AI Practitioner Exam

Inferencing at the Edge

- Process of performing machine learning inference directly on edge devices rather than relying on centralized cloud computing. - Low latency. - Bandwidth Efficiency. - Edge devices are usually devices with less computing power that are close to where the data is generated, in places where internet connections can be limited.

SageMaker - Model Deployment: Batch

- Processes multiple predictions at once, allowing for efficient handling of large volumes of data. - Ideal for scenarios like generating predictions for an entire dataset or periodically processing data. - Prediction for an entire dataset (multiple predictions). Request and responses are in S3. High latency.

AWS AI Service Cards

- Provide important information on the ethical considerations, transparency, and intended usage of AWS AI services. - Form of responsible AI documentation. - Help understand the service and its features. - Find intended use cases and limitations. - Responsible AI documentation that provides customers with a single source of information on the intended use cases and constraints, responsible AI design choices, and deployment and performance optimization best practices for AWS AI services.

Continued Pre-training

- Provide unlabeled data to continue the training of an FM. - A process of further training a pre-trained model on a new, relevant dataset. - is a method that provides unlabeled data to a foundation model (FM) so the model can train on a specific domain or topic.BLU

Prompt Engineering Technique: Few-Shot Prompting

- Provides a few examples of the desired output along with the prompt. Helps the model understand the context and expected format, improving response accuracy. Effective for tasks that benefit from specific guidance or examples.

SageMaker Clarify - Model Explainability

- Provides insights into how machine learning models make predictions by explaining the influence of features on the output. - A set of tools to help explain how machine learning (ML) models make predictions. - Helps increase the trust and understanding of the model.

Explainability

- Providing clear explanations and interpretability of how AI systems make decisions, enabling accountability and trust. - Understand the nature and behavior of the model. - Being able to look at the inputs and outputs and explain without understanding exactly how the model came to the conclusion.

SageMaker Ground Truth

- RLHF (Reinforcement Learning From Human Feedback). - A fully managed data labeling service that helps users build high-quality labeled datasets for machine learning. - Supports the use of human labelers for various data types (images, text, videos) to ensure accurate annotations. - Incorporates machine learning algorithms to pre-label data, which can then be refined by human labelers, speeding up the labeling process.

Balanced

- Refers to a dataset where the classes or categories are represented in approximately equal proportions. - Improves model performance and reduces bias towards the majority class. - Enhances the ability to accurately predict minority class instances. - Neither overfitting or underfitting (low bias and low variance).

Multi-Model Model

- Refers to architectures that combine multiple models or modalities (e.g., text, images, audio) to perform complex tasks. - Enhances performance by leveraging the strengths of different models, allowing for richer understanding and generation of content. - Used in applications like multimodal chatbots, where understanding and generating responses involve integrating information from various sources. - Does NOT rely on a single type of input (text, image, audio).

Transparency

- Refers to the ability to understand and inspect the inner workings, decision-making processes, and outputs of an AI system. - It involves making the AI system's behavior, decisions, and underlying logic visible and comprehensible to relevant stakeholders, such as developers, regulators, and end-users.

Privacy and Security

- Safeguarding the privacy and security of the data used by AI systems and handling personal information responsibly.

BertScore

- Semantic similarity between text. - Compare the actual meaning of the text and see if the meaning is similar. - Looking at context and nuance.

Hyperparameters

- Settings or configurations that govern the training process of a machine learning model but are not learned from the data. - Learning Rate: How fast a model can incorporate new data. Determines the step size at each iteration while moving toward a minimum of the loss function. - Batch Size: How many data points to consider at a time. The number of training examples used in one iteration of model training. - Number of Epochs: How many times you want to iterate on your model until you say you've converged to a good result. The number of complete passes through the training dataset during training. - Regularization Parameters: See how flexible the model should be. Settings like L1 or L2 regularization that control overfitting. - Hyperparameter Tuning: finding the best hyperparameter values to optimize the model performance.

BERT (Bidirectional Encoder Representations from Transformers)

- Similar intent to GPT, but reads the text in two directions.

Stop Sequences

- Specific sequences or tokens that signal the model to stop generating further output, allowing for controlled termination of responses. - Tokens that signal the model to stop generating output.

