Brainscape's Knowledge GenomeTM


Top Flashcards (Certified)
All Flashcards

AIGP > Section 5 > Flashcards

Section 5 Flashcards

(76 cards)

Start Studying
1
Q

Four stages of AI development lifecycle

A

Plan, Design, Develop, Deploy

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

Stages involved in planning (6)

A

Business problem, Mission, Gaps, Data, Scope, Governance

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

Stages involved in design (2)

A

Implement data strategy (Data quality, Data gathering, Data wrangling/preparation, Data cleansing, Data labeling, Data privacy), Determine system architecture

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

Five “V’s” of data wrangling and preparation

A

Variety, Value, Velocity, Veracity, Volume

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

Components of determining a system architecture (6)

A

Choose appropriate algorithm or architecture, Consider desired accuracy and interpretability, Objective of data, Business problem, Compliance/business requirements, Constraints

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

Feature engineering

A

Crucial pre-processing step to transform raw data into relevant information and create predictive model features

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

Base data pile

A

Final dataset including training data, testing data, and validation data

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

Data poisoning

A

An attack where a malicious actor intentionally compromises a training dataset

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

Stages involved in development (5)

A

Define features, Engineer features, Train model, Test model, Validate model

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

Purposes of feature engineering (3)

A

Improve model performance, Reduce computational costs, Boost model explainability

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q

Building an impact assessment involves (4)

A

Adapting existing frameworks, incorporating AI principles, addressing key risks, customizing by use case

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

Confusion matrix

A

Measures where a model is confused, assessing classification model performance

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

False positive

A

Type I error

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
14
Q

False negative

A

Type II error

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
15
Q

HUDERIA Risk Index Number (RIN)

A

Value from context-based risk analysis to guide proportionate risk management and stakeholder engagement

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
16
Q

RIN calculation variables

A

Gravity potential, Rights-holders affected, Severity, Likelihood

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
17
Q

Gravity potential

A

Worst-case scenario harm severity (1=minor, 2=serious, 3=critical, 4=catastrophic)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
18
Q

Rights-holders affected

A

0.5 = 1-10K; 1 = 10K–1M; 2 = over 1M individuals

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
19
Q

RIN scoring levels

A

Low: ≤5; Moderate: 5.5–6; High: 6.5–7.5; Very High: ≥8

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
20
Q

Risk mitigation hierarchy

A

Avoid, Reduce, Restore, Compensate

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
21
Q

Avoid (risk mitigation)

A

Make changes to design or development to avoid impacts

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
22
Q

Reduce (risk mitigation)

A

Implement actions to minimize potential impacts

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
23
Q

Restore (risk mitigation)

A

Rehabilitate affected individuals to pre-impact condition

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
24
Q

Compensate (risk mitigation)

