AWS Certified AI Practitioner

Artificial Intelligence

encompasses the development of intelligent systems capable of performing tasks that typically require human intelligence, such as perception, reasoning, learning, problem-solving, and decision making

Machine Learning

type of artificial intelligence for understanding and building methods that make it possible for machines to learn

Deep learning

subfield of machine learning that uses artificial neural networks to analyze data and learn complex patterns

Generative AI

subset of deep learning because it can adapt models built using deep learning, without retraining or fine tuning

capable of generating new data based on the patterns and structures learned from training data

Neural networks

lots of connected nodes that are organized into layers (input layer, hidden layer(s), and output layer)

when data is inputted, it figures out how to identify patterns by adjusting the connections between its nodes

Computer vision

a field of AI that makes it possible for computers to interpret and understand digital images and videos

ex: image classification, object detection, and image segmentation

Natural language processing (NLP)

branch of AI that deals with the interaction between computers and human languages

ex: text classification, sentiment analysis, machine translation, and language generation

Model

systems built using neural networks, system resources, data, and prompts, all working together to process input and produce outputs

Foundation model

models that are pretrained on internet-scale data

adapt a single FM to perform multiple tasks such as text generation, text summarization, information extraction, image generation, chatbot, and question answering

Algorithm

a set of instructions to be followed in calculations or other operations

Prompts

specific set of inputs to guide LLMs to generate an appropriate output or completion

Inference

process of using a trained model to generate results based on new input data (e.g., turning your prompt into an answer, image, or prediction)

Completion

the output the model produces in response to your prompt; often a continuation or expansion of the given input

Training

the process of feeding an AI model curated data sets to evolve the accuracy of its output

Labeled data

dataset where each instance/example is accompanied by a label or target variable that represents the desired output or classification

Unlabeled data

dataset where the instance/example do not have any associated labels or target variables (only input feature without corresponding output/classification)

Structured data

data that is organized and formatted in a predefined manner, typically in the form of tables or databases with rows and columns

Tabular data

structured data stored in spreadsheets, databases, or CSV files, with rows representing instances and columns representing features or attributes

Time-series data

structured data that consists of sequences of values measured at successive points in time, such as stock prices, sensor readings, or weather data

Semi-structured data

data that can have different attributes or missing attributes, such as a text file that contains JSON

Unstructured data

data that lacks a predefined structure or format, such as text, images, audio, and video

Text data

unstructured data that includes documents, articles, social media posts, and other textual data

Image data

unstructured data that includes digital images, photographs, and video frames

Inferencing

using the information that a model has learned to make predictions or decisions

Batch inferencing

analyzes a large amount of data all at once to provide a set of results

suitable for offline processing when data can be processed in batches. batch transform can support processing times of days therefore highest latency

Real-time inferencing

computer has to make decisions quickly, in response to new information as it comes in

real-time inference offers the lowest latency requirements because of the 60-second processing times

Asynchronous inferencing

can queue incoming requests for inference processing

asynchronous inference provides moderate latency requirements because of the processing times of up to 1 hour

Bias

model is missing important features of the datasets

measured by the difference between the expected predictions of the model and the true values we are trying to predict

Fairness

the equitable and impartial treatment of individuals, or data subjects, by AI systems

Fit

the process of adjusting the parameters of a model to best capture the patterns and relationships in the input data

Large language model (LLM)

powerful models that can understand and generate human-like text

use tokens, embeddings, and vectors to capture complex relationships in language

Tokens

basic units of text that the model processes (e.g. words, phrases, or characters) to provide standardization of input data

Vector

list of numbers representing data (such as text or an image) in a mathematical form the model can work with

Vector database

collection of data stored as mathematical representations

store structured and unstructured data, such as text or images with the vector embeddings

Embeddings

numerical representations of tokens, where each token is assigned a vector that captures its meaning and relationship with other tokens

a numerical vectorized representation of any entity that captures the meaning or semantic relationships of data

In-context learning

giving the model examples or extra information within the same prompt (inside the context window) so it can learn the task or style on the fly without retraining

Supervised learning

training on labeled data

learn a mapping function that can predict the output for new, unseen input data

Unsupervised learning

learn from unlabeled data

discover inherent patterns, structures, or relationships within the input data

Reinforcement learning

give only a performance score and portion of labeled training data

learns from feedback (rewards or penalties) to improve its decision-making over time

Classification model

a supervised learning technique used to assign labels or categories to new, unseen data instances based on a trained model

Use-cases: fraud detection, image classification, customer retention, and diagnostics

Regression model

a supervised learning technique used for predicting continuous or numerical values based on one or more input variable

Use-cases: advertising popularity prediction, weather forecasting, market forecasting, estimating life expectancy, population growth prediction

