MIS3536 Final Exam

The number of parameters Size of context window Training size Accuracy Inference speed Memory usage

How are LLMs compared and talked about? (Name three ways)

Described in terms of an "Agent", "Environment", "Reward", and "Policy"

How is reinforcement learning described (terminology, 4 words)

Decision tree

The simplest classification model

Feature irrelevance

This model also assumes that no other features of the problem are influencing the outcome

Learning, reasoning, and self-correction

What 3 aspects does modern AI focus on?

Pre-design Design and development stage Deployment stage

What are the stages of the AI life cycle

1. Data quality 2. Generation method 3. Input context

What are three common causes for LLMs to hallucinate

Transformer model Transformer architecture

What can the modern explosion in LLM functionality be traced to

He believes that people are underestimating how radical the upside of AI could be just how people are underestimating how bad the risks could be

What does Dario Amodei believe about the view on AI?

Local Interpretable Model-Agnostic Explanations

What does LIME stand for?

Creating LLMs

What does Mistral AI focus on

Train workers to adapt to technical roles and provide better machinery to take on the physically demanding tasks

What is Osomuglu's solution to the economic challenge?

Institutional or historical bias

What is systemic bias also referred to as?

Robotic Process Automations focus on automating repetitive, rule-based tasks, such as data entry. It operates based on predefined instructions without "learning" AI involves systems that simulate human intelligence, capable of learning, reasoning, and adapting. Can handle complex, non-linear problems and improve over time through training Key difference - RPA is rule-based static, AI is dynamic and capable of adapting to new scenarios

What is the difference between RPAs and AI?

General agents are multipurpose, while vertical agents focus on specialized, domain-specific tasks

What is the difference between vertical agents and general agents?

Vertical software targets a single industry, horizontal software serves a wide range of users across various fields

What is the difference between vertical software and horizontal software?

Determine previously unknown patterns in the data

What is the goal of unsupervised data

They have geared towards nuclear power plants to power their AI ventures

What recent move has Google made to power their AI

There was a mixed reaction, some supporting his essay and others saying it was overly optimistic

What was the reaction from Dario Amodei's essay

Data entry, document extraction, information transfer, system migrations, and web scraping

What were some promises of RPAs? Name three

1. You don't need a more complicated model to get better attention, you need a model that can grow with more hidden layers 2. Your attention process must evaluate input sequences in parallel rather than sequentially

What were the key ideas from A Vaswani's paper "Attention is All You Need"? (2)

OpenAI To pursue working on AGI

Where did VP of AI at Google, Sébastien Bubeck, leave to go to? Why?

Every step

Where does managing bias fit into the ML steps?

Clusters are not known in advance

Why is K Means clustering "unsupervised"

1. Assumes that all the factors outside the test have remained constant 2. Is assumed that the prior tests don't influence the current test

Why is Naive Bayesian classification called Naive? (2)

Trust Transparency Fairness and Bias Reduction Compliance Better (Human) Decision Making

Why is interpretability important? (5 reasons)

The AI life cycle

Consists of three stages is an iterative, multi-stakeholder approach

Generative AI

Refers to a type of AI that can create new content, such as text, images, audio, or video by learning patterns and structures from existing data. An example would be ChatGPT

Inference time compute

Refers to the computational resources and time required for an AI model to make predictions or generate outputs based on new input data. Is a critical factor in deploying AI systems efficiently for real time applications

Human bias

Reflect systematic errors in human thought. Are omnipresent in the institutions, groups, and individual-decision making processes across the AI lifecycle, and the use of AI applications once deployed

True

T/F: Each individual parameter is a function of a weight, and another value called bias

False - neural networks are NOT exclusively unsupervised ML models

T/F: Neural networks are exclusively unsupervised ML models

False - is not bias in the traditional sense but is a form of bias

T/F: Techno solutionism is bias

AI

The ability of machines to perform tasks that typically require human intelligence. Includes ML and DL

Conditional independence

The model assumes that all the events don't depend on each other

Systemic bias

Type of bias that results from procedures and practices of institutions that operate in ways which result in certain social groups being disadvantaged or favored and others being disadvantaged or devalued

Confirmation bias

Type of human bias where individuals tend to seek out information that confirms their existing beliefs or hypotheses while ignoring or dismissing evidence that contradicts those beliefs

Anchoring bias

Type of human bias, is when a person's judgement is influenced by an initial piece of information, called the "anchor"

Social desirability

Type of human bias, is when individuals answer questions in a way that they believe will be viewed favorably by others

Attention mechanism

Used by LLMs, seeks to determine the significance of a word (or token) as it appears in a string

Supervised learning

Uses data sets that have been labelled to build a model which can be used to predict an outcome

It does not matter who is flipping the coins. One person is as good as another. It does not matter if the coin flipper is alone or in a group, or if it is sunny outside, or if the month has the letter "R" in it

Using a coin analogy, give an example of feature irrelevance

Flipping 3 coins simultaneously is the same as flipping 3 coins one at a time

Using a coin flip analogy, give an example of conditional independence

1. Data collection 2. Data cleaning 3. Model training

What are the steps in ML?

Factual contradictions Prompt contradictions Sentence level

What are the three levels of hallucinations for AI?

