AWS Certified AI Practitioner flashcards

An AI technique that enhances the capabilities of a Large Language Model (LLM) by allowing it to retrieve relevant information from an external knowledge base before generating a response.

• Combats LLM hallucinations by grounding responses in facts.
• Enables LLMs to access and use up-to-date or private information.
• Reduces the need for continuous model retraining for new data.

Memory trick: RAG Retrieves Answers, Grounding Generations.

A type of foundation model engineered to process, understand, and generate data across multiple distinct modalities, such as text, images, audio, and structured data.

• Capable of learning relationships between different data types.
• Enables richer understanding and more complex applications.
• Often more computationally demanding than single-modality models.

Memory trick: Multi-modal Models Master Many Modalities.

Human-in-the-Loop (HITL) is an approach to AI development and deployment where human intelligence is combined with machine intelligence to achieve better outcomes, often involving human review and intervention.

• Crucial for high-stakes AI applications.
• Allows for human judgment and ethical oversight.
• Can improve model accuracy and reduce bias over time.

Memory trick: Humans and AI, a powerful pair.

AI fairness ensures that an AI model's decisions are free from bias and do not lead to systematically different or discriminatory outcomes for different groups of people.

• Addresses disparate impact and treatment.
• Critical for ethical AI deployment.
• Requires careful data selection and model evaluation.

Memory trick: F.I.R.E. P.R.O.T.E.C.T.S. AI

New, often surprising, capabilities that appear in foundation models as their scale (parameters, data) increases significantly, which were not present in smaller versions.

• Not explicitly trained for, but 'emerge' from scaling.
• Often include complex reasoning, instruction following, and world knowledge.
• Indicate a non-linear relationship between scale and capability.

Memory trick: Emergent Abilities Emerge, Expanding Expectations.

AI privacy refers to the ethical and legal responsibility to protect personal and sensitive data used by AI systems, ensuring compliance with data protection regulations and individual rights.

• Includes anonymization, pseudonymization, differential privacy.
• Crucial for trust and regulatory compliance.
• Balances data utility with individual protection.

Memory trick: P.S.F.A.T. for AI

A technique where a pre-trained foundation model is further pre-trained on a large dataset specific to a target domain, allowing it to learn domain-specific knowledge and representations.

• Also known as continued pre-training.
• Enhances model performance on specialized tasks without full retraining from scratch.
• Effective for adapting models to medical, legal, or proprietary enterprise data.

Memory trick: DAPT Delivers Domain-Specific Deepness.

AI robustness refers to an AI system's ability to maintain its performance and integrity when faced with novel, adversarial, or noisy input data, or changes in its operating environment.

• Crucial for security and reliability.
• Mitigates adversarial attacks and data drift.
• Enhanced through techniques like adversarial training.

Memory trick: Strong AI for Secure Futures

The phenomenon where a Large Language Model generates information that is plausible-sounding but factually incorrect, nonsensical, or fabricated.

• Arises from the probabilistic nature of next-token prediction.
• A significant challenge in deploying LLMs for factual tasks.
• Mitigated by techniques like Retrieval Augmented Generation (RAG).

Memory trick: Next-Token Prediction Powers Plausible Problems.

The ability of a pre-trained foundation model to perform well on new, unseen data or tasks that differ from its original training data, often with minimal fine-tuning.

• Achieved through pre-training on massive, diverse datasets.
• Enables few-shot or zero-shot learning for downstream tasks.
• Reduces the need for extensive labeled data for specific applications.

Memory trick: Generalization Grants Great Gains.

Hosting and running a foundation model entirely within a private, local data center or dedicated private cloud infrastructure, rather than using a public cloud service.

• Provides maximum control over data sovereignty and security.
• Essential for highly sensitive data and strict regulatory compliance (e.g., HIPAA, GDPR).
• Requires significant internal IT infrastructure and expertise.

Memory trick: Private Placement Protects Patient Privacy.

A type of foundation model trained on massive text datasets, capable of understanding, generating, and processing human language.

