- example: random values -- F1; F2 = random();
- how do we know if values are actually needed in a problem?
- use some data reduction to remove some columns from the dataset
- mostly for columns which are transformations of other columns
- principle component analysis (PCA)
- look for columns that are orthogonal to one another
- looks for columns that are the most different
- discards columns that are linear combinations
- large variances may be interesting -- these are kept
Clustering
- grouping data based on similarity
- usually involves a distance between groups
- usually involves a centre for each group
- the idea -- finding the groups provides us with some knowledge of the data
- later elements can be compared to clusters to identify for which they are most similar
- can label the clusters to provide meaning
- providing a label is classification (providing meaning = classifying)
- only if cluster actually represents something
- clustering requires at least one parameter
- minimum distance
- density
- number of clusters
- clusters are seldom separable -- often overlap
- classification
- labelling data
- build a model which labels records
- once trained, the system can be find a label for previously unseen data
- normally requires some labelled data which is used to build the model
- a kind of classification where we attack labels to the classifications
Prediction
sequential or temporal
- normally with a sequence, we want to know what comes next
- sometimes more interested in modelling characteristics of the sequence without prediction
- identifies dynamics of the system
- example -- heart rhythm -- does an ECG trace period have a meaning for cardiac health?