Statistical/computational bias Human bias Systemic bias

What are the three types of bias

He believes that within the next 5 to 10 years AI will eliminate a majority of diseases, will lift people out of poverty, and create a renaissance of liberal democracy and human rights

What is Dario Amodei's belief about the possibilities of AI in the future

Robotic Process Automation, is a software technology that uses virtual robots to perform repetitive tasks that are typically done by humans

What is RPA?

It emphasizes the importance of taking risks, embracing competition, and fostering individual agency as catalysts for technological and societal growth.

What is the Techno-Optimist Manifesto's view on risk, competition, and individual responsibility in driving innovation and societal advancement?

Robotics and/or agents that interact with the physical world

What is the application of reinforcement learning

Even if we have very little data, we can still forecast the future by making assumptions and then skewing them based on what we observe about the world

What is the core of Bayes' rule?

Instead of hard coding each individual response, AI will be prompted with an end goal

Why might RPAs be possible with LLMs?

Moore's Law

the observation that computing power roughly doubles every two years.

AlphaGo

Is an AI system developed by DeepMind that mastered the game of Go using reinforcement learning and Monte Carlo tree search

Techno solutionism

Is the belief that technology can solve complex social, political, and economic problems without considering broader societal implications or addressing underlying structural issues

Input context

Is the only cause of a LLM hallucination that we (the user) can do something about. It refers to the information that is given to the model as a prompt

K Means clustering

Is the simplest unsupervised classification model

Bias

Is used in a different way then expected, here it means an extra parameter added to the activation function

ML

A subset of AI that involves training algorithms to learn patterns in data and make predictions or decisions based on that data

Artificial Narrow Intelligence (ANI) / Weak AI

AI that does one specific thing

LLMs are an implementation of Neural Networks LLMs rely on Neural Networks

Describe the difference between LLMs and Neural Networks (2)

True

T/F: AI is any automation that appears to be intelligent

DL

A subset of ML that involves training artificial neural networks to discover patterns in data. Good at handling large, complex datasets and can be used for image recognition, speech recognition. and natural language processing

Artificial General Intelligence (AGI)

A theoretical form of AI that aims to create machines that can perform any intellectual tasks that a human can. Is a mix of ML and DL

Sentence level

AI hallucination level Is when a LLM generates a sentence that contradicts one of the previous sentences

Prompt contradiction

AI hallucination level Is when the LLM doesn't do as it is asked

Backward propagation

Adjusts the weights based on the errors

Neural networks

Are designed to learn from data and make predictions or decisions without explicitly programmed to perform the task

Tokens

Are the text units the model processes

Weights

Are the values learned during training that guide how the model process information

Reinforcement learning

Can be considered a third kind of ML. Its terminology and its focus/application make it distinct

Forward propagation is where the prediction is made, no learning occurs here Backward propagation is where the learning occurs

Describe the difference between forward propagation and backward propagation

The Techno-Optimist Manifesto

Document that advocates for humanity's boundless potential to create a prosperous, adventurous future through the transformative power of technology, rejecting fear-based stagnation and embracing innovation and progress.

Supervisory signal

Each set of data that has the inputs and the expected output

LIME

Explains how the machine learning model works

Semi-supervised learning

Is a type of ML that falls in between supervised and unsupervised learning. Uses a small amount of labeled data and a large amount of unlabeled data to train a model.

Output layer

In a LLM where is error calculated

Anchor

In anchoring bias, this is the initial piece of information that influences a person

Linear regression model

In supervised learning if the outcome is a value that exists on a continuous scale what model would be used?

Both the input and output values

In supervised learning what does the training data contain?

Classification mode

In supervised learning, if the outcome is categorical what type of model would be used?

His view is that LLMs are fundamentally incapable of reaching AGI no matter how many GPUs the hyper scaler uses. Saying they are no more smarter than our pets

In the podcast episode "Is AI Optimism in the Air?" what is LeCun's view about LLMs?

The United State's aging population and the mishandling of AI growth

In the podcast episode, "AI and the looming economic challenge", what is the looming economic challenge?

Parameters

Include weights and other tunable elements (such as bias) that define the model's structure and performance

Factual contradictions

Is a AI hallucination level Is fundamentally incorrect statements

Naive Bayes Classification

Is a model that simulates Bayesian cognition. It is a relatively small, efficient model that makes good estimations with little information

Data collection

Is a step in ML, is the process of gathering observations or measurements. Involves identifying data types, the sources of data, and the methods used.

Training

Multiple iterations of forward propagation, error calculation, and backward propagation

Weights

Numeric measure of strength between neurons. Are also called parameters

Design and development stage

Stage in the AI life cycle, is about choosing the correct model and pre-defining limits before deployment

Pre-design stage

Stage in the AI life cycle, is the initial planning, problem framing, and decision-making step

Deployment stage

Stage in the AI life cycle, requires continuous monitoring and reassessment

Statistical/computational bias

Stem from errors that result when the sample is not representative of the population. Arise from systematic as opposed to random error and can occur in the absence of prejudice, partiality, or discriminatory intent

Model training

Step in ML, is a process in ML where an algorithm learns from a given dataset

Data Cleaning

Step in ML, is the process of fixing or removing incorrect, corrupted, incorrectly formatted, duplicate, or incomplete data within a dataset

Artificial Super Intelligence (ASI)

a hypothetical form of AI that is capable of surpassing human intelligence by manifesting cognitive skills and developing thinking skills of its own

Unsupervised learning

attempts to discover unknown patterns in data

Forward propagation

computes the predicted output based on the current state of the network's weights and biases