Intro to Ai

Question 1

What is AI?

Answer 1

AI is a branch of computer science focused on building smart machines capable of performing tasks requiring human intelligence.

Question 2

What are algorithms in the context of AI?

Answer 2

Algorithms are finite sequences of instructions that terminate when run, necessary for computers to perform tasks.

Question 3

What is the Von Neumann Model?

Answer 3

A model that describes how algorithms are structured and executed in computers.

Question 4

What is an NP-Hard problem?

Answer 4

A problem that lacks efficient algorithms or has algorithms that are computationally expensive.

Question 5

What is Narrow AI?

Answer 5

Specialized AI that excels in specific tasks but operates within a limited context.

Question 6

What is General AI?

Answer 6

Human-level AI capable of reasoning, understanding its environment, and transferring knowledge across domains.

Question 7

What is Super AI?

Answer 7

AI that surpasses human intelligence in all fields, including creativity and reasoning.

Question 8

What is the AI Singularity?

Answer 8

A hypothetical point in the future where technological growth becomes uncontrollable, resulting in superintelligence.

Question 9

What is Machine Learning?

Answer 9

A field that enables computers to learn how to solve problems from data without needing explicit algorithms.

Question 10

What is required for effective learning in Machine Learning?

Answer 10

A high volume of high-quality data.

Question 11

What are Neural Networks?

Answer 11

Inductive learning models loosely based on biological brains, trained using labelled datasets.

Question 12

What is the structure of Artificial Neural Networks (ANNs)?

Answer 12

Consists of artificial neurons organized in layers, with each neuron having:
- Multiple inputs and one output
- Processes inputs using a summation unit and a transfer (activation) function.
- Uses weights (real numbers) associated with each input, they can be positive or negative.

Question 13

What is a Perceptron?

Answer 13

The first neural network capable of simple computations like Boolean AND/OR. Training involves finding weights to produce correct outputs from given inputs. For simple networks, weights can sometimes be determined visually.

Question 14

What is the purpose of the Error Backpropagation Algorithm (Backprop)?

Answer 14

To efficiently update weights iteratively to minimize error in output and improve network performance.

Question 15

What is Deep Learning?

Answer 15

An advanced form of neural networks with many layers, automating complex learning tasks.

Question 16

What challenges did deep learning address?

Answer 16

Manual feature extraction, limited hardware, and algorithm inefficiencies.

Question 17

What are Convolutional Neural Networks (CNNs)?

Answer 17

Networks that convert images to feature vectors without human intervention for tasks like image recognition.

Question 18

What is the Turing Test?

Answer 18

A test to evaluate a machine’s ability to demonstrate intelligent behavior indistinguishable from a human’s.

Question 19

What are the strengths of the Turing Test?

Answer 19

• Simple and clear criterion for assessing AI intelligence
• Focus on human interaction
• Historical significance in AI research

Question 20

What are criticisms of the Turing Test?

Answer 20

• Narrow scope
• Dependence on deception
• Cultural bias
• Advances in AI challenge its relevance
• Philosophical limitations

Question 21

What is the Chinese Room Experiment?

Answer 21

A thought experiment that challenges the idea that machines can possess a mind or consciousness like humans.

Question 22

What is the Computational Theory of Mind (CTM)?

Answer 22

CTM posits that the mind is analogous to a universal computer with cognitive processes as computational operations.

Question 23

What are the critiques of CTM?

Answer 23

• Emotional and intuitive cognition
• Consciousness issues
• Embodied cognition overlooks the role of the body

Question 24

What is a Transformer in AI?

Answer 24

A specialized form of deep neural network that revolutionized natural language processing by understanding context and relationships at unprecedented scales.

Question 25

What are some key areas impacted by Generative AI?

Answer 25

• Natural Language Processing
• Content Creation
• Search and Information Retrieval
• Education and research

Question 26

What are some ethical and social implications of AI?

Answer 26

• Misinformation or hallucinations
• Impact on employment
• Bias and fairness
• Existential concerns

Question 27

Are Large Language Models (LLMs) sentient?

Answer 27

No, they process and generate responses based on patterns in their training data without understanding or emotion.

Question 28

Training a Multi-Layer Perceptron

Answer 28

An MLP is an ANN with hidden layers. Training involves finding weights to correctly classify inputs. For example, classifying images of cats or dogs. It outputs [0,1] for cat and [1,0] for dog. The training set is based on 50 labelled images of cats and 50 labelled images of dogs. Images are represented as vectors via feature extraction. The weights are initially randomized to a real number between 0 and 1.

Question 29

Deep Learning Challenges

Answer 29

Neural networks with more than 3 layers were hard to train, manual feature extraction was time consuming and subjective. These were old problems (~2006).
Then in 2006 there was the deep learning breakthrough which enabled multi-layer networks, automating complex learning tasks. It became feasible due to advancements in algorithms and hardware.

Question 30

Factors that enabled deep learning

Answer 30

Technological factors such as Computing Power (Multi-core CPUs, GPUs for parallel processing etc.), Faster RAM, SSDs, and High-speed connections with unlimited bandwidth

Question 31

Deep learning has transformed AI fields

Answer 31

Machine Vision: Enhanced image recognition and object detection.
Machine Speed: Improved speech synthesis and recognition.
Natural Language Processing: Better language understanding and generation.
It leveraged computer hardware/software for advanced AI libraries.

Question 32

problems with MLPs

Answer 32

that they struggle with geometric transformations in images. Convolutional Neural Networks address these issues.

Question 33

CNN Structure

Answer 33

Structure:
- Feature Extraction: 2 convolutional layers, each followed by pooling/downsampling
- Classification: Two fully connected layers, Output is one hot encoded

Question 34

The Thought Experiment

Answer 34

Setup: Imagine an English speaker locked in a room with a rulebook explaining how to translate Mandarin Chinese to Cantonese Chinese without understanding either language. Mandarin input is passed into the room, and the person uses the rulebook to output Cantonese translations.
Key Insight: The person in the room doesn’t understand they just follow rules mechanically.
Replace the person with a computer program, does the program understand Chinese?

Question 35

Criticisms of the Chinese Room

Answer 35

Misunderstandings of AI: Assumes AI operates like the person in the room, following predefined rules. It is argued that modern AI especially ML adapts and learns from data, making this comparison inadequate.

Systems Reply: While the person in the room doesn’t understand Chinese, the system as a whole (person+instructions) could be said to understand. Suggests that understanding emerges at a system level not at an individual level.

Nature of Understanding: The experiment’s narrow definition of understanding is criticised. If a system can reliably interact and produce meaningful responses, should that not qualify as understanding?

Robot Reply: Some argue that embedding AI in a robot capable of perceiving and interacting with the world could lead to genuine understanding, as intelligence and understanding may be rooted in sensory and environmental interaction.