In this exercise you should aim to improve the recognition rate on the MNIST dataset using SVM. Discuss a good architecture for your framework so that you can reuse software components in later exercises. Reminder: As already discussed. From now on you are free to either implement algorithms on your own or use any kinds of libraries. Use the provided training set to build your SVM. Apply the trained SVM to classify the test set. Investigate at least two different kernels and optimize the SVM parameters by means of cross-validation.
- Access to your github so that we can inspect your code.
- Average accuracy during cross-validation for all investigated kernels (e.g. linear and RBF) and all parameter values (e.g. C and γ).
- Accuracy on the test set with the optimized parameter values.
Use the provided training set to train an MLP with one hidden layer. Apply the trained MLP to classify the test set. Perform cross-validation with the following parameters:
- Optimize number of neurons in the hidden layer (typically in the range [10, 100]).
- Optimize learning rate c (typically in the range [0.001, 0.1]).
- Optimize number of training iterations. Plot a graph showing the error on the training set and the validation set, respectively, with respect to the training epochs.
- Perform the random initialization several times and choose the best network during cross- validation.
- Access to your github so that we can inspect your code.
- Plot showing the error-rate on the training and the validation set with respect to the training epochs.
- Test accuracy with the best parameters found during cross-validation.
In this exercise, you should train and test a basic CNN on the MNIST dataset. We recommend you use the DeepDIVA framework, which is build on top of PyTorch. Complete the provided CNN implementation. Use the provided training set to train a CNN. Apply the trained CNN to classify the test set. Perform validation: • Optimize learning rate (typically in the range [0.001, 0.1]). • Optimize number of training iterations. Plot a graph showing the accuracy on the training set and the validation set, respectively, with respect to the training epochs. • Perform the random initialization several times and choose the best network during validation.
Use the provided training set (permutated MNIST) to train your MLP (from task 2b) and your CNN (from task 2c). Apply the both trained models to classify the test set (permutated MNIST). Compare the results with the results from the normal MNIST. Is there a difference from before? If yes, why? Try to explain what you observe with your own words and the reason of this happening.