INTD 161 Pt 3

• AI systems can encode biases if not taken care of.
• Dall-E, an AI image generation model, produces images reflecting stereotypes of successful people as white, male, young, dressed in Western business attire, working in urban offices, and having common hairstyles.

Announcements

• Assignment #2 is live and due Next Tuesday, with three possible submissions.
• Three submissions are allowed, to mitigate accidental submissions, internet outages, etc.
• There will be no feedback provided upon submitting.
• There will be no feedback provided upon submitting.
• Asignment #2 is due tonight
• Asignment #1 grades have been released
• Asignment #3 opens tonight
• Asignment #3 is due the following Tuesday and will be like A2. It will be a multiple choice quiz in Python

Prerequisites and Learning Objectives

• From last lecture, learn the building blocks of machine learning, data and models.
• Data is the information used to train the model, like text and images
• The model is a computer program that processes input and creates output that's usually a number or a collection of numbers.
• Supervised learning learns to imitate, and learns to predict the label of data, which are created by human labelers
• Key learning objectives for this content include:
• Listing the strengths and weaknesses of different models.
• Distinguishing between generative and supervised learning
• Listing examples of generative learning systems and what they are used for

Machine Learning Building Blocks

• Machine learning has 2 building blocks, data and models
• Supervised learning is the basic paradigm
• The goal is to list and describe the building blocks of ML systems
• The goal is to identify problems that can be thought of as supervised learning

AI Caution, Fear, Excitement

• “Success in creating effective AI, could be the biggest event in the history of our civilization. Or the worst. We just don't know. So we cannot know if we will be infinitely helped by AI, or ignored by it and side-lined, or conceivably destroyed by it,” ~Stephen Hawking
• “The rise of AI will free people up to do things that software never will—teaching, caring for patients, and supporting the elderly, for example.” ~Bill Gates
• “AI is far more dangerous than nukes.” ~Elon Musk
• There are three reasons why people fear AI
• Cynicism is the belief that it is rational not to cooperate
• Humanism/racism is systematic bias against machines, denial of their potential moral worth and personhood
• Conservatism is the fear of change, fear of the other tribe
• None of these fears reflect well on those who hold them- Rich Sutton (UoA)

AI Versus ML

• AI mimics intelligence, it includes systems with goals that make decisions.
• ML is a subset of AI where knowledge is extracted from data to build predictive models.
• Expert system MYCIN is given as an example of AI without ML.
• A simulated path-finding robot is another example of AI, not ML.
• Spam filters, YouTube recommendations, Google Lens, and ChatGPT are examples of both AI and ML.
• ML is at the core of most modern AI systems.

Deciding which models to use

• All basic models have different strengths and weaknesses
• Modelling Power is whether the model can learn complex patterns
• Data requirements is the amount of data needed to train the model
• Trainability is whether it is difficult to train, and may require ML Expertise
• Interpretability is understanding what the model learns

Data as a building block

• In ML, data trains models to distill information and knowledge.
• Data is a collection of discrete or continuous values conveying information, quantity, quality, facts, and statistics.

Course Map

• The topics in this deck will include the Machine Learning Building Blocks 2 as well as Generative AI Systems, and choose a model.

Lecture Topics

• This lecture will cover how to choose a model, and another important type of paradigm in MI: Generative Learning
• Generative Learning will be broken down into Imitating Data, generative learning systems (Dall-E), Generative Adversarial Networks, and Sequence Generation Models (ChatGPT)
• There will also be discussion of the power of scaling with data and computation over scaling with people.

Machine Learning Pipeline

• The steps are training and prediction.
• Training extracts info/knowledge from data to build the model.
• Prediction uses the model to answer an input query.
• Labels in ML examples provide the model with the question it should answer.
• The label is the correct answer/output/prediction for a given input.

