ENGR 421 • Midterm I • Introduction to Machine Learning

Introduction to Machine Learning

Machine Learning Notation Explained

Machine Learning Preliminaries

Why Supervised?

Hypothesis Space & Occam's Razor

Loss Functions: Measuring Mistakes

Example: Least-Squares Linear Regression

Counting and Probability

Conditional Probability and Independence

Bayes' Rule

Discrete Random Variables

Continuous Random Variables

Expected Value and Variance

Bernoulli and Binomial Distributions

Continuous Uniform Distribution

Exponential Distribution

Normal Distribution

Laplace and Logistic Distributions

Maximum Likelihood Estimation(MLE)

Bernoulli Likelihood

Multinomial Likelihood and Smoothing

Bayes' Theorem

Parametric Classification

Unequal Variances → Quadratic Boundary

Gaussian Classification Boundary

Parametric & Polynomial Regression

Modeling Multivariate Data: Estimation, Normal Distributions, and Naive Bayes

Multivariate Classification: Linear, Quadratic, and Model Selection

Discrete Features & Multivariate Regression

Sample Mean and Covariance Matrix

Mahalanobis Distance

LDA vs QDA Classification

Naive Bayes Classification (Discrete Features)

The Problem

Linear Discriminant

Two Classes/Multiple Classes/Pairwise Seperation

From Discriminants to Posteriors

Gradient Descent

Gradient Descent Update

Linear Regression + MSE: Gradient Descent Step Size Effects

Logistic Discrimination: Two Classes

Logistic Discrimination: K>2 Classes

Generalizing the Linear Model

Discrimination by Regression

Learning to Rank

Perceptron

Training a Perceptron

Limitation: XOR

MLP Architecture & Representation View

Backpropagation

Regression

Discrimination

Should we initialize all MLP weights to zero?

Introduction to Deep Learning & Activation Functions

Training Deep Networks

Regularization Techniques

Tuning Network Structure

Learning Time

Time-Delay Neural Networks (TDNN)

RNN / LSTM / GRU

Generative Adversarial Networks (GANs)

Pass Rates & Majors (Bayes; Law of Total Probability)

Weighted Least Squares (Closed-Form Solution, Matrix View & Interpretation)

MLE for α (positive support, exponential tail)

Linear Discriminant with Equal Variance

MLP with Hard-Threshold Units

One Shared Network vs. Three Separate Networks

Naive Histogram Estimator vs. Parzen Windows (Kernel)

Kernel Smoother

Naive Density Estimator (Bandwidth effect & validity)

Why Not Regression for Classification?

From Binary to Multiclass: One-vs-All / One-vs-One with a Binary Classifier

Max-shift for SoftMax

Why Initialize Weights Near Zero?

Adaptive Learning Rates in Gradient Descent

When Do Direct Input Output Links Help in an MLP?

Mahalanobis vs. Euclidean: Why and When?

Regularized Least Squares

Gaussian Generative Model → Logistic Posterior

Naive Bayes Text Classification with Binary Features

Derivative of Softmax

Model Selection Using Validation Performance and Test MSE

k-NN Decision Boundaries and the Effect of k

Mean Square Error for Linear Regression

Gradient Descent Update

k-NN Regression Prediction

Decision Boundary and Building a Network for Binary Classification

Derivative of Squared Error

True/False Reasoning on Activation, Linear Networks, and Gradient Descent

Computing Total Probability

True/False on Scaling, k-NN, Intrinsic Error and Model Complexity