Description
For the first part, answer the following questions in a Jupyter notebook. Label each question with a heading with the answer in a cell below it.
1. Explain the differences between artificial intelligence, machine learning, and deep learning.
2. Explain the difference between supervised learning and unsupervised learning. Can you have a combination of both?
3. Suppose you train a classification model with a SVM classifier. After fine-tuning, you find the accuracy of the test set is 85%. You decide to adjust one of the hyperparameters to get the test set accuracy above 90%. It works but then you think it is possible to adjust another hyperparameter to get the test set accuracy even higher. You keep adjusting hyperparameters in this way. At what point should you stop adjusting hyperparameters to improve the test set accuracy? Explain your answer.
4. Suppose we have a deep neural network for predicting the blood glucose value (quantitative) given 50 features for 10,000 patents. You have six hidden layers each with a relu activation function. The number of neurons for the hidden layers are 100, 100, 100, 50, 50, 20. What would be the code for the output layer in keras?
For the second part, there is a dataset called breast_cancer (sklearn.datasets.load_breast_cancer()). Load the data as you did the diabetes dataset:
X, y = load_breast_cancer (Links to an external site.)(return_X_y=True)
Train a classifier model (any one you want but not a neural network). Report the accuracy of the training data and the test data. Find and discuss the precision and recall for the training data (you decide the threshold). For the third part, use a convolution neural network to fit the MNIST data (the handwriting data, not the fashion data). Use the test data we have been using in the past as the validation set (we will not have a test set here). You decide the number of convolution layers, pooling layers, and dense layers. Report your validation accuracy after you train the model.