A

Provide restitution when other mitigations are not possible

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
25
Approaches to global regulation (3)
Amend existing laws, Create comprehensive regulations, Focus on specific risk areas
26
Brazil AI regulation
Comprehensive, risk-based legislation requiring human oversight for rights-implicating AI
27
India AI guidance
AI must not facilitate illegal content; label content from in-development or unreliable models
28
Japan AI governance
Non-binding guidelines with national strategy deferring to private sector
29
Canada Bill C-27 (3)
Digital Charter Implementation Act with three components: AIDA, Consumer Privacy Protection Act, Data Protection Tribunal Act
30
Canada AIDA
Risk-based; includes anonymization, high-impact assessment, risk mitigation, monitoring, audits
31
Canada Consumer Privacy Protection Act
Replaces PIPEDA privacy parts; requires explanation of automated decision-making
32
Canada Data Protection Tribunal Act
Creates tribunal to hear privacy appeals
33
Canada Voluntary Code of Conduct for GenAI
Common standards for responsible AI development and deployment
34
Cyberspace Administration of China (CAC)
Regulates cyberspace, cybersecurity, and the internet
35
China's Generative AI Measures
Regulates public GenAI use; includes transparency, anti-monopoly, consent, IP, training, and labeling requirements
36
Singapore Model AI Governance Framework
Risk-based, industry-led; principles: human-centric, explainable, transparent, fair
37
Singapore sector-specific AI guidance
Health: AI in Healthcare Guidelines; Finance: FEAT principles and Veritas framework
38
Singapore National AI Strategy 2.0
Goals: Maximize value creation and empower AI adoption with trust across individuals, businesses, and communities
39
Article 13 (Singapore)
Requires government to regularly review and update AI laws and standards
40
Singapore AI governance regulators
IMDA (Info-communications & Media Development Authority), PDPC (Personal Data Protection Commission)
41
Singapore Model AI Governance Framework Deliverables
Self-assessment guide, Use case compendium
42
AI Verify
AI governance testing framework and toolkit to demonstrate compliance
43
Generative AI Framework and Sandbox (Singapore)
Tools for testing and developing GenAI under regulatory guidance
44
Data provenance
Documents origin, history, and authenticity of data
45
Data lineage
Maps flow, transformations, and dependencies of data in systems
46
Semi-structured data
Data with organizational properties but no rigid structure (e.g., emails, digital health records)
47
Assurance in AI governance
body of frameworks, policies, processes and controls to measure, evaluate, and promote trustworhty AI. Includes conformity and impact assessments, risk assessments, audits, certifications, valiation and testing
48
System architecture
Structure, components, and organization of an AI model
49
Feedforward neural network
Data flows in one direction; simple model structure
50
Convolutional neural network (CNN)
uses multiple layers to filter and extract distinctive features from input data. Generally excel in classification, object detection and intricate visual tasks
51
Recurrent neural network (RNN)
process data bi-directionally to handle tasks that require memory, such as speech recognition or time-series analysis
52
Graph neural network (GNN)
process data represented in graph structures by understanding and analyzing how data points are connected (e.g. LinkedIn uses GNN to analyze social network connections)
53
Ensemble method
Combines multiple models into one system for improved performance
54
Types of ensemble methods
Stacking, Bagging, Boosting
55
Things to test for in AI systems (7)
Bias, Accuracy, Reliability, Robustness, Privacy, Interpretability, Safety
56
Stacking
train multiple models; meta data synthesizes outputs
57
Bagging (or bootstrap aggregation)
train same model on different subsets of data and aggregate the outputs by taking an average
58
Boosting
sequential building of simple models, each improving on the last
59
Types of testing (4)
Repeatability, Adversarial testing, Threat modeling, Counterfactual explanations
60
Counterfactual explanations
Show how small input changes can lead to different model outputs
61
Pseudonymization
Replaces personal identifiers in data
62
De-identification
Removes some personal identifiers
63
Anonymization
Removes all personal identifiers
64
Data obfuscation/masking
Modifies data to make it useless to unauthorized users
65
Four privacy-enhancing technologies
Homomorphic encryption, Secure multi-party computation, Differential privacy, Federated learning
66
Homomorphic encryption
Enables computation on encrypted data without exposing it
67
Secure multi-party computation
Allows joint computation across parties without revealing inputs
68
Differential privacy
Adds noise to prevent reidentification in datasets
69
Federated learning
Trains models across data silos without centralizing data
70
Model card regulatory check
OECD tool for assessing model documentation compliance
71
Datasheet for dataset
Describes dataset composition, collection, motivation, and use
72
Repeatability assessment
Same team can replicate results under identical conditions
73
Ways LLMs can fail (4)
Brittleness, Embedded bias, Uncertainty, Catastrophic forgetting
74
Acceptable use policy key aspects
Optionality, Redress
75
Optionality
users ability to opt out or opt in
76
Redress
organization's ability, willingness to correct or address a wrong

AIGP (4 decks)

Brainscape helps you reach your goals faster, through stronger study habits.
© 2025 Bold Learning Solutions. Terms and Conditions