Paper 3

Question 1

What is Latency in the context of Natural Language Processing?

Answer 1

The delay that occurs in processing user input and generating a response

Question 2

Define Dependency Reduction in the Critical Path.

Answer 2

Strategies for streamlining processes by minimizing dependencies that cause delays

Question 3

What is the Critical Path?

Answer 3

The shortest and most efficient sequence of linked machine learning models required to process a user’s message

Question 4

What are the main types of analysis challenges in Natural Language Processing?

Answer 4

• Syntactic Analysis Failures
• Semantic Analysis Failures
• Pragmatic Analysis Failures

Question 5

What is Syntactic Analysis Failures?

Answer 5

Problems with understanding sentence structure

Question 6

What is Semantic Analysis Failures?

Answer 6

Challenges in understanding meanings, such as idioms and homonyms

Question 7

What does Pragmatic Analysis Failures refer to?

Answer 7

Issues with understanding context, tone, or cultural differences

Question 8

What is an example of a Syntactic Analysis Failure?

Answer 8

Misunderstanding a request due to poor sentence structure

Question 9

What is Data Augmentation in NLP?

Answer 9

The process of artificially generating new data based on existing training data

Question 10

What are some hardware types used in NLP?

Answer 10

• CPUs
• GPUs
• TPUs

Question 11

What is the role of CPUs in machine learning?

Answer 11

They are cost-effective for non-parallel processing tasks

Question 12

What are GPUs particularly good at?

Answer 12

Matrix operations and tasks requiring parallel processing

Question 13

What is the von Neumann bottleneck?

Answer 13

The limitation of memory access speed compared to calculation speed in CPUs

Question 14

What is a common challenge faced by Natural Language Processing models?

Answer 14

Biases such as Historical Bias, Linguistic Bias, and Sampling Bias

Question 15

What is Historical Bias in NLP datasets?

Answer 15

Bias that arises from historical inequalities reflected in training data

Question 16

What is an example of Sampling Bias?

Answer 16

Algorithms predicting healthcare needs favoring one demographic over another

Question 17

What is an ethical challenge in AI?

Answer 17

Accountability for decisions made by AI systems

Question 18

True or False: Data augmentation can help reduce bias in AI training datasets.

Answer 18

True

Question 19

What is an example of a real-world scenario where NLP failed?

Answer 19

McDonald’s Drive Thru Chatbot Beta Test leading to incorrect orders

Question 20

What strategies can be used to solve tokenization problems in NLP?

Answer 20

• Querying large tables of edge cases
• Using context-aware tokenization models
• Applying more complex heuristics

Question 21

Fill in the blank: _______ are problems with understanding context, tone, or cultural differences.

Answer 21

Pragmatic Analysis Failures

Question 22

What is the impact of cultural biases in NLP models?

Answer 22

Models trained on English datasets may not understand biases relevant to other cultures

Question 23

What is a potential solution to mitigate biases in NLP models?

Answer 23

Data augmentation and training on diverse datasets

Question 24

What is a common issue with tokenization in languages like Chinese?

Answer 24

Defining a ‘word’ is difficult due to the lack of spaces

Question 25

What technology advancements have GPUs made beyond high-performance graphics?

Answer 25

High-speed data processing and massively parallel computations ## Footnote GPUs are now used in various applications, including machine learning.

Question 26

What type of calculations do GPUs handle in graphics rendering?

Answer 26

Complex mathematical and geometric calculations ## Footnote This includes operations like matrix multiplication with floating point numbers.

Question 27

Why are GPUs considered ideal for models that require parallel processing?

Answer 27

They are capable of handling lots of computation efficiently ## Footnote This is particularly true for models like Transformers.

Question 28

What are TPUs specialized for?

Answer 28

Neural network workloads ## Footnote TPUs are designed specifically for matrix processing, which is essential for deep learning.

Question 29

What architecture do TPUs utilize for matrix operations?

Answer 29

Systolic array architecture ## Footnote This consists of thousands of accumulating units connected on a single processor.

Question 30

What is the grid size of Google's Cloud TPU v3?

Answer 30

128 x 128 grid of ALUs ## Footnote This design allows for efficient propagation of products through the processor.

Question 31

From a capitalist viewpoint, what should happen to a chatbot providing ethically problematic responses?

Answer 31

It should not be profitable ## Footnote Users would likely take their business elsewhere if the chatbot is not enjoyable to use.

Question 32

What ethical challenge is associated with chatbots that are not transparent about their use?

Answer 32

They are perceived as less ethical by customers ## Footnote This perception can lead to reputational damage for companies using such chatbots.

Question 33

What was one of the errors made by Google’s Bard chatbot?

Answer 33

Claiming the James Webb Space Telescope captured the first images of a planet outside our solar system ## Footnote Such misinformation can damage the reliability of chatbot companies.

Question 34

What is prompt hacking?

Answer 34

A hacking technique that manipulates prompts to exploit vulnerabilities of LLMs ## Footnote This can lead to unintended responses from language models.

Question 35

What are the three types of prompt hacking?

Answer 35

* Prompt jailbreak * Prompt injection * Prompt leaking ## Footnote Each type involves different methods of exploiting LLM vulnerabilities.

Question 36

What is required for accountability in AI systems according to the European Commission?

Answer 36

Fair impact on people's lives, adherence to values, and suitable accountability processes ## Footnote These principles ensure responsible use of AI.

Question 37

What are the three conditions necessary for accountability?

Answer 37

* Authority recognition * Interrogation * Limitation of power ## Footnote These conditions must be met for an agent to be held accountable for actions.

Question 38

Fill in the blank: TPUs cannot run a word processor or perform the same tasks as a CPU because they are not _______.

Answer 38

Turing complete ## Footnote This limitation means TPUs are specialized and not general-purpose processors.

Question 39

What ethical consequences can arise from chatbot errors or misinformation?

Answer 39

Loss of reliability and client support ## Footnote This can have serious implications for businesses using chatbots.

Question 40

True or False: Prompt injection is a type of attack that can produce undesired input by overriding the original chatbot prompt.

Answer 40

True ## Footnote This type of attack can lead to malicious outputs from chatbots.