YouTube Video Recommendation example

• For YouTube video recommendations, input data might include user ID, current viewing video, watch history, liked videos, and disliked videos.
• The output is a list of top-K recommended videos

Google Lens example

• The input is an image like a 2x2 Rubik's cube,
• The output provides links to online stores selling similar items via object detection

MNIST example

• MNIST, a commonly used ML dataset for research, consists of handwritten digits from 0 to 9,
• It's used in real-world applications to read ZIP codes in postal services and check amounts in bank accounts.

Predicting text articles and movie reviews

• Language tasks include classifying news articles by topic, classifying movie reviews as positive or negative.
• There is also text generation based on input text.

Language Data

• Before Large Language Models, ML focused on language, labeled "Natural Language Processing" (NLP).
• Large Language Models are huge Neural Nets with Billions of connections & neurons trained mostly on text data.
• They then output sentences

Bag of Words

• The "Bag of Words" representation converts text to numbers for processing in ML models.
• The goal is to provide calculations
• Word order and importance can matter in language data.
• "Near" can have many word associations

Count-Based Models and Vector Creation:

• Count-based models convert words to numbers based on relationships, building a data table with rows and columns for each word.
• We increment counts based on observed proximity, creating a vector representing word relationships.
• This list of numbers creates a number vector.
• Count-based models can show how word meanings shift over time.

Machine Learning Models

• The model output is a number or a collection of numbers.
• Classification models output a category, or a discrete value.
• Regression models output some value that is a continuous value of the input.
• Models answer questions based on input data

Classification Model

• In classification, the output is a category, or a discrete label.
• Examples include digit recognition (0-9), object detection (cube), and sentiment analysis (positive/negative).

Regression Model

• Regression models work give output with a continuous value of the input, such as scoring digit writing, pricing, or rating a movie.

Model Output

• The output of these models can be a variety of categories and values

Chat GPT

• In ChatGPT data refers to the information used to train the model, like large collections of text from the internet, books, and other written sources

Machine Translation

• Machine translation involves generating text output based on a given input.

Large Language models

• LLMs convert large amounts of text into manageable word vectors and use context windows to predict the next word.

Attention

• LLMs learn to focus on key words through a process, "attention", improving understanding.

Supervised Learning

• Supervised learning, an ML framework, learns from labeled data where each input pairs with a label or correct answer.
• The labels provide descriptive tags or values.
• Supervised learning takes direction from the people selecting the answers

Categories

• Labels in classification are categories, like categories for digits, objects, or sentiment.

Regression

• Labels in regression are numerical values, like scores, prices, or ratings.
• Supervised learning models have no agency and cannot be better than the data they are given.

Imitation

• Supervised learning systems and expert systems are imitation machines.
• ML Models are much easier to update.

Features

• Features, like temperature, precipitation, day of the week, and clothing, are numbers or categorical data inputted into the model.
• Models should generalize predictions on new data.

Generalization

• Generalization makes predictions on new unseen data

Decision Trees

• To get around this, decision trees, a classical machine learning model, can be implemented to find trends in data
• In this model temperature is noted against the data.
• They can be used for both classification and regression.

Decision Trees Pros and Cons

• Decision trees are noted as having high explainability but struggle against higher, more complex data implementation and learning
• It is important to remember how and when the systems work for this, and those of you in INT-D 161 will use powerful ML models

Other Learning

• Supervised learning is the most basic paradigm in ML but there are other learning systems.
• Other learning systems include Generative and Reinforcement Learning.

Lecture Topic Summaries

• Data and models are crucial ML building blocks.
• Data encoding depends on data type.
• Supervised learning imitates human answers.
• Decision trees are explainable classical models.

Sample Basic Models

• Artificial Neural Network: The model is composed of layers of interconnected nodes, or neurons, that process and transmit information, it is comprised of an Input Layer, Hidden Layers, and an Output Layer
• Decision Tree: A decision tree is a structured model that uses a series of binary decisions to classify or predict outcomes based on input features or attributes, it navigates based on conditions like temperature or weather, it then derives certain conditions
• Perceptron: A perceptron is a single-layer neural network that performs binary classification by applying weights to input features and using a threshold to make a decision, it includes inputs, a Heavyside Step-function, and an output