BUAD 2001 Midterm Terms

Data

Raw facts and figures

Information

Data presented in a context so that it can answer a question or support decision-making.

Knowledge

Insight derived from experience and expertise (based on data and information)

Primary Data

data that you have collected for your own use

Secondary Data

data somebody else has collected and made available for others to use

Time series data

values of a specific measurements (subject) taken over a range of time periods

Cross Sectional Data

values collected from a number of subjects during a single time period

Panel Data

consists of a time series for each cross-sectional member

Artificial Intelligence

Computer software that can mimic or improve upon functions that would otherwise require human intelligence.

Neural Networks

A statistical techniques that hunt down and expose patterns, building multilayered relationships that humans can’t detect on their own.

Supervised Machine Learning

• Uses data which are labeled by humans (hence supervised)

• Thanks to labeled data, we can measure accuracy of
  supervised machine learning algorithms

Unsupervised Machine Learning

Discover hidden patterns in data without the need for human
intervention (hence unsupervised)

Machine Learning

A machine learning system is trained, rather than explicitly programed.

Deep Learning

A subset of ML that uses neural network algorithms for predictive applications

Generative AI

Algorithms that generate original content, such as images, text, voice, video, or music, by learning patterns from a dataset

Symbolic AI

Refers to methods in AI research that utilize high-level symbolic representations of problems, logic, and search.

Regression

Predict a continuous numerical value

Classification

Predicts discrete categories or labels

Underfitting

• Model is too simple.

• Performs poorly on training data.

• Performs poorly on testing data.

• Associated with high bias

Overfitting

• Model learns noise or random patterns in the training data.

• Performs very well on training data.

• Performs poorly on testing data.

• Associated with high variance.

Test Data

where the trained model is applied to for assessment of performance, and model comparison

Training Data

From which the model/algorithm “learns”

A ____ makes a statement

Decision Tree

Decision Tree classifies things into categories

Classification Tree

When Decision Tree Predicts numerical values

Regression Tree

The very top of the tree

Root Node or The Root

Internal Node / Branch

Any node in a tree structure that has at least one node

The inference of a decision tree runs by routing an example

Root to the leaf

Do all conditions in a decision tree involve only a single feature?

No

KNN Pros

• Classification

• Easy to implement

• Few hyper parameter

KNN Cons

• Can be slow

• Curse of dimensionality

• How to choose k?

• Does not scale very well

• Overfitting

Deep Learning

Advanced subset of machine learning that uses artificial multi-layered
neural networks to simulate the human brain’s decision-making process.

Neural Network

A system made of connected layers of simple neurons that learn patterns from data to make predictions.

Node

A small math unit that takes numbers in, combines them, and passes it forward.

Weight

A number that controls how important an input is to a node.

Bias

A number added to shift node’s input up or down.

Forward Passing

This is the process of the data going from the input layer, to the hidden layers and the output layer.

Loss Function

• Answers how wrong the model is

• Lower Loss = Good

• Higher Loss = Bad

Backpropagation

The process of sending errors backward through the network to update weights and
improve predictions.

Optimizer

The method that updates a model’s weights to reduce the loss and improve predictions.

Activation Function

Decides whether a node should pass its information forward and how strong that signal should be.

Large Language Model

A type of AI that tries to predict the next word/token in a sentence based on what came before it.

Tokenization

The process of breaking down text into smaller units called tokens

Embeddings

A vector (list of numbers) that represents the meaning of the token.

Transformer

A special type of neural network designed to understand context.

Parameters

Broad category of internal variables in a deep learning model system learns automatically from training

Token Embeddings

A high-dimensional space where each token is mapped to a unique position called a vector (a series of number).

Positional Embedding

This preserves the information about the token order and you
don’t lose the relevance of the position of the token in the prompt

Fine Tuning

Retraining an existing LLM on smaller, domain specific dataset.

Instruction Fine Tuning

Teaching the model to follow human instructions better.

Reinforcement Learning from Human Feedback

A crucial innovation for aligning AI systems with human preferences, creating a feedback loop that bridges the gap between what models can do and what humans actually want them to do.

Direct Prefrence Optimization

Streamlines the preference learning process by directly training the model to produce human-preferred responses without requiring the complex intermediate reward-modeling step that RLHF demands.

Prompt Engineering

The skill of crafting clear text inputs (prompts) to guide GenAI models such as LLMs
toward specific outputs.

Zero Shot Prompting

Relies entirely on the model’s pre-trained knowledge to interpret and respond. It works well for straightforward tasks where the model can reasonably infer the requirements without additional context.

Few Shot Prompting

• Help establish a pattern, guiding the model to interpret the task requirements with more precision.

• Effective when the desired output format needs to be demonstrated.

Role Playing

Assigning the AI a specific role or persona enabling it to engage with tasks in
a realistic and contextually appropriate manner.

Chain Of Thought Prompting

• This technique is designed to guide an LLM through a
  multi-step reasoning process.

• Useful for tasks that require logical inference,
  detailed reasoning, or decision based on multiple
  variables.

Guided Zero Shot COT Prompting

• Prompting the model with a structured sequence of steps without specific examples.

• This technique is useful for tasks that require a precise sequence but may not need specific examples to explain the pattern.

Open Ended Zero Shot COT Prompting

This technique uses minimal instruction to prompt the model to think through the problem independently, without any structured steps or examples.

Few Shot COT Prompting

A technique where you provide a small number of example problems, each with step‑by‑step reasoning, to guide an LLM to produce its own multi‑step reasoning on new queries

What to include for custom instructions

• Who you are

• What good looks like

• How you want it to push back

Custom Instructions

Provide standing preferences and information that guide how an AI
assistant responds

Temperature

Controls randomness

Top P

Filters candidate tokens using a cumulative probability threshold.

Max Length

Limits response length

Frequency Penalty

Reduces repeated words/phrases

Presence Penalty

Discourages repeated ideas.