CS33: Exam 1 Study Guide

Definitions

You should generally be familiar with the definitions of the terms we've used in class, and be comfortable using them. Here are a few that are especially important:

• Artificial Intelligence Making computers better at tasks that humans are good at.
• AI is made up of many subfields such as robotics, planning, game playing, and machine learning.
• ML has become the dominant approach.
• Machine Learning Any computer program that improves its performance at some task through experience.
• There are three main approaches to machine learning: supervised learning, reinforcement learning, and unsupervised learning.
• Supervised Learning When we train a model using data sets that pair inputs with correct answers (targets).
• Examples we've seen in the class include linear regression, logistic regression, and deep learning.
• Deep Learning A term coined to describe the type of artificial neural network in modern use that has many hidden layers.
• Large Language Model (LLM) A specific type of deep learning model designed to produce text by sequentially generating a plausible-next-word given the context (i.e. the words so far).

Should I use AI for this?

• Be able to reason about real-world problem scenarios to assess possible impacts or dangers
• Be able to reason about what sorts of problems AI models are appropriate for, and how we might go about deciding
• Understand both the strengths and the limitations of LLMs, and why their capabilities might be called "jagged"
• Understand how long the term AI has been in use, and how its use has shifted over that time
• Practice problems:
1. Consider an AI system designed to recognize the faces of students to automatically unlock doors around campus when a student who has permission to enter that door approaches.
• What types of mistakes are possible?
• What are the consequences of each type of mistake?
• What sort of tradeoff between these types of mistakes is reasonable?
• What sort of accuracy rate (if any) would be sufficient to make such a system safe for use?
• What sorts of concerns might such a system present even if it had a high accuracy?
2. Consider the same questions, but for a facial recognition system intended for different environments. For example, airport security cameras, or police body cameras.

Working with Data

• Be able to define 'data'
• Understand the relationship between data and the world it is intended to reflect
• Understand some of the factors that can result in data not accurately reflecting the world
• Know the different types of data that can be stored and how they are categorized
• Understand the process that data goes through before we can apply AI/ML techniques
• Understand what Exploratory Analysis is
• Be able to apply reasoning and analysis to data in order to gain understanding and draw conclusions
• Practice: try working through the Titanic data cleaning example from class, or try doing some Exploratory Analysis on a new data set

Models, Linear and Logistic Regression with Gradient Descent

• Be able to define 'model'
• Be able to explain what makes models useful? What makes models dangerous?
• Understand why we add a column of 1's to the example vectors
• Understand how to interpret the weights of a model once it's been fitted to some data
• Understand the assumptions Linear Regression makes about the data or problem
• Understand the assumptions Logistic Regression makes about the data or problem
• Understand when to use Linear and when to use Logistic regression
• Be able to evaluate a problem and decide whether Linear (or Logistic) Regression is likely to be useful
• Understand why we use Logistic Sigmoid instead of Step-Edge for Logistic Regression
• Understand the difference between Stochastic and Batch gradient descent
• Understand how to handle multi-class problems using Logistic Regression
• Understand how the update equation for SGD is calculated and applied (NOTE: you do not need to be able to do the derivation, but you should have a conceptual understanding of what is being done and why)
• Understand the role of the learning rate (alpha), and the dangers of having it be too big or too small.
• Be able to apply gradient descent by hand. Try working through some practice problems:
1. Consider a sample problem with two examples: ([5,3], 1) and ([-2,4], 3)
2. Assume the initial weights are initialized to 0. Calculate the predictions (should be 0 for both examples), and then calculate the sum squared error.
3. Do one iteration of gradient descent using a learning rate (alpha) of 0.01
4. Re-calculate the predictions and the error using the updated weights
5. For more practice, try doing several more steps.
6. Try doing SGD with one update per example, and compare it to doing batch-updates
7. Try using a different set of data points (i.e. come up with your own examples to practice on)
8. Try doing a similar problem using the Logistic Regression framework; you'll need to ensure the target values are 0 or 1, e.g. ([5,3], 0) and ([-2, 4], 1)

Artificial Neural Networks

• Understand the operation of nodes (simulated neurons) in an ANN
• Describe the important differences between the Step-Edge, Logistic Sigmoid, and Rectified Linear activation functions
• Understand the relationship between ANNs and linear/logistic regression models
• Describe the difference between shallow and deep networks
• Understand the similarities and the differences between ANNs and biological neural networks
• Know the differences between input, output, and hidden nodes
• Be able to decide how many input and output nodes are necessary for a given problem
• Be capable of calculating the number of learnable parameters in a given network
• Understand the similarities and differences between SGD for logistic regression and SGD for an ANN
• Know how the Error Backpropagation Algorithm calculates error, both for output nodes and for hidden nodes (NOTE: you do not need to be able to do the derivation, but you should have a conceptual understanding of what is being done and why).
• Practice problems:
1. Consider an ANN with one hidden layer; it has 4 input nodes, 6 hidden nodes, and 3 output nodes. How many total learnable parameters does it have?
2. Try doing this for several other network architectures
3. Imagine we want to create an ANN to fit a data set with 7 columns, one of which is the target. The target is a class label, which have the values 'cat', 'dog', or 'bird'. How many input nodes would the network have? How many output nodes?