1.1. Basic architecture [1 pt]
Implement the CNN architecture shown above in Figure 1. Use ReLU activation functions for all layers except the final layer, which should use the Softmax activation function. Use the Adam optimiser and SparseCategoricalCrossentropy loss. Train this on the yoga32 dataset – what do you observe?
1.2. Regularisation and data augmentation [2 pt]
Modify the basic architecture by adding some form of (a) regularisation and (b) data augmentation. Train your new network on the yoga32 dataset – how does the training performance change?
Your write-up should include a brief description and justification of your choice of regularisation and data augmentation schemes. It should also show the plots of training and validation accuracy for the original network (without regularisation+data augmentation) and the network with these modifications and explain any differences that you observe in the training behaviour.
2. Error analysis [2 pt]
Evaluate your network from part 1.2 on the yoga32 test set. In your write-up, present the overall classification accuracy and the average accuracy for each of the 10 classes. Explain the performance of the CNN model, using example images from the test set to illustrate your discussion. What classes/images were difficult for this model, and why?
3. Visualisation [2 pt]
Visualise the feature space that your network uses to classify images by implementing a nearest neighbour analysis. Use the embeddding from the last convolutional layer of your network from part 1.2 after it has been maxpooled (e.g., extract this layer at the point at which it is flattened and sent to the classification layer). To visualise how images are organised in this feature space, implement a nearest neighbour analysis. For each test image, find the 5 nearest neighbours in the training set. Use Euclidean distance to compare the feature vector from the test image to the feature vectors of the training images. In your write-up, show nearest neighbours for multiple test images to illustrate the feature space and explain your model’s performance. Critically evaluate your model – has it learned a good feature space for this classification task?
