Reference no: EM133915964

Instruction

The MNIST database is a dataset with handwritten digits (from 0 to 9). The digits have been size-normalised and centred in a fixed-size image (28 × 28 pixels) with values from 0 to 1. You can use the following code with TensorFlow in Python to download the data.

Every MNIST data point has two parts: an image of a handwritten digit and a corresponding label. We will call the images x and the labels y. Both the training set and test set contain x and y.

Each image is 28 pixels by 28 pixels.

As mentioned, the corresponding labels in the MNIST are numbers between 0 and 9, describing which digit a given image represents. In this assessment, we regard the labels as one-hot vectors; that is, 0 in most dimensions, and 1 in a single dimension. In this case, the ??-th digit will be represented as a vector, which is 1 in the? dimensions. For example, 3 would be [0,0,0,1,0,0,0,0,0,0].

The assessment aims to build NNs for classifying handwritten digits in the MNIST database, train it on the training set and test it on the test set. Since the main object of this assessment is for you to understand the relationship between input, model and output, you are not expected to achieve very high accuracy in model performance; instead, for each task, you should be able to identify how you can improve model performance with the change of network structure.

There are two parts to this assessment.
Part 1
Part 1 is comprised of three main tasks.

Task 1
Build a neural network without convolutional layers to do the classification task (hint: you will need the use of dense layers). Then you can change the model structure (i.e. number of dense layers, number of neurons in dense layers or activation functions) to be able to improve network performance.

Task 2
Build a neural network with the use of convolutional layers (you can decide other layer types you want to include in your network). Then you can change the number of convolutional layers and the number of filters or activation functions in the convolutional layers to be able to improve network performance. Get top-rated assignment help now.

Task 3
Change the type of optimiser or learning rate that you applied in the previous tasks and see how these changes can influence model performance. (You can keep the final network structure you applied in task 2 and try at least one different optimiser setting.)
Please read the following comments and requirements very carefully before starting the assessment:

The assessment is based on the content of labs and Weeks 1-3.

In Week 1 we talked about the use of training set, validation set and test set in machine learning. In this assessment, you are asked to train the NN on the training set and test the NN on the test set, without any given validation set. (If you want to monitor the training process, you can also try what we did in Week 3: you can consider the validation set is the same as the test set in this assessment.)

In the assessment, the performance of an NN is measured by its prediction accuracy in classifying images from the test set (i.e. number of the correctly predicted images/number of the images in the test set).

Since the MNIST dataset is a black-and-white image dataset, the shape of dataset is (dataset_length, 28,28). But to fit it into a conv2d layer, we need to make the input shape comply with its required format: (batch_size, image_width, image_depth, image_channels). Although batch_size can be decided later when you train it, you will still need to tell the number of image channels here. You can consider reshaping the dataset into (dataset_length, 28,28,1) or add one more dimension at the end with np.newaxis.

You are expected to show at least two models in for tasks 1 and 2: one for the model you start with, and another model is the model that you identified to have better accuracy. For task 3, you need to show what optimiser and/or learning rate you applied.

Part 2
Your report must at least contain the following content:
Your name and student number.
Architectures of the NNs, with figures for tasks 1 and 2.
Description on the optimiser and learning rate you applied in the final model of task 2 and the optimiser or change of learning rate you used in task 3.
Experiments and performances, with parameter setting.
Discussion on the improvement/deterioration of the NN's performance after changing the architecture and parameter setting for each task and findings of comparing the results from all three tasks.
The ranking of all NNs' performances from all the three tasks.

Assessment criteria
This assessment will measure your ability to:

Part 1:
describe the two models, experiment settings and compare the results for task 1
describe the two models, experiment settings and compare the results for task 2
describe the two optimisers or learning rates, experiment settings and compare the results for task 3

Part 2:
demonstrate correct code quality
research extensively and demonstrate depth of thinking; produce a well-structured report.