Give two possible causes for a ML model being unfair.

Part (b): Consider a single neuron/unit with weight vector , bias , and activation function . The input is a vector . (If the math does not display: weights w, bias b, activation function f, input x.) What is the output of the neuron/unit as a function of the parameters and input?

How can attention be used for interpretability?

Part (f): What change(s) would we need to make if we instead wanted to implement Ridge regression as a neural network? (Be specific.)

Mallory selects two different locations on the ROC curve and you help her produce a confusion matrix for each given the test dataset of 100 instances. Below are the two resulting confusion matrices. Tradeoff 1 Tradeoff 2 Actual 1 (+) 0 (-) Predicted 1 (+) 9 2 0 (-) 41 48 Actual 1 (+) 0 (-) Predicted 1 (+) 40 34 0 (-) 10 16

What are deepfakes?

What is the purpose of the positional encoding in a transformer? Would you use positional encoding in an RNN?

Why shouldn’t we use the test set for hyperparameter tuning?

Part (e): What do you conclude about of such a system?

Errors in binary classification: Your classmate Charlie is a UF CS student who is doing undergraduate research at the Super Duper Awesome Deep Transformers Learning Lab (SDADTLL). Researchers at the lab have trained a neural network model (based on a powerful LLM) to predict the validity of any hypothesis formulated as an English sentence. Specifically, given any hypothesis such as: “job applicants who wear red on the day of the interview are 2.37% more likely to be hired than those who wear blue,” or “left-handed individuals are faster at reading nonfiction books than right-handed individuals,” the model outputs TRUE (+1) or FALSE (0). The network has 27651 layers and 7.1 quadrillion parameters and it was trained on 112 Exabytes of data. In this problem, you will help Charlie evaluate the performance of the model on real-world data.