Intro to AI

Question 1

Artificial Intelligence (AI)

Answer 1

Broad capability to imitate human abilities.

Question 2

What is AI? History/origins? Milestones?

Answer 2

AI involves machines mimicking human cognitive functions.

Started in 1950s with Turing Test. And in 1956 the term ‘AI’ was introduced by John McCarthy.

Milestones:
- 1997: chess-playing computer (Deep Blue) defeats chess champion Garry Kasparov
- 2011: Apple integrates Siri into iPhone
- 2014: Amazon Alexa virtual assistant
- 2020: GPT-3 tool
- 2022: ChatGPT available to public

Question 3

What is the Turing Test?

Answer 3

An intelligent test for machines. - if a machine tricks humans into thinking it is a human, then it has intelligence.

Question 4

Machine Learning (ML)

Answer 4

Subset of AI; algorithms improve through experience and learning.

Question 5

Deep Learning (DL)

Answer 5

Subset of ML; uses neural networks with many layers

Question 6

Generative-AI (Gen-AI)

Answer 6

subset of DL; generates new data and outputs (text, e.g. ChatGPT; image, e.g. Dall-E, Stable Diffusion, music)

Question 7

Large Language Models (LLMs)

Answer 7

subset of Gen-AI; GPT-series (ie. GPT-4, Bard, Claude)

Question 8

Top-Down Approach to AI

Answer 8

Aka - Symbolic Reasoning

• Models the way a person reasons to solve a problem; we, humans, model our reasoning process and program it into computers
• Involves extracting knowledge from a human being, representing it in a computer-readable form.
• Develop a way to model reasoning inside a computer.
• uses logic and reasoning to work
• old way/approach

Question 9

Bottom-up approach to AI

Answer 9

aka - Neural Networks

• Learns from examples
• Models the structure of a human brain, using neurons to form understanding of patterns
• each neuron acts like a weighted avg of inputs - we can train a network of neurons to solve useful problems by training data; learning from examples.
• new approach - this is what is used now

Question 10

Types of ML

Answer 10

• Supervised
• Unsupervised
• Reinforcement

Question 11

What is supervised ML?

Answer 11

learning with labeled data

Question 12

What is unsupervised ML?

Answer 12

Identifies patterns in data w/o labels
- e.g. spam emails contain certain patterns that are common

Question 13

What is reinforcement ML?

Answer 13

Learning via feedback and rewards

Question 14

What is big data?

Answer 14

large, complex, diverse and high quality datasets

Question 15

Connection between big data and AI?

Answer 15

• Data as the foundation: Ai systems learn from data.
• The more data an AI system has, the better it can learn and make accurate predictions, decisions, recognise patterns, improve performance over time.
• the data must be cleaned, labelled, and prepared for use in AI models
• But not always.. Memory must be sufficient for AI to decipher big data

Question 16

What is self-supervised learning?

Answer 16

LLMs learning - use inherent data properties and label the data themselves

Question 17

List of LLMs (newest to oldest)

Answer 17

Newer Models:
1. OpenAI: GPT-4o (May 2024)
2. Anthropic: Claude 3.5 (March 2024)
3. Google DeepMind: Gemini (December 2023)
4. xAI: Grok-1 (November 2023)
Mistral AI: Mistral 7B (September 2023)
Original List:
Google: Bard (March 2023)
Anthropic: Claude (March 2023)
OpenAI: ChatGPT (November 2022)
Meta: OPT-IML (2022)
Baidu: Ernie 3.0 Titan (December 2021)
DeepMind: Gopher (December 2021)
NVIDIA: Megatron-Turing NLG (October 2021)
AI21 Labs: Jurassic-1 (August 2021)
Google: LaMDA (May 2021)

Question 18

Basic terms for neural networks

Answer 18

• Neurons
• Weights
• Activation functions
• paths
• Bias

Question 19

Convolutional Neural Network (CNNs)

Answer 19

used for image data

Question 20

Recurrent Neural Networks (RNNs)

Answer 20

suitable for sequence data

Question 21

How do neural networks work?

Answer 21

Each neuron acts like a weighted average of its inputs, and we can train a network of neurons to solve useful problems by providing training data.

Question 22

What is the activation function?

Answer 22

Function that generates output based on input signal and their weights.
- uses a threshold to send info

Question 23

What is Bias (B)?

Answer 23

= a number to calibrate the summation function; a fine tuning function that helps achieve better results
- E.g.: Facial recognition, forecasting by identifying patterns, music composition

Question 24

What is Natural Language Processing (NLP)?

Answer 24

Making sense of and generating human language

• used in chatbots, LLMs, translators

Question 25

What are the challenges to NLP?

Answer 25

sarcasm, context, cultural nuances
- different word, same meaning
- Different expression, same meaning
- different grammar same meaning
- same word, different context

Question 26

How Dall-E 2 works?

Answer 26

1. Image Caption: encoding - starts with textual description which is the prompt
2. Text embedding: textual description converted to numerical representation
3. Prior: Randomness called the “prior” is utilised to introduce diversity in generated images / noise
4. Encoder: text embedding and prior are combined and processed through an encoder (basc. Mix text and numbers together)
5. Decoder: encoder’s immediate representation is fed into a decoder - transforms it into a synthesised image
6. Image Generation: through the decoder’s iterative process, final img is generated.