You are on page 1of 4

EE 142 Project: MNIST Competition Akila Ganlath

Introduction
I attempted the competition using Convolution Neural Network’s (CNN’s) from Matlab
Deep Learning Toolbox (DLT). A 3−d enumeration was performed over the number of
parameters, number of training epochs, and the number of training examples to capture
the sensitivity of classification accuracy to these parameters.

Approach
A deep network from the DLT can be defined as a vector collection of layer objects, such
as imageInputLayer, convolution2dLayer, batchNormalizationLayer, reluLayer,
maxpooling2dLayer, fullyConnectedLayer, softmaxLayer, classficationLayer.
Of these layers, only convolution2dLayer, batchNormalizationLayer, and
fullyConnectedLayer have learnable parameters. Each of the convolution2dLayer
have (h×w×c+1)×Nf ilters parameters, where h, w are the height and width of each filter,
respectively, and Nf ilters is the number of filters in the layer. The batchNormalizationLayer
has two vector parameters γ and β the scale factor and offset. The fullyConnectedLayer
has matrix of parameters W vector of parameters b.
To capture the performance of the network on the number of parameters, 5 different
network configurations are defined. These can be found in
layer_details.m
were made with different numbers of hidden layers. Each of these network configurations
has a different number of learnable parameters.
The accuracy of a trained network depends on how many epochs the training was
conducted, how many data samples were used in training, and the number of parame-
ters(network topology). Although there are many other hyper-parameters that determine
the classification accuracy, we restrict to variation in these three by enumerating 5 changes
in each and train 5 × 5 × 5 different networks, exploring a discretized cubic search space.
A script was prepared to conduct the experiments and plot the results:
Project_Script.m
Each experiment conducted corresponds to a test data classification accuracy of each
element of the cubic.

Results
The test classification accuracy of all experiments was visualized using an intensity map
as a slice in the ’number of training samples’ dimension, as seen in the figure below,
where the accuracy is captured by the intensity colors where the accuracy range is given
by the adjacent colorbars.

1
We can see the interdependence of the three parameters in these slices and the non-
convex(non-concave) accuracy function with respect to the ’network topology’ and
’number of epochs’ variations.
We choose the leftmost figure, that corresponding to the best performing network
topology’ and ’number of training samples’ slice to visualize how the accuracy de-
pends on ’number of epochs’ and plot the results below. We follow the same process
for the other corresponding slices

2
All experiments run were completed in 3124.69 seconds.

3
References
[1] Mathworks, Create Simple Deep Learning Network for Classification

You might also like