Pattern Recognition Toolkit

Vector classification is a process by which a feature vector is classified or labelled by partitioning the multidimensional space into regions corresponding to the different populations. Ideally, the regions should be such that the decisions are never wrong. Since this is not always possible, one should try to minimise the probability of error.

Typically the classification is determined by using information encoded in probability distributions or more commonly in a set of training samples. The training samples can be labelled in which case the classification process is supervised. When no labels are provided, the classification process is unsupervised and is referred to as clustering where the distributions are automatically determined.

Pattern Recognition Toolkit classes

The pattern recognition toolkit contains a list of class hierarchies which should prove to be useful in pattern recognition application domains as well as for general numerical data processing.

All the classes have been designed in an integrated fashion to improve their flexibility, ease of use and intuitiveness. This means that the following things hold true throughout the toolkit:

• With the exception of the data storage objects, all the classes that will be used by the user are AMMA Big Objects and naturally have corresponding behaviour.
• The names of handle classes which the user will use terminate with a C (eg NumModelPolynomialC). The underlying implementation class names terminate with a BC (eg NumModelPolynomialBC). These classes should not be used directly.

Toolkit Hierarchy

Several class hierarchies are contained within the toolkit. The following list provides access to a description of each of the hierarchies including its purpose and access to subsection pages:

• NumData Data storage types for training and test sets etc.
• NumFunc Class hierarchy sharing basic Y=f(X) functionality where Y and X are vectors. Different subhierarchies include:
  • NumMagnitude Measures |X| using various metrics.
  • NumModel Is able to learn the relationship Y=f(X) from a training set of examples.
  • NumCost Classes for representing cost functions for optimisation. Can use anything derived from NumFuncC for optimising the expression |Yd-f(X)| where Yd is a desired output and || is estimated using any of the NumMagnitude classes.
  • NumPDF Represents probability density functions for use with classifiers.
  • NumClassify Supervised and unsupervised classification algorithms.
  • NumReduce Feature space dimensionality reduction classes.
• NumOptimise Optimisation classes to use in conjunction with cost functions derived from NumCostC.

IO and the Pattern Recognition Toolkit

• The old situation with the Pattern Recognition Toolkit used to be that:
  • All objects can be saved to a stream using either a Save function call or the output stream operator. They can be read back in using the input stream operator:

    NumClassifyKMeansC classifier(10, NumMagnitudeSqrEuclideanC());
    ofstream out("MyKmeansFile");
    // this ...
    classifier.Save(out);
    // ...is equivalent to this:
    out << classifier;
    
    ifstream in("MyKmeansFile");
    NumClassifyKmeansC othclassifier(3, NumMagnitudeSqrEuclideanC());
    in >> othclassifiers; // overrides parametres given in constructor (3<--10)
    

    Virtual constructors are used so that an object of base class type can be constructed from a stream as any of its derived types:

    ifstream in ("MyKmeansFile");
    NumClassifyC generic;
    in >> generic;
    

    However only the NumClassifyC interface would be available.

  • Output and input stream operators are there for all classes. You don't see them documented nor in the header file of the class you are using because they are implemented via a macro in the header file of the class NumLabelC, like this:

    #define NUM_IO_OPERATORS(DataC) \
    ostream & operator<< (ostream &out, const DataC &val); \
    istream & operator>> (istream &in, DataC &val);
    

• The current IO situation should be for all classes in AMMA to work in this way. Please try it first, and if it doesn't work, please complain.

Subtopics: