Project Part 3: Classification Using Neural Networks and Deep Learning

Description

Project Part 3: Classification Using
Neural Networks and Deep Learning
Objective:
The main objective of this project is learning and understanding the concept of convolutional
neural network in an experiment with different parameters and settings.
Dataset:
The dataset is the same as that of Project 1 – MNIST dataset and can be easily brought into the
workspace using libraries in Keras.
Tasks:
The major 3 tasks to be implemented are:
1. Run the baseline code.
2. Change the kernel size to 5 X 5 and redo the experiment again and plot the learning errors vs
epoch and report the testing error and testing accuracy.
3. Change the number of feature maps in the first and second convolutional layers and repeat
the experiment. Plot the learning errors vs epoch and report the testing error and testing
accuracy.
Results:
The plot of learning error vs epoch and accuracy are plotted at each stage of the experiment
conducted.
Task 1:
Kernel size = 3 X 3. The test loss and accuracy of the experiment with the given (default)
parameters are:

Task 2:
Change the kernel size to 5X5 and redo the experiment. The test loss and accuracy of the
experiment with the changed parameters are:
Because the Project Overview said: The pooling is 2×2 with stride 1
So, I add the strides= (1,1) and padding=” same”, I made a comparison between adding
and not adding, and finally the accuracy of adding strides is higher.
Result without strides:
The accuracy is 98.2%,when the kernel size increased.
With strides:

The accuracy seems to increase to 98.9% when the kernel size is increased and add strides.
Task 3:
Change the number of feature maps in the first and second convolutional layers and redo the
experiment. [ # feature maps for 1st layer – 8, 2nd layer – 32]
I also add strides to increase the accuracy and do the compare.
Result without strides:

The accuracy seems to increase to 98.3% when the feature maps are increased in the first and
second layers.
Result with strides:

The accuracy seems to increase to 98.7% when the feature maps are increased in the first and
second layers and add strides.
Conclusion:
The overall observation is that the model performs better when increasing the kernel size and
the number of feature maps and is evident in the above-mentioned obtained test accuracies.
And add strides also can increase the accuracy obviously.