Length

- Specifies the maximum number of tokens in the generated output, controlling the response length. - Maximum length of answer.

Bias

- The error introduced by approximating a real-world problem with a simplified model. - Difference or between predicated and actual value.

Variance

- The error introduced by the model's sensitivity to fluctuations in the training dataset. - How much the performance of a model changes if trained on a different dataset which has a similar distribution.

Prompt Engineering

- The practice of designing and refining input prompts to optimize the performance of AI models, especially in natural language processing. - Involves crafting specific questions or instructions to elicit desired responses from generative models. - Key techniques include experimenting with phrasing, context provision, and example-driven prompts. - Developing, designing, and optimizing prompts to enhance the output of the FMs of your needs. - Improved prompting techniques consists of: Instructions, Context, Input data, Output indicator.

Tokenization

- The process of breaking down text into smaller units called tokens (e.g., words, phrases, or characters). - Essential step in natural language processing (NLP) to prepare text for analysis.

Watermark Detection

- The process of identifying and extracting watermarks embedded in digital media (images, videos, audio). - Used to verify the authenticity and ownership of content or to prevent unauthorized use. - Can be visible (e.g., logos, text) or invisible (embedded within the data). - Check if an image was generated by Amazon Titan.

Feature Engineering

- The process of selecting, modifying, or creating new features (variables) from raw data to improve model performance. - Transformation: Scaling, normalization, or encoding categorical variables. - Creation: Generating new features based on existing data (e.g., combining features or creating interaction terms). - Selection: Identifying and retaining the most relevant features while discarding irrelevant or redundant ones. - Essential for enhancing model accuracy, interpretability, and efficiency. - Plays a crucial role in both supervised and unsupervised learning. - The process of using domain knowledge to select and transform raw data into meaningful features. - Helps enhancing the performance of machine learning models.

Context Window

- The range of tokens (words or characters) that a model considers when processing text. - The number of tokens an LLM can consider when generating text.

Training Set

- The subset of data used to train a machine learning model. - The model learns patterns and relationships from this data, adjusting its parameters accordingly. - Typically contains the majority of the available labeled data. - Used to train the model. - Typically 60-80% of the dataset.

Regression

- Used to predict a numerical value based on input data. - The output variable is continuous meaning it can take any value within a range (whats the weight and height of a person). - A type of supervised learning task that predicts continuous numerical values. - Examples include predicting house prices, stock prices, or temperature forecasts. - Models aim to minimize the difference between predicted and actual values.

Classification

- Used to predict the categorical label of input data. The output variable is discrete, which means it falls into a specific category or class. (E.g. dog and cat vs giraffe and spam vs not spam). - A type of supervised learning task that predicts categorical labels or classes. - Examples include spam detection (spam vs. not spam), image recognition (cat vs. dog), and sentiment analysis (positive vs. negative). - Models output probabilities for each class, often using thresholds to determine the final prediction.

Instruction-Based Fine Tuning

- Uses labeled examples that are prompt responses pair. - A method to customize AI models by training them on specific instructions or prompts.

Automatic Evaluation

- Utilizes metrics and algorithms to assess model outputs. - Common metrics include BLEU, ROUGE, and perplexity. - Provides quick feedback on model performance.

Supervised Learning

- Where the model is trained to map inputs to known outputs. - Learn a mapping function that can predict the output for new unseen input data. - A machine learning paradigm where models are trained on labeled data to learn the mapping from inputs to outputs. - Common tasks include regression and classification. - The goal is to make accurate predictions on unseen data based on learned patterns.

F1 Score

- You can use this to evaluate a model's accuracy for binary classification. - Use precision and recall to evaluate how accurate a model correctly classifies the correct class.

SageMaker - Automatic Model Tuning (AMT)

- feature that automatically searches for the best hyperparameters for machine learning models.

Bilingual Evaluation Understudy (BLEU)

- is a metric used to evaluate the quality of text generated by machine translation systems, but it's also applicable to other text generation tasks. - Especially for translations. - Evaluate the quality of generated text. - Primarily measures precision. - Translation tasks.

