Nearest Neighbour Classifier, Cont'd
||Nearest Neighbour Classifier, Cont'd|
||Natan Kruglyak <email@example.com>|
||2021-11-30 – 2022-07-01|
This is an extension of the ongoing project LiU-compute-2021-41. The computations conducted in that project have been fairly successful and have resulted in a research paper submitted for publication. They have also revealed the need for "cleaning" of the training set for the constructed neural network as the next step in improving the suggested algorithm. However, investigating and validating the new algorithm would require a higher amount of computing hours per month.
A windowed version of the Nearest Neighbour (NN) classifier for images is investigated. While its construction is inspired by the architecture of artificial neural networks, the underlying theoretical framework comes from approximation theory. The well known MNIST dataset of handwritten digits, and its recent extension EMNIST, are used to investigate the described concepts. In combination with extensions through shifts, rotations, and non-uniform scalings of the available training images, this novel classifier is expected to outperform previously published NN algorithms on these datasets.