• Excels at natural language processing (NLP) tasks.
• Used for text generation, translation, summarization, and question answering.
• Forms the basis for many AI assistants and chatbots.

Memory trick: LLMs Learn Language, Lifting Limits.

Recall, also known as sensitivity, measures the proportion of actual positive cases that were correctly identified by an AI model. It is crucial when the cost of false negatives is high.

• Calculated as True Positives / (True Positives + False Negatives).
• High recall means fewer actual positive cases are missed.
• Important for rare disease detection, fraud detection, etc.

Memory trick: P.R.A.F. for Balanced Views

AI transparency refers to the ability to understand how an AI system works, including its data, algorithms, and decision-making processes.

• Aids in debugging and trust.
• Different levels: model, data, and process.
• Often linked to interpretability and explainability.

Memory trick: See-Through AI for Trust

Addressing data bias involves identifying and mitigating systematic errors or imbalances in training data that can lead to unfair or inaccurate AI model predictions.

• Data quality is paramount for fair AI.
• Underrepresentation leads to poor performance on specific groups.
• Solutions include data augmentation, re-sampling, and collecting new data.

Memory trick: DATA CLEANUP for AI Trust

An Ethical Impact Assessment (EIA) is a systematic process of identifying, analyzing, and evaluating the potential ethical, societal, and human rights impacts of an AI system throughout its lifecycle.

• Proactive risk identification before deployment.
• Considers fairness, privacy, accountability, and safety.
• Involves stakeholders and multidisciplinary expertise.

Memory trick: Plan, Build, Assess, Deploy, Monitor

AI Interpretability, or explainability, refers to the extent to which a human can understand the cause of a decision made by an AI model. It allows stakeholders to comprehend the rationale behind an AI's output.

• Crucial for building trust and debugging.
• Techniques include LIME, SHAP, and feature importance.
• Helps identify bias and ensure compliance.

Memory trick: Explain Your AI, Build Trust.

AI accountability involves establishing clear responsibility for the design, development, deployment, and outcomes of AI systems, especially in cases of error, harm, or misuse.

• Crucial for trust and legal compliance.
• Requires clear governance structures.
• Encompasses both technical and ethical considerations.

Memory trick: A.C.T. F.A.I.R. AI

Systematic and unfair prejudice in the outputs of a foundation model, often stemming from biases present in its vast training data.

• Can lead to discriminatory or suboptimal outcomes.
• Manifests in various forms: gender, racial, cultural, or even technical preferences.
• Mitigation involves diverse data collection, bias detection, and ethical fine-tuning.

Memory trick: Bias Blocks Better Behavior.

The process of modifying a foundation model to better align its outputs with human values, intentions, instructions, and specific desired behaviors or styles.

• Often involves techniques like Reinforcement Learning from Human Feedback (RLHF).
• Crucial for ensuring models are helpful, harmless, and honest.
• Addresses issues like brand voice, tone, and ethical guidelines.

Memory trick: Alignment Achieves Aspirational Aims.

Algorithmic bias refers to systematic and repeatable errors in an AI system that lead to unfair or discriminatory outcomes, often as a result of biased training data, flawed model design, or inappropriate use.

• Can amplify existing societal biases.
• Leads to disparate impact on certain groups.
• Requires careful design, data, and evaluation to mitigate.

Memory trick: S.A.M.E. Bias in AI

The suitability of a foundation model's architecture and pre-training for different types of data, such as text, images, audio, or structured numerical data.

• Most current FMs excel with unstructured data (text, images, audio).
• Direct application to structured tabular data can be less optimal.
• Requires specific adaptation techniques or different models for tabular data.

Memory trick: Unstructured is Usual, Structured is Stumbling.

The computational resources (e.g., GPU memory, processing power) and associated financial cost required to run a pre-trained foundation model to generate predictions or responses.

• Directly correlated with model size (parameter count).
• Impacts deployment feasibility, especially for real-time applications.
• A major consideration for budget-constrained projects.

Memory trick: Cost Counts, Capacity Controls.

A machine learning service that automatically extracts text, handwriting, and data from scanned documents, forms, and tables.

