AWS AI Practitioner Flashcards
Where can I get free AWS AI Practitioner Certification training?
AI Practitioner Practice Exam Questions
What level is the AWS AI Practitioner certification?
Foundational
What are the AWS AI Practitioner Testing specifics?
120 minutes for 85 questions
5 Domains of AI Practitioner Exam
AI & ML (20%)
Generative AI (24%)
Applications of Foundation Models (28%)
Responsible AI (14%)
AI Security Compliance & Governance (14%)
Define AI
Any way a computer can mimic human intelligence
Define ML
AI subset. Identifies patterns within data. Uses patterns to predict new patterns.
Define Deep Learning
ML subset. Neural Networks. Trying to replicate how the brain works.
Define Generative AI
DL subset. Uses DL to generate content.
How is traditional programming different than ML?
Traditional uses human to write the code.
ML uses a learning algorithm to create a model based on its learnings to produce a prediction.
When to use ML vs. Traditional programming
- When you cant code it (too complex to write the code)
- When you cant scale it (e.g. fraud detection, spam, recommendations)
- When you have to adapt/personalize (e.g. book recommendations)
- When you cant track it (data coming in to quickly, e.g. automated driving)
Conditional ML is _______ in its predictions
Deterministic
Generative IA is ________ in its predictions
Non-Deterministic. Can make things up
Why use ML over Gen AI?
Transparency * Interpretability
Explainability & Fairness
Robustness & Consistency
Data Efficiency
Specific Task Performance
3 Machine Learning Types
Supervised, Unsupervised, Reinforcement
What is ML/Supervised Learning
Data is assigned a label. Supervised learning uses the labels and repetition to learn. (E.g. how little kids learn).
Labeled data is put into a ML algorithm which makes a prediction. A human corrects /adjusts the model if the prediction is incorrect.
What is unsupervised learning?
No labeled data. Uses grouping data to predict other relationships. E.g. these type customers bought x so they will likely buy Y
What is ML/Reinforcement learning?
Learns through reinforcement. E.g. Dog training.
Key areas of ML/Supervised Learning
Regression (Numeric e.g. predict house prices), Classification (Binary (a or b class) or Multiclass (can fit into several classes)
Key area of ML/Unsupervised Learning
Grouping/Clustering (Labels unknown or Finds data in Patterns)
Key area of Reinforcement learning
Correct actions rewarded (if agent does x, do y)
ML/Self-Supervised Learning
The basis of Generative AI. Unlabled data has a word removed. The model then predicts the word that was removed. Do this over and over and the prediction becomes accurate.
Label/Target
Dependent variable: what you are attempting to predict
Feature
Independent variable: data that helps you make predictions
Feature Engineering
Data Transformation: process of reshaping data to get more value
Feature Selection
Variable/subset selection: process of using the most valuable data
ML Development Lifecycle
Problem/ML Problem Framing/Data Collection/Data Integration/Data Prep/Data viz & analysis
Key Benefit of Sagemaker
It manages the entire ML lifecycle
Rekognition
ML service: Image and video analysis
Textract
ML service: extracts text and data from documents (e.g. OCR)
Comprehend
ML service: discovers insights and relationships in text
Kendra
ML service: machine learning search service that allows you to use natural language questions
Personalize
ML service: create personalized experiences (e.g. retail site - prompt new product discovery)
Fraud Detector
ML service: helps id fraudulent activity
What would you use to develop a learning model to analyze customer feedback on products. Customer will train the ml model
Supervised Learning
Ecommerce company needs to integrate tailored recommendations to customers based on browsing and purchase history.
Amazon Personalize
A ML engineer wants to implement a ml pipeline on AWS to automate training and deploying models. Pipeline should include data preprocessing, training and model deployment
Sagemaker
Zero-Shot Prompting
The larger the LLM, the more likely the zero-shot prompt will yield effective results
What improves zero-shot tuning
Instruction Tuning
Few-Shot Prompting
Provide several examples to help the model
Chain-of-thought prompting
Use CoT prompting when the task involves several steps or a requires a series of reasoning
What is prompt injection?
When an attacker
What is prompt leaking?
When we leak sensitive information
What are two common malicious activities for AI
Prompt Injection and Prompt Leaking
What is a prompt template?
Use when using prompts for large data sets when you are applying several variations of a prompt.
RAG
Retrieval Augmented Generation`
What is RAG?
Framework for building generative AI applications that can make use of data sources
What is a Transformer?
What is a Token?
What is prompt engineering?
What are the costs with the different prompt engineering options?
What is the use case for Bedrock guardails?
What makes up responsible AI?
Fairness, Explainability, Controllability, Safety, Privacy & Security, Governance, Transparency, Veracity & Robustness
What is dataset Bias?
The imbalance in data used to train a ML model
What are common types of dataset Bias?
Sampling, Historical, Measurement
What is Sample Bias?
Data does not represent the true population
What is historical Bias?
Data reflects past biases and inequities in society
What is Measurement Bias?
the data collection process introduces errors
I can you identify imbalances in data sets
Calculate the ration of the smaller class vs. total data
What are model generalization problem examples
Underfitting, Overfitting, Appropriate Fitting
What is Underfitting?
model is too simple
How do you identify Underfitting?
Poor training and test results
How do you fix Underfitting?
Increase Training Data or passes through the existing data
What is Overfitting?
Model picks up noise instead of underlying relationships
How do you identify Overfitting?
Model training does good. When you add more data, the model stops working.H
How do you fix Overfitting?
Reduce model flexibility: fewer features, smaller groups of data per training job, adjust weights
What type of bias and variance (H,M,L) is ideal for a model
Low bias and low variance
AWS ML visually explained
https://mlu-explain.github.io/
What is a linear regression?
What is
What is Reinforcement Learning?
What
what are the stages in governing the model lifecycle?
Onboard, Build, Train, Deploy, Monitor,
What is ROC?
What is A2I?
Amazon Augmented AI. Provides a human review of ML predictions
What is SageMaker Clarify?
Helps to detect Bias in ML.
ID data imbalances, check trained model for bias, explain overall behavior
SageMaker Model Monitor
Monitors ML models in production. Detects drift and quality issues
What is SageMaker Role Maker?
Define minimum permissions using premade permission sets.
What are SageMaker Model Cards?
Document, retrieve and share model information
What is SageMaker Model Dashboard?
Monitor model performance through a unified view.