Polly

-Turn text into lifelike speech using deep learning. - Allowing you to create applications that talk. - Lexicons: define how to read certain pieces of text (AWS => "Amazon Web Services") - SSML: how your text should be pronounced. Markup for your text to indicate how to pronounce it. (E.g., "Hello, <break> how are you?") - Voice Engine: generative, long form, neutral, standard - Speech mark: encode where a sentence/word starts or ends in audio.

ResNet (Residual Network)

- (Image) used for image recognition tasks, object detection, facial recognition.

WaveNet

- (audio) model to generate raw audio waveform, used in speech synthesis.

GAN (Generative Adverbial Network)

- (data augmentation). Models used to generate synthetic data such as images, videos, or sounds that resemble the training data. Helpful for data augmentation.

Feature Store

- A centralized repository for storing, managing, and serving machine learning features for model training and inference. - Ingests features from a variety of sources. Ability to define the transformation of data into feature from within Feature Store. - Can publish directly from SageMaker Data Wrangler into SageMaker Feature Store. - Features are discoverable within SageMaker Studio.

Multi-Turn Messaging

- A conversational model interaction involving multiple exchanges between user and model. - Context from previous turns is maintained to facilitate a more natural conversation. - Enables more complex dialogues and nuanced understanding. - Useful for applications like chatbots and virtual assistants that require context.

Single-Turn Messaging

- A conversational model interaction where the user provides a single input and receives one output. - No context is carried over from previous exchanges. - Often used in straightforward question-and-answer scenarios. - Simpler implementation with less computational overhead. - Part of instruction-based fine-tuning. System, messages, role, content. - Alternate between user and assistant to provide more context.

Model Parallelism

- A feature designed to help train large deep-learning models that cannot fit into the memory of a single GPU. - This feature automatically partitions the model across multiple GPUs, efficiently training very large models. - This distribution helps reduce the training time and allows for handling models that are otherwise too large to be trained on a single GPU.

Party Rock

- A feature or initiative within the Amazon Q ecosystem that emphasizes collaborative and engaging user experiences. - May refer to specific tools or components designed to enhance social interaction in applications.

Bedrock - CloudWatch Integration

- A feature that allows monitoring and logging of Bedrock applications. - Enables tracking of key performance metrics, such as model response times, usage patterns, and error rates. - Helps developers set up alerts and dashboards for real-time insights into application performance. - Facilitates troubleshooting by providing logs and metrics that can identify issues or bottlenecks. - Supports maintaining operational health and optimizing resource utilization in AI applications.

Semi-Supervised Learning

- A machine learning approach that combines a small amount of labeled data with a large amount of unlabeled data. - Aims to improve learning accuracy when acquiring labeled data is expensive or time-consuming. - Utilizes the strengths of both supervised and unsupervised learning methods. - Use a small amount of labeled data and a large amount of unlabeled data to train systems. After that, the partially trained algorithm itself labels the unlabled data.

Reinforcement Learning from Human Feedback (RLHF)

- A machine learning approach that combines reinforcement learning with human feedback to improve model performance. - Use human feedback to help ML models to self-learn more efficiently. - Involves training agents to make decisions based not only on predefined rewards but also on feedback provided by human evaluators. - Human Feedback: Insights or corrections given by humans, often used to guide the learning process. - Reward Model: A model trained to predict human preferences, which helps the agent learn more effectively.

Unsupervised Learning

- A machine learning paradigm where models are trained on unlabeled data to identify hidden patterns or groupings. - Does not require labeled outputs, focusing instead on discovering inherent structures within the data. - The goal is to discover inherent patterns, structures, or relationships within the input data. The machine must uncover and create the groups itself, but humans still put labels on the output groups.

Transfer Learning

- A machine learning technique where a model developed for one task is reused for another related task. - Reduces the need for extensive data and computational resources for training from scratch. - Widely used for image classification. - The broader concept of reusing a pre-trained model to adapt it to a new related task.

Amazon Q

- A platform designed for building and deploying conversational AI applications. - Facilitates the creation of chatbots and virtual assistants that leverage AWS services.