• Extracts text, forms, and tables.
• Supports various document types and layouts.
• Can process handwritten text.

Memory trick: Textract pulls out the exact text from documents.

Cultural Sensitivity in AI refers to designing and deploying AI systems that understand, respect, and adapt to diverse cultural norms, values, and contexts to avoid offense or misinterpretation.

• Goes beyond linguistic accuracy to address social nuances.
• Crucial for global AI applications and user acceptance.
• Requires diverse training data and cultural expertise in development.

Memory trick: AI must be a 'cultural chameleon', adapting to its environment.

A component in neural networks that allows the model to weigh the importance of different parts of the input data when processing a specific element, enabling the capture of long-range dependencies.

• Crucial for Transformers
• Assigns weights to input elements
• Improves contextual understanding

Memory trick: ATTENTION is like a spotlight, highlighting important details across the whole scene.

A technology that enables private connectivity between VPCs and services hosted on AWS, without exposing data to the public internet.

• Keeps network traffic within the AWS network
• Enhances security and compliance for sensitive data
• Supports various AWS services and partner services

Memory trick: PrivateLink: A 'private link' for your VPC to AWS services, no public internet allowed.

The practice of protecting data within AWS AI/ML services by converting it into a coded format, using AWS Key Management Service (KMS) for key management.

• Crucial for data privacy and compliance.
• Covers data at rest (storage) and in transit (network).
• KMS provides centralized key management.
• Customer-managed keys (CMKs) offer greater control.

Memory trick: KMS holds the keys to unlock your data's privacy.

The practice of designing and refining input prompts for foundation models to achieve desired outputs, explore different behaviors, or steer generation towards specific styles or content.

• No model weight changes
• Rapid iteration
• Crucial for leveraging pre-trained FMs

Memory trick: PROMPT Engineering is like giving the AI clear INSTRUCTIONS.

A fully managed service that provides every developer and data scientist with the ability to build, train, and deploy machine learning models quickly.

• Supports popular open-source ML frameworks (TensorFlow, PyTorch).
• Offers managed notebooks, training, and inference.
• Handles infrastructure management (e.g., GPU instances).

Memory trick: SageMaker builds ML models wisely.

A foundation model capable of processing, understanding, and generating data from multiple modalities (e.g., text, images, audio, video) simultaneously, enabling cross-modal tasks.

• Handles diverse data types
• Enables cross-modal understanding
• Used for tasks like image captioning, video generation, complex music

Memory trick: MULTIMODAL is like a conductor leading an entire ORCHESTRA of data types.

An AI-powered coding companion that generates real-time code recommendations in integrated development environments (IDEs).

• Suggests code snippets, functions, and even full files.
• Supports multiple programming languages.
• Integrates with popular IDEs like VS Code, IntelliJ, and AWS Cloud9.

Memory trick: CodeWhisperer helps write code, CodeGuru reviews it.

A fully managed service that allows you to easily analyze streaming data in real time with Apache Flink or SQL.

• Processes data from Kinesis Data Streams or Kinesis Data Firehose.
• Supports real-time anomaly detection and aggregations.
• Can invoke external endpoints for ML inference.

Memory trick: Kinesis Analytics streams ML insights.

An IoT edge runtime and cloud service that helps you build, deploy, and manage intelligent device software.

• Extends AWS capabilities (e.g., Lambda, ML inference) to edge devices.
• Enables local processing, even offline.
• Supports over-the-air (OTA) updates for applications and ML models.

Memory trick: Greengrass brings the cloud to the edge.

A common issue in generative foundation models where the output lacks diversity, producing similar or identical phrases, ideas, or structures despite prompts for variety.

• Can be due to sampling strategies or model architecture
• Reduces perceived creativity and utility
• Addressed through techniques like diverse decoding or fine-tuning

Memory trick: Generative FMs can Repeat, Hallucinate, or Bias.

A fully managed AWS service that provides pre-trained computer vision capabilities to analyze images and videos.

