Here are a list of potential topics for each exam. You should make sure read the textbook, review slides/handouts, and take advantage of the office hours.
Midterm Exam 1
- For a problem, identify
- What type of learning you would apply (supervised vs. unsupervised)?
- What model is appropriate (linear regression, decision trees, k-nearest neighbors, k-means clustering) and why?
- What type of evaluation makes sense (MAE, MSE, accuracy, precision, recall, f1) and why?
- Data
- Shape of input in a problem
- k-fold cross validation
- batch, mini-batch, stochastic gradient descent
- Metrics
- MSE equation
- precision/recall/f1 equation
- precision/recall trade-off
- How to calculate Jaccard Similarity
- KNN
- Describe
- Calculate prediction from data for new point
- k-means
- Describe
- Different initialization schemes
- Procedure for determining cluster membership for new point
- Linear regression
- formula
- interpretation of formula (e.g., the gender pay gap lab)
- trade-offs between closed-form vs. gradient descent optimization
- Decision trees
- Apply a decision tree
- Apply the CART algorithm
- Gini-impurity equation and calculation for a decision boundary
- Recommender Systems
- Pros and cons to different approaches
- Predict ratings for user and item-based collaborative filtering (both simple average and similarity weighted)
Midterm Exam 2
- Problem Sketch - Given a situation, identify
- What type of neural network is appropriate (RNN, CNN) and why?
- What would your input/output look like?
- What loss would you use?
- How would you evaluate your trained system (e.g., precision)?
- Apply CNN kernel, detector, pooling to a concrete example
- Logisitic Regression
- formula
- use cases
- how it relates to decision boundaries
- Describe hyperparameter tuning
- Propose a kernel to do a specific thing
- Attention
- Aims of attention (database metaphor)
- Reason about similarity, normalization, and application of attention weights
- Feedforward Neural Network
- parameters
- hyperparameters
- Reccurrent Neural Network
- parameters
- hyperparameters
- Convolutional Neural Network
- parameters
- hyperparameters
- Neural Networks in general
- preventing overfitting