Test Set

- A separate subset of data used to evaluate the final performance of a trained model. - Provides an unbiased assessment of how well the model generalizes to unseen data. - Should not be used during the training or validation phases to ensure a fair evaluation. - Used to evaluate final model performance. - Typically 10-20% of the dataset.

Vector Database

- A specialized database designed to store and manage vector embeddings. - Optimized for efficient similarity search and retrieval of high-dimensional data.

Validation Set

- A subset of data used to tune model hyperparameters and evaluate model performance during training. - Helps in selecting the best model configuration and prevents overfitting by providing feedback on how well the model generalizes. - Typically smaller than the training set and not used for training the model directly. - Used to tune model parameters and validate performance. - Typically 10-20% of the dataset.

Negative Prompting

- A technique used in prompt engineering to specify undesirable outputs or behaviors to avoid. - Involves instructing the model on what not to generate, guiding it away from unwanted content or responses. - Helps refine the quality of model outputs by minimizing irrelevant, biased, or harmful results. - A technique where you explicitly instruct the model on what not to include or do in its response (avoid unwanted content, maintain focus, enhance clarity).

Association Rule Learning

- A technique used to discover interesting relationships or associations between variables in large datasets. - Commonly used in market basket analysis to identify products that frequently co-occur in transactions (e.g., customers who buy bread also buy butter). - Utilizes metrics like support, confidence, and lift to evaluate the strength of associations. - E.g., which products are frequently bought together (bread and butter).

Reinforcement Learning

- A type of Machine Learning where an agent learns to make decisions by performing actions in an environment to maximize cumulative reward. - Operates on the principle of trial and error, where the agent takes actions, receives feedback (rewards or penalties), and adjusts its strategy accordingly.

Embedding Model

- A type of model designed to convert categorical data (like words, sentences, or images) into vector embeddings. - Transforms complex data into a numerical format that captures semantic meaning and relationships. - Enable efficient similarity search and retrieval in large datasets.

Clustering

- A type of unsupervised learning that groups similar data points into clusters based on defined similarity measures. - Examples include k-means clustering, hierarchical clustering, and DBSCAN. - Useful in applications like customer segmentation, image grouping, and anomaly detection.

Transformer Model (LLM)

- Able to process a sentence as a whole instead of word by word. Fast and more efficient text processing (less training time). A type of deep learning architecture designed for handling sequential data, primarily used in natural language processing (NLP). - Utilizes self-attention mechanisms to weigh the importance of different words in a sequence, allowing for better context understanding. - Forms the basis for many large language models (LLMs) like BERT, GPT, and T5, enabling tasks such as text generation, translation, and summarization.

Reinforcement Learning Components

- Agent: The learner or decision-maker. - Environment: The context within which the agent operates. - Actions: The choices available to the agent. - Rewards: Feedback signals that guide learning, indicating the success of actions.

SageMaker - Model Deployment: Serverless

- Allows for automatic scaling and management of inference resources without the need for server provisioning. - Ideal for applications with variable traffic patterns, as users only pay for the actual compute resources consumed during inference. - Simplifies deployment, especially for infrequent or unpredictable workloads. - Idle period between traffic spikes. Can tolerate more latency (cold starts). Low latency. No infrastructure.

Top P (0 to 1)

- Also known as nucleus sampling. - Determines the cumulative probability threshold for token selection. The model considers the smallest set of tokens whose probabilities sum to P (e.g., 0.9) for generating the next token. - Low _ consider the 25% most likely words, will make a more coherent response. - High _ consider a broad range of possible words, possibly more creative and diverse output.

Vector Search

- Amazon OpenSearch Service - Amazon Aurora PostgreSQL-Compatible Edition - Amazon Relational Database Service (Amazon RDS) - Amazon Neptune ML - Amazon MemoryDB - Amazon DocumentDB (with MongoDB compatibility)

XGBoost (Extreme Gradient Boosting)

- An implementation of gradient boosting.

Bedrock Studio

- An integrated development environment (IDE) for building and deploying generative AI applications using AWS Bedrock. - Provides tools for designing, testing, and iterating on models and applications. - Features a user-friendly interface that simplifies the process of connecting to foundational models and managing data.