• Offers object, scene, and activity detection.
• Can perform facial analysis and celebrity recognition.
• Requires no machine learning expertise to use.

Memory trick: Rekognize what you see, no ML needed!

A fully managed service that makes foundation models (FMs) from Amazon and leading AI startups available via an API, with capabilities for private customization.

• Offers a choice of FMs including Amazon's Titan models and third-party models.
• Supports customization (fine-tuning) of FMs with proprietary data.
• Simplifies development of generative AI applications without managing infrastructure.

Memory trick: Bedrock is the foundation for creative generative AI, letting you customize your output.

A fully managed machine learning service that helps developers add real-time personalization and recommendation capabilities to applications.

• Uses the same ML technology as Amazon.com.
• Requires historical user behavior data (interactions, items, users).
• No deep ML expertise needed, abstracts away complex model training.

Memory trick: Personalize your shopping experience with recommendations.

A fully managed service for delivering real-time streaming data to destinations like Amazon S3, Amazon Redshift, Amazon OpenSearch Service, and Splunk.

• No servers to manage
• Automatically scales to match data throughput
• Integrates with other AWS services for analytics and storage

Memory trick: Kinesis Firehose: A rapid river of data for your machine's brain.

Unexpected capabilities that appear in large-scale foundation models as they increase in size (parameters) and are trained on more data, often without direct explicit training for those specific tasks.

• Not present in smaller versions of the same model
• Can include multi-step reasoning, arithmetic, translation
• A key characteristic distinguishing FMs from traditional ML models

Memory trick: FMs have Scale, Generalize, and Emerge.

Overfitting occurs when a machine learning model learns the training data too well, including its noise and specific details, leading to poor performance on new, unseen data.

• High training accuracy, low validation/test accuracy.
• Model is too complex for the data.
• Common in models with many parameters, like FMs.

Memory trick: Models can be too specific or too general.

Foundation models are highly adaptable due to their pre-training on massive datasets, allowing them to perform a wide range of tasks with minimal task-specific fine-tuning.

• Broad applicability across tasks
• Leverages vast pre-training data
• Reduces need for extensive task-specific data

Memory trick: Foundation models are like a SWISS ARMY KNIFE for AI tasks.

An AWS natural language processing (NLP) service that uses machine learning to find insights and relationships in text.

• Performs sentiment analysis
• Identifies entities (people, places, organizations)
• Extracts key phrases and detects language

Memory trick: Comprehend: Understand the 'compre'hensive meaning of text.

The practice of choosing a specific AWS Region for deploying resources to ensure data storage and processing occurs within defined geographic boundaries.

• Each AWS Region is a distinct geographical area.
• Data stored in a Region stays in that Region unless explicitly moved.
• Crucial for compliance with data residency laws (e.g., GDPR).

Memory trick: Region choice keeps data where it belongs.

Large Language Models are a class of foundation models trained on massive text datasets, capable of understanding, generating, and processing human language for various NLP tasks.

• Trained on vast text corpora.
• Excels at tasks like text generation, summarization, translation.
• Forms the backbone of many AI applications involving language.

Memory trick: Models are like specialists for different data types.

AI interpretability (also known as explainability) is the degree to which a human can understand the cause and effect of an AI system's decisions, especially for complex models.

• Crucial for building trust and accountability.
• Helps justify decisions, particularly in high-stakes domains.
• Can involve techniques like LIME, SHAP, or decision trees.

Memory trick: Explain the AI's thought process.

PEFT refers to a collection of techniques that enable efficient adaptation of large pre-trained foundation models to downstream tasks by fine-tuning only a small number of additional parameters, rather than all of them.

• Reduces computational cost and memory usage.
• Mitigates catastrophic forgetting of pre-trained knowledge.
• Effective with limited task-specific data.

Memory trick: Small changes, big impact, save resources.

Encryption keys created and managed by the customer within AWS Key Management Service (KMS), providing full control over key lifecycle and permissions.

• Customers define key policies and grant permissions.
• Provides an audit trail of key usage.
• Used with AWS services for data encryption at rest.

Memory trick: CMKs give YOU the key, not AWS.

