Qs 1) Using kNN - Classifier[Top]

Solution

1. The kNN - Classify was done on a Training Set of 3000 images, and a Test Set of 1000 images. For larger values of training set the percentage drops down significantly.
2. The percentage error is least for when k=4 nearest neighbours are used.
   Part B
   

   Construct the 2-D Isomap model for the first 3000 samples of the MNIST training set using Euclidean and Tangent distance. Show the clusters of the following groups:

   Qs 1) Using Euclidean Distance[Top]

   Solution A : Cluster of 1 & 7

   

   Solution B : Cluster of 4 & 9

   

   Solution C : Cluster of all digits

   
   
   
   
   

   Qs 2) Using Tangent distance[Top]

   Solution A : Cluster of 1 & 7

   

   Solution B : Cluster of 4 & 9

   

   Solution C : Cluster of all digits

   
   
   
   
   Deep Learning
   [Top]

   Method Used	Learning Rate	Epoch	BatchSize	Error Percentage
   Deep Belief Network	1	1	100	8.91
   Deep Belief Network	3	1	100	6.94
   Deep Belief Network	4	25	25	2.62
   Convolution Neural Network	1	10	30	2.14
   Neural Network	5	1	20	2.27

   • In Most Cases Error Percentage decreases with:
     • Increasing the learning rate to a limit.
     • Decrease in Batchsize. This increases computation time but decreases the error.
     • Increase in number of epochs.
   • Convolution Neural Network worked the best.