Amazon Q Developer

- Answer questions about the AWS documentation and AWS service selection. - Answer questions about resources in you AWS account. - AI code companion to help you code new applications. Supports many languages. - Real-time code suggestions and security scans. - Integrates with IDE to help with software development needs. - Refers to developers who create and manage applications on the Amazon Q platform. - Focuses on integrating AI capabilities and ensuring optimal performance of conversational applications.

Large language models (LLMs)

- Are extensive deep learning models pre-trained on massive datasets. They utilize a transformer architecture, which includes neural networks composed of an encoder and a decoder with self-attention mechanisms. These components work together to derive meaning from text sequences and comprehend the relationships between words and phrases.

SageMaker Model Cards

- Are structured documents that provide essential information about machine learning models. - Capture key details about model performance, training data, hyperparameters, and intended use cases. - Essential model information. - Example: intended uses, risk ratings, and training details

Transcribe Medical

- Automatically convert medical-related speech to text (HIPPA Compliant)

Transcribe

- Automatically convert speech into text. - Automatically remove PII, language identification for multi-lingual audio. - Custom Vocabularies (for words): add specific words, phrases, domain-specific terms. - Custom Language Models (for context): train transcribe model on your own domain-specific text data.

Textract

- Automatically detects and extracts text, handwriting, and data from any scanned documents using AI and ML. - Fully managed service that can detect and extract text and data from scanned documents, PDFs, and images.

SageMaker Clarify - Detect Bias

- Automatically identifies and measures bias in datasets and model predictions to ensure fairness. - Ability to detect and explain biases in your datasets and models.

Bedrock - Fine Tuning

- Bring your own data and update foundation model with your own data to make sure its more adapted to your use case. - Process of customizing pre-trained models for specific tasks or datasets. - Training data must adhere to a specific format and be stored in S3. - You must use "provisioned throughput" pricing model.

Lex

- Build chatbots quickly for your applications using voice and text. - Integration with lambda, connect, comprehend, Kendra.

Phases of Machine Learning Project

- Business Problem - ML Problem Framing - Data Collection and Preparation - Feature Engineering - Model Training and Parameter Tuning - Model Evaluation - Model Testing and Deployment - Monitoring and Debugging

Bedrock - GuardRails

- Control the interaction between users and Foundation Models (FM). - Filter undesirable and harmful contents. - Remove PII. - Enhance privacy and reduce hallucinations. - A feature designed to enhance the safety and reliability of generative AI applications. - Provides mechanisms to set boundaries and guidelines for model behavior. - Helps prevent the generation of inappropriate or harmful content. - Enables developers to define rules for acceptable outputs based on business requirements. - Can be used to enforce compliance and ethical standards in AI applications.

Temperature (0 to 1)

- Controls randomness in model outputs. Lower values (e.g., 0.2) yield more deterministic responses, while higher values (e.g., 0.8) produce more varied outputs. - Creativity of the models output. - Low outputs are more conservative, repetitive, focused on most likely response. - High outputs are more diverse, creative, and unpredictable, maybe less coherent.

Amazon Q Apps (Q Business)

- Create Gen AI-powered apps without coding by using natural language processing. - Applications built on the Amazon Q platform aimed at business use cases. - Includes tools for customer support, sales assistance, and other business-oriented tasks.

Mechincal Turk

- Crowdsourcing marketplace to perform simple human tasks. - Distributed virtual workforce. - Humans around the world will work for a reward (E.g., maybe 10 cents to label data, or image classification, data collection, etc).

Labeled Data

- Data includes both input features and corresponding output labels. - Data that is annotated with meaningful labels or tags indicating the correct output. - Essential for training supervised learning models, as it provides the ground truth for predictions.

Unlabeled Data

- Data includes only input features without any output labels. - Data that does not have any associated labels or annotations indicating the correct output. - Commonly used in unsupervised learning tasks where the goal is to find patterns or structures in the data without prior knowledge of the outcomes.

Structured Data

- Data is organized in a structured format, often in rows and columns. Tabular data. - Data that is organized and formatted in a predefined manner, typically in rows and columns (e.g., databases, spreadsheets). - Easily searchable and analyzable using traditional data processing tools. - Examples include customer information in a CRM system, transaction records, and sensor data.