Clustering

an unsupervised learning technique that groups data into different clusters based on similar features or distances between the data point to better understand the attributes of a specific cluster

Use-cases: customer segmentation, targeted marketing, and recommended systems

Dimensionality reduction

an unsupervised learning technique used to reduce the number of features or dimensions in a dataset while preserving the most important information or patterns

Use-cases: big data visualization, meaningful compression, structure discovery, feature elicitation

Data collection

process of gathering and measuring information on variables of interest

should accurately reflect the diverse perspectives and experiences required for the use case of the AI system

Exploratory data analysis (EDA)

an analysis approach that identifies general patterns in the data

Data pre-processing

preprocess the data to ensure it is accurate, complete, and unbiased

techniques such as data cleaning, normalization, and feature selection can help to eliminate biases in the dataset

Feature engineering

the process of creating, transforming, extracting, and selecting variables from data

Model training

the process of training an ML algorithm with adequate training data to demonstrate correlation between the outcome and the influencing variables

Hyperparameter tuning

the problem of choosing a set of optimal hyperparameters for a learning algorithm

hyperparameter: a parameter whose value is used to control the learning process, which must be configured before the process starts

Model evaluation

the process of using different evaluation metrics to understand a machine learning model's performance, as well as its strengths and weaknesses

Model deployment

the process of putting machine learning models into production

Model monitoring

Ensures the model is maintaining a desired level of performance through early detection and mitigation

MLOps

refers to the practice of operationalizing and streamlining the end-to-end machine learning lifecycle from model development and deployment to monitoring and maintenance

Accuracy

model performance metric that measures the percentage of correct predictions or classifications

Area Under the ROC Curve

represents the probability that the model, if given a randomly chosen positive and negative example, will rank the positive higher than the negative

F1 score

the harmonic mean of precision and recall

Precision

proportion of positive predictions that are actually correct

Recall

proportion of correct sets that are identified as positive

Prompt engineering

focuses on developing, designing, and optimizing prompts to enhance the output of FMs for your needs

selects appropriate words, phrases, sentences, punctuation, and separator characters to effectively use LLMs for a wide variety of applications

Multi-modal models

can process and generate multiple modes of data simultaneously

learn how different modalities like images and text are connected and can influence each other

use-cases: automating video captioning, creating graphics from text instructions, answering questions more intelligently by combining text and visual info

Diffusion models

deep learning architecture system that starts with pure noise/random data, gradually adding more meaningful information to end with a clear and coherent output

Forward diffusion

system gradually introduces a small amount of noise to an input image until only the noise is left over

Reverse diffusion

noisy image is gradually introduced to denoising until a new image is generated

Stable diffusion

does not use the pixel space of the image, uses a reduced definition latent space

Generative adversarial networks (GANs)

a class of generative models that uses two neural networks — a generator that creates synthetic data, and a discriminator that tries to tell real data from fake data. The two compete in a feedback loop until the generator produces data that is indistinguishable from real examples

Variational autoencoders (VAEs)

A type of generative model that learns to encode input data into a compressed latent representation and then decode it back to reconstruct the data. By sampling from the latent space, VAEs can generate new, similar data

Intrinsic analysis

can be applied to models with low complexity or simple relationships between the input variables and the predictions

high model interpretability -> lower model performance

Post hoc analysis

can be applied to simple relationship models and more complex models that capture nonlinear interactions

Foundation model lifecycle

1. Data selection (FMs required training on massive datasets from diverse sources)

2. Pre-training (use self-supervised learning - make use of structure to autogenerate labels)

   • Can learn the meaning, context, and relationship of words in the dataset

3. Optimization

   • Techniques like prompt engineering, retrieval-augmented generation (RAG), and fine-tuning on task-specific data

4. Evaluation (measured using appropriate metrics and benchmarks)

5. Deployment

6. Feedback and continuous improvement (collected from users, domain experts, or other stakeholders to identify areas for improvement, detect potential bias, and inform future iterations of the model)

Advantages of generative AI

Adaptability, responsiveness, simplicity, creativity and exploration, data efficiency, personalization, and scalability

Toxicity

can generate content that is inflammatory, offensive, or inappropriate

Hallucinations

model generates inaccurate responses that are not consistent with the training data

Interpretability

the access into a system so that a human can interpret the model’s output based on the weights and features

Nondeterminism

model might generate different outputs for the same input, which can cause problems in applications where reliability is key

Agents

software components or entities designed to perform specific actions or tasks autonomously or semi-autonomously, based on predefined rules or algorithms

task coordination, reporting and logging, scalability and concurrency, integration and communication

Factors to select appropriate generative AI models

model types, performance requirements, capabilities, constraints, compliance

Business metrics for generative AI

User satisfaction, accuracy, customer lifetime value, average revenue per user, cross-domain performance, conversion rate, and efficiency

Context window

a model property that describes the number of tokens that the model can accept in the context

Temperature

controls the randomness or creativity of the model's output (set between 0 and 1)

Low temperature: Outputs are more conservative, repetitive, and focused on the most likely responses.

