Project Title
Dimensionality Reduction Techniques: Algorithm Analysis and Investigation

Project Goal
Dimensionality Reduction and Manifold Learning are critical concerns for many of the subfields of Artificial Intelligence, from applied visual pattern recognition to knowledge-base data compression.  As such, there has been much attention to improving the classical algorithms used for minimising the dimensionality of a data set.  Many of these algorithms work impresively well on complicated-seeming examples presented in seminal papers and some seem to be designed for special situations which demand algorithmic modifications-- we would like to be able to test these algorithms on a neutral data set including realistic distributions and determine their applied comparative usefulness as well as attempting, if possible, to combine the greatest strengths of the algorithmic developments thus far in the field, knowledge of when exactly certain techniques are most useful, and algorithmic techniques from related statistical analysis subfields in order to create a strong hybrid algorithm or toolbox.  Also under consideration are any applications such a hybrid technique might motivate in some subfields of Artificial Intelligence (of specific concern are agent-based knowledge systems).

To specify the goal-concerns of this project,:

• Research and implement (in Matlab) classical algorithms (PCA, MDS, Isomap, LLE)
  
• Investigate and implement, where useful, recent modifications of classical algorithms
• Compare implemented methods with other related statistical analysis fields and techniques
• Create neutral test suite
• Compare and analyse behaviours of implemented algorithms
• Conjecture and test hybrid algorithms and techiniques
  

Project Timeline

Project Results & Conclusions
Thesis