Questions for self-monitoring

• What's the difference between generative and discriminative models.
• Why is the sigmoid function used in the logistic regression classifier?
• Which role does the bias term play?
• Why should feature values be standardized or normalized? How is this accomplished?
• How can multiple data points be classified in parallel?
• What are the advantages/drawbacks of a logistic regression classifier compared to a naive Bayes classifier?
• What's the difference between the data representations for a binary and a multinomial classifier?
• How can the output be transformed into a probability distribution for the class labels?
• What's the purpose of the loss function?
• How are the optimal weight values determined?
• Will the optimum (the weight assignment that produces the smallest loss) always be found? What could prevent the search from finding the optimum?
• How can the largest gradient of the loss function be determined?
• How are the weight values updated?