Study on the Pattern Recognition Enhancement for Matrix Factorizations with Automatic Relevance Determination
Machines learning the parts of objects have become more attention in computer science recently, and they have been playing the important role in computer applications such as object recognition, self-driving cars, and image processing, etc… However, the existing research such as traditional non-negative matrix factorization (NMF), principal component analysis (PCA), and vector quantitation (VQ)  has not been discovering the ground-truth bases which are basic components representing objects. For example, in face recognition application, it is supposed that a human face is composed of four basic components: mouth, nose, eyes, and eyebrows that are ground-truth bases to represent a face. If an algorithm could discover correctly four above components, it can represent a face. In contrast, if an algorithm extracts components rather than four, it means that a face is composed by other parts that are not intrinsic features . Indeed, PCA and VQ only discovered a whole face instead of ground-truth bases while traditional NMF discovered basic components that are redundant. In practice, an algorithm fails to extract basic components leading to not recognize correctly objects, not detect motions in video, and camera processing. If it is applied in real time applications: self-driving car, face recognition, it will cause serious issues related to security and safety. Therefore, finding correctly the number of ground truth bases is significant in extracting the hidden structures of investigated data, and improving a performance.
"Study on the Pattern Recognition Enhancement for Matrix Factorizations with Automatic Relevance Determination,"
OSR Journal of Student Research: Vol. 5
, Article 88.
Available at: https://scholarworks.lib.csusb.edu/osr/vol5/iss1/88