Unstructured Data

- Data that doesn't follow a specific structure and is often text-heavy or multimedia content. - Data that does not have a predefined structure or format, making it more challenging to analyze. - Often includes text, images, audio, and video, which require specialized processing techniques. - Examples include social media posts, emails, photographs, and web pages.

Foundation Model

- Describe machine learning models that are trained on a diverse range of generalized and unlabeled data. These models are capable of performing a wide array of tasks, including language comprehension, text and image generation, and natural language conversation.

Comprehend Medical

- Detects and returns useful information in unstructured clinical text. - Uses NLP to detect protected health information (PHI).

Amazons Hardware for AI

- EC2 Instance types (GPU Based Types): - AWS Trainium: ML chip built to perform Deep Learning on 100B+ parameter models. - AWS Inferentia: ML chip built to deliver inference at high performance and low cost.

SageMaker - Model Deployment: Real-time

- Enables immediate predictions by deploying models as endpoints. - Suitable for applications requiring instant responses, such as chatbots or online recommendations. - Offers low-latency inference, but may involve costs for idle endpoints. - One prediction at a time. Low latency.

Prompt Engineering Technique: Chain of Thought Prompting

- Encourages the model to articulate its reasoning process step-by-step before arriving at a final answer. - Aims to improve reasoning and problem-solving capabilities by breaking down complex tasks. - Useful in tasks requiring logical deduction or multi-step reasoning.

Fairness

- Ensuring AI models are unbiased and do not discriminate against individuals or groups based on protected characteristics. - Promote inclusion and prevent discrimination.

Veracity and Robustness

- Ensuring AI systems operate reliably and consistently and are resilient to potential failures or adversarial attacks. - Reliable even in unexpected situations.

Governance

- Establishing clear governance frameworks, policies, and processes for the responsible development and deployment of AI systems.

RAG (Retrieval-Augmented Generation)

- Fetching data from external data sources. - Combines retrieval of relevant information with generative models. - Enhances response quality by using external data sources.

Rekognition

- Find objects, people, text, scenes in images and videos using ML. - Facial analysis and facial search to do user verification.

Comprehend

- For Natural Language Processing (NLP). Uses machine learning to find insights and relationships in text. - You cannot use this to process textual information from images that are provided in PNG format. Requires only text as input.

Personalize

- Fully managed ML service to build apps with real-time personalized recommendations. - E.g., personalized product recommendations. - Same technology used by Amazon.com to recommend products. - Recipes: algorithms the are prepared for specific use cases. - Recipes and personalize are for recommendations.

Kendra

- Fully managed document search service powered by machine learning. - Extract answers from within a document. - Natural language search capabilities.

SageMaker

- Fully managed machine learning service provided by AWS that enables developers and data scientists to build, train, and deploy machine learning models quickly and easily. - Example: predicting your AWS exam score. - Collect and prepare data, build and train machine learning models, deploy the models and monitor the performance of the predictions.

Bedrock

- Fully managed service for building generative AI applications. - Provides access to various pre-trained foundational models. - Allows customization using your own data. - Primarily designed for accessing and customizing pre-trained foundation models.

Forecast

- Fully managed service that uses ML to deliver highly accurate forecasts.

GPT (Generate Pertained Transformer)

- Generate human text or computer code based on input prompts.

SageMaker - Model Deployment: Asynchronous

- Handles predictions where immediate responses are not necessary. - Submits requests for inference that are processed in the background, with results available later. - Useful for large datasets or complex models that require longer processing times. - For large payload sizes up to 1GB. Long processing times. Near real time latency requirements. Request and responses are in S3. Medium to high latency. Near real-time. For one record.

DeepRacer

- Helps you learn reinforcement learning. - Fully autonomous 1/18th scale car race driven by Reinforcement Learning (RL).

Augmented AI (A2I)

- Human oversight of Machine Learning predictions in production. - Low confidence predictions are sent for human review. Reviews are consolidated using weighted scores. - Can be your own employees, or contractors from AWS, or AWS Mechanical Turk.

