ImageClassification_RadialBasisFunctions_SelfOrganizingMaps
Image classification on MNIST Hand-written Digits Dataset and function aproximation (regression) via full custom Radial Basis Functions (Exact Interpolation + Fixed Centers Selected at Random Method + Regularization) and Self Organizing Maps implementation.
For usage instructions please check Usage README
For full documentation - system design, experiments & findings please read ImageClassification_RadialBasisFunctions_SelfOrganizingMapsDoc Report
Introduction
Few experiments are undertaken:
Q1. Function Approximation with RBFN
๐ฆ = 1.2sin(๐๐ฅ)โcos(2.4๐๐ฅ), ๐๐๐ ๐ฅ โ [โ1.6,1.6]
(a) Exact interpolation method (Gaussian, standard deviation of 0.1).
(b) Fixed Centers Selected at Random
(c) Regularization study
Q2. Handwritten Digits Classification using RBFN
(a) Exact Interpolation Method + regularization (Gaussian function, standard deviation of 100)
(b) Gaussian function with standard deviation of 100 + varying width
(c) Xlassical โK-Mean Clusteringโ with 2 centers.
Q3. Self-Organizing Map (SOM)
(a) Map a 1-dimensional output layer of 40 neurons to a โhatโ (sinc function) + visualisation.
(b) Maps a 2-dimensional output layer of 64 (i.e. 8ร8) neurons to a โcircleโ + visualisation.
(c) Cluster and classifiy handwritten digits.
+ Visualisation of conceptual/semantic map of the trained SOM and the trained weights of each output neuron on a 10ร10 map.
Figures:
Radial Basis Functions
Self Organizing Maps