A component of Amazon SageMaker that provides a fully managed service for training machine learning models.

• Supports various instance types, including GPU-accelerated.
• Automatically provisions and de-provisions resources.
• Allows use of custom Docker images for training environments.

Memory trick: SageMaker trains smart, saves costs.

A natural language processing (NLP) service that uses machine learning to extract health data from unstructured medical text.

• Identifies medical conditions, medications, treatments, and Protected Health Information (PHI).
• HIPAA-eligible service.
• Optimized for clinical notes, discharge summaries, and patient records.

Memory trick: Comprehend Medical understands health words.

A technology that allows you to privately connect your VPC to AWS services and VPC endpoint services powered by PrivateLink, without using an internet gateway, NAT device, VPN connection, or AWS Direct Connect.

• Keeps traffic entirely within the AWS network.
• Enhances security and compliance (e.g., HIPAA, GDPR).
• Uses private IP addresses for service access.
• Crucial for scenarios requiring strict network isolation.

Memory trick: PrivateLink keeps data private, never touching the public net.

A neural machine translation service that provides fast, high-quality, and affordable language translation.

• Supports many languages.
• Offers real-time and batch translation.
• Can be integrated with other AWS AI services.

Memory trick: Translate bridges language gaps, making words global.

Deep semantic understanding in foundation models refers to their ability, gained through extensive pre-training, to grasp the meaning, context, and relationships within language or other data modalities beyond surface-level patterns.

• Acquired from massive pre-training datasets.
• Enables nuanced interpretation of text, images, etc.
• Facilitates complex reasoning and generation tasks.

Memory trick: Big data makes models smart and adaptable.

The practice of encoding sensitive data to prevent unauthorized access, both when stored (at rest) and when being transmitted across networks (in transit).

• Data at rest: Encrypted using services like AWS KMS with S3, EBS, RDS, etc.
• Data in transit: Encrypted using Transport Layer Security (TLS) for network communication.
• Crucial for compliance (e.g., HIPAA, GDPR) and protecting intellectual property.

Memory trick: Encrypt data to keep it private, whether it's sitting still or moving around.

An AI framework that combines a retrieval system with a generative foundation model to improve the accuracy and factual grounding of generated responses by retrieving relevant information from an external knowledge base.

• Reduces hallucinations
• Enables source citation
• Handles dynamic knowledge bases

Memory trick: RAG is like having a SMART RESEARCH ASSISTANT for your AI.

RLHF is a technique used to align large language models with human preferences and safety guidelines by training a reward model on human feedback and then using reinforcement learning to optimize the language model.

• Uses human preferences to create a reward signal.
• Enables models to follow complex instructions and avoid harmful outputs.
• Key for safety and helpfulness alignment in FMs.

Memory trick: Align models with human values through feedback.

An AWS text-to-speech (TTS) service that converts text into lifelike speech.

• Supports multiple languages and a wide selection of voices.
• Offers Neural Text-to-Speech (NTTS) for highly natural-sounding voices.
• Allows creation of custom Brand Voices for unique identity.

Memory trick: Polly speaks, Transcribe hears.

A machine learning service that uses data from industrial sensors to detect abnormal equipment behavior and predict future machine failures.

• Specifically designed for industrial predictive maintenance.
• Requires no machine learning expertise.
• Automatically builds, trains, and deploys models from sensor data.

Memory trick: Lookout for Equipment sees future failures.

A networking service that provides private connectivity between VPCs and AWS services, or between VPCs and on-premises networks, without exposing data to the public internet.

• Uses interface VPC endpoints.
• Traffic remains within the Amazon network.
• Enhances security and compliance for sensitive data.

Memory trick: SageMaker's VPC Link keeps fraud models safe.

Transfer learning with foundation models involves taking a model pre-trained on a large, general dataset and adapting it to a new, specific task with a smaller, target dataset.

• Leverages knowledge from massive pre-training.
• Reduces need for large task-specific datasets.
• Accelerates model development for new tasks.

Memory trick: General models learn, then transfer knowledge.