High temperature: Outputs are more diverse, creative, and unpredictable, but might be less coherent or relevant

Top K

limits the number of words to the top k most probable words, regardless of their percent probabilities

Low top k: With a low setting, like 10, the model will only consider the 10 most probable words for the next word in the sequence

High top k: With a high top k setting, like 500, the model will consider the 500 most probable words for the next word in the sequence, regardless of their individual probabilities.

Top P

a setting that controls the diversity of the text by limiting the number of words that the model can choose from based on their probabilities (set between 0 and 1)

Low top p (e.g 0.250): model will only consider words that make up the top 25 percent of the total probability distribution (more focused and coherent)

High top p (e.g 0.99): model will consider a broad range of possible words for the next word in the sequence, because it will include words that make up the top 99 percent of the total probability distribution (diverse and creative)

Stop sequences

special tokens or sequences of tokens that signal the model to stop generating further output

particularly useful in tasks where the desired output length is variable or difficult to predict in advance

Retrieval Augmented Generation (RAG)

a natural language processing (NLP) technique that combines the capabilities of retrieval systems and generative language models to produce high-quality and informative text outputs

reduces hallucinations

business applications: Building intelligent question-answering systems, Expanding and enriching existing knowledge bases, Generating high-quality content

Fine tuning

a supervised learning process that involves taking a pre-trained model and adding specific, smaller datasets

Elements of a prompt

Instructions: This is a task for the large language model to do. It provides a task description or instruction for how the model should perform.

Context: This is external information to guide the model.

Input data: This is the input for which you want a response.

Output indicator: This is the output type or format.

Negative prompting

involves providing the model with examples or instructions about what it should not generate or do

used to guide the model away from producing certain types of content or exhibiting specific behaviors

Zero-shot prompting

a technique where a user presents a task to a generative model without providing any examples or explicit training for that specific task

relies on the model's general knowledge and capabilities to understand and carry out the task without any prior exposure, or shots, of similar tasks

Few-shot prompting

technique that involves providing a language model with contextual examples to guide its understanding and expected output for a specific task

supplement the prompt with sample inputs and their corresponding desired outputs, effectively giving the model a few shots or demonstrations to condition it for the requested task.

Chain-of-thought prompting

a technique that divides intricate reasoning tasks into smaller, intermediary steps (use phrase “think step by step” to initiate)

recommended to use CoT prompting when the task requires multiple steps or a series of logical reasoning

Prompt tuning

the actual prompt text is replaced with a continuous embedding backer that is optimized during training

technique helps the prompt to be fine-tuned for a specific task, while keeping the rest of the model parameters frozen

Latent space

the encoded knowledge of language in LLMs or the stored patterns of data that capture relationships and reconstruct the language from the patterns when prompted

Amazon Bedrock Guardrails

provide safety and privacy controls to manage interactions in your generative AI applications

define threshold for content filters for hate, insults, sexual content, or violence

Poisoning

refers to the intentional introduction of malicious or biased data into the training dataset of a model

Hijacking and prompt injection

refer to the technique of influencing the outputs of generative models by embedding specific instructions within the prompts themselves

hijack the model's behavior and make it produce outputs that align with the attacker's intentions, such as generating misinformation or running malicious code

Exposure

refers to the risk of exposing sensitive or confidential information to a generative model during training or inference

Prompt leaking

refers to the unintentional disclosure or leakage of the prompts or inputs (regardless of whether these are protected data or not) used within a model

Jailbreaking

refers to the practice of modifying or circumventing the constraints and safety measures implemented in a generative model or AI assistant to gain unauthorized access or functionality

attempts involve crafting carefully constructed prompts or input sequences that aim to bypass or exploit vulnerabilities in the AI system's filtering mechanisms or constraints. The goal is to "break out" of the intended model limitations

Reinforcement learning from human feedback (RLHF)

an ML technique that incorporates human feedback to help models learn more efficiently

uses a supervised fine-tuning of a language model and a reward model

benefits: enhances AI performance, supplies complex training parameters, increases user satisfaction

Parameter-efficient fine-tuning (PEFT)

a process and set of techniques that freeze or preserve the parameters and weights of the original LLM and fine-tune a small number of task-specific adaptor layers and parameters

Low-Rank Adaptation (LoRA)

a popular PEFT technique that also preserves or freezes the original weights of the foundation model and creates new trainable low-rank matrices into each layer of a transformer architecture