Saftey

- Identifying and mitigating potential risks and unintended consequences associated with AI systems.

Techniques to prevent overfitting

- Increase the training data size. - Early stopping the training of the model. - Data augmentation (to increase diversity in the data set). - Adjust hyperparameters

Techniques to prevent underfitting

- Increasing Model Complexity: Using a more complex algorithm or model architecture. - Feature Engineering: Adding relevant features or transforming existing ones to better capture data relationships. - Reducing Regularization: Allowing more freedom in the model to fit the training data.

ML Features

- Individual measurable properties or characteristics used as input for machine learning models. - Features are inputs to ML models used during training and used for inference. Important to have high quality features across your datasets in your company for reuse.

Bedrock - Agents

- Intelligent components that leverage foundational models to perform specific tasks or functions. - Capable of interacting with users or systems to provide automated responses or actions. - Can utilize retrieval-augmented generation (RAG) techniques to enhance decision-making by accessing external data. - Instead of asking questions to model now the model can start thinking and perform multi step tasks.

Prompt Engineering Technique: Zero-Shot Prompting

- Involves asking the model to perform a task without any prior examples. - Relies on the model's general knowledge and understanding to generate responses. - Useful for tasks where no specific context or training examples are provided.

Intelligent Document Processing (IDP)

- Involves automating the process of manually entering data from paper-based documents or document images to integrate with other digital business processes. - It extracts, categorizes, and generates insights from unstructured data. - Its ready-to-use APIs efficiently process unstructured data at scale, extract critical information, and generate insightful summaries and reports.

Human Evaluation

- Involves human reviewers assessing the quality of model outputs. - Evaluates aspects like relevance, coherence, and overall quality. - Offers nuanced insights that automated metrics may miss.

Inferencing - Real-time

- Involves making predictions instantly or with minimal delay as new data arrives. - Typically used in applications requiring immediate responses, such as chatbots, fraud detection, and online recommendations. - Requires efficient processing and low latency to handle incoming requests in real time. - Computers have to make decisions quickly as data arrives. Send prompt with immediate response. Speed is preferred over perfect accuracy.

Inferencing - Batch

- Involves processing a large volume of data at once, rather than making predictions on individual instances. - Often used for tasks where immediate results are not critical, such as generating reports, analyzing historical data, or periodic predictions. - Can be more resource-efficient, as it allows for optimizing compute resources and can take advantage of parallel processing. - Large amount of data that is analyzed all at once. Give large amounts of data and get it when it's ready. Speed of the results is usually not a concern, and accuracy is.

SageMaker - Model Registry

- Is a feature that helps manage and organize machine learning models throughout their lifecycle. - Supports tracking multiple versions of models, allowing users to manage updates and rollbacks efficiently. - Enhances collaboration among data science teams by providing a centralized repository for model assets. - Monitor the quality of your model in production: continuous or on-schedule.

SageMaker Jumpstart

- Is a feature that provides quick access to a variety of pre-built machine learning models and solutions. - Offers a library of pre-trained models for common use cases, such as image classification, text analysis, and time series forecasting. - ML Hub to find pre-trained Foundation Model (FM), computer vision models, or natural language processing models. - Pre-built ML solutions for demand forecasting, credit rate prediction, fraud detection, and computer vision.

Stable Diffusion

- Is a generative AI model that creates distinctive photorealistic images based on text and image prompts.

Recall-Oriented Understudy for Gisting Evaluation (ROUGE)

- Is a metric that you can use to evaluate the quality of text summarization and text generation. - You can use this to assess the performance of an FM for text generation. - Primarily measures recall. - Summarization tasks.

SageMaker - Model Monitor

- Is a service that enables continuous monitoring of machine learning models in production. - Centralized portal where you can view, search, and explore all your models. - View all your models in one place.

SageMaker Pipelines

- Is a service that enables the creation, automation, and management of end-to-end machine learning workflows. - Allows users to define and orchestrate steps in the ML workflow, such as data preparation, model training, and deployment. - Centralized repository allows you to track, manage, and version ML models. - Manage approval status of a model, automate model deployment, share models. - Centralized repository to manage ML model versions.

SageMaker Clarify

- Is a tool designed to help detect bias in machine learning models and improve transparency in AI. - Evaluate foundation models. Evaluating human factors such as friendliness or humor. - Leverage an AWS-managed team or bring your own employees.

Data Wrangler

- Is a tool designed to simplify and accelerate the data preparation process for machine learning. - Prepare tabular and image data for machine learning. - Data preparation, transformation and feature engineering. - Single interface for data selection, cleansing, exploration, visualization, and processing. - Data quality tools.

SageMaker Role Manager

- Is a tool that simplifies the management of IAM roles for Amazon SageMaker. - Define roles for personas. Example: Data Scientists, MLOps Engineers

SageMaker Model Dashboard

- Is a visual interface for monitoring and managing deployed machine learning models. - Provides real-time insights into model performance metrics such as latency, accuracy, and throughput. - Allows users to set up alerts for performance degradation or anomalies in model behavior. - Centralized repository. - Information and insights for all models.

SageMaker Canvas

- Is a visual interface that enables business analysts and data scientists to build machine learning models without writing code. - Provides a user-friendly drag-and-drop interface for data preparation, model building, and evaluation. - Build ML models using a visual interface (no coding required). - No code interface. - Ready-to-use models.

SageMaker Studio

- Is an integrated development environment (IDE) for machine learning provided by AWS. - End-to-end ML development from a unified interface. - Team collaboration. - Tune and debug ML models. - Deploy ML models. - Automated workflows.

MLFlow on Amazon SageMaker

- Is an open-source platform for managing the machine learning lifecycle, integrated with Amazon SageMaker. - Allows users to log and track experiments, capturing parameters, metrics, and artifacts associated with model training. - An open-source tool which helps ML teams manage the entire ML lifecycle.

Top K

- Limits the number of highest probability tokens from which the model can sample. For example, with Top _ set to 50, only the top 50 tokens are considered for selection. - Limits the number of probable words. - Low _ more coherent response, less probable words. - High _ more probable words, more diverse and creative.

SVM (Support Vector Machine)

- ML algorithm for classification and regression.

Controllability

- Maintaining appropriate human control and oversight over AI systems, particularly in high-stakes decision-making scenarios.

MlOps

- Make sure models aren't just developed but also deployed, monitored, retrained systematically and repeatedly.

RNN (Recurrent Neural Network)

- Meant for sequential data such as time-series or text, useful in speech recognition, time series prediction.

Multimodal

- Models are AI systems that can process and generate content across multiple modalities, such as text, images, and audio. T - these models are designed to understand and integrate information from diverse data types, enabling more comprehensive and contextually rich outputs.

Bedrock - Batch

- Multiple predictions are processed at once; responses are received later than in real-time. - Responses are received later than real-time but come at a discounted rate. - Necessary for fine-tuned or custom models, ensuring predictable performance and capacity.

Translate

- Natural and accurate language translation.

Underfitting

- Occurs when a machine learning model is too simple to capture the underlying patterns in the data. - Model performs poorly on training data. Could be a problem of having a model too simple or poor data features. - Poor performance on both the training set and unseen data (test set). - High bias, meaning the model makes strong assumptions about the data that are not true.

Overfitting

- Occurs when a machine learning model learns the training data too well, capturing noise and outliers instead of the underlying patterns. - Performs well on training data but doesn't perform well the evaluation data. - High accuracy on the training set but poor performance on unseen data (test set). - The model becomes overly complex, with too many parameters relative to the amount of training data.

Bedrock - On-Demand

- Pay-as-you-go model. - Ideal for variable workloads and testing. - Charges based on usage without long-term commitments.

Bedrock - Provisioned Throughput

- Pre-purchase model units for a specified duration (e.g., 1 month, 6 months). - Necessary for fine-tuned or custom models, ensuring predictable performance and capacity.

Model Evaluation Metrics

- Precision: best when false positives are costly - Recall: best when false negatives are costly - F1: best when you want a balance between precision and recall, especially in imbalanced datasets - Accuracy: best for balanced datasets. - AUC-ROC: values from 0 to 1 (1 being perfect model).