DEV Community: Marwa Talaat

AWS Reinvent 2023: Unleashing the Power of ML and Generative AI

Marwa Talaat — Wed, 20 Dec 2023 14:46:03 +0000

Several interesting announcements were made within the just-ended AWS (Amazon Web Services) Reinvent 2023 conference, especially in the domains of machine learning (ML) and generative artificial intelligence (AI). Cutting-edge improvements and brand-new services were announced at the event, showing AWS's focus to expanding boundaries of innovation in these revolutionary technologies. We'll explore some of the main announcements and consequences for ML and generative AI in the future in the next section.

1. Introducing Amazon SageMaker Studio for Generative AI:

Among the most interesting announcements was the launch of Amazon SageMaker Studio, a feature-rich programming environment designed to facilitate workflows using generative AI. AWS intends to make it easier for data scientists and developers to design, test, and refine a variety of generative AI projects by automating the creation and implementation and deployment of generative models with SageMaker Studio. With the help of this new toolkit, users may now discover and take use of the possibilities of generative models in a variety of industries, including design, content creation, and the arts.

For more details check Amazon SageMaker Studio adds web-based interface, Code Editor, flexible workspaces, and streamlines user onboarding

2. Faster model deployment with guided workflows in Amazon SageMaker :

AWS recently released new features and updates for services in Amazon SageMaker. New interactive model deployment processes provide step-by-step guidance on which instance type to select in order to identify the most suitable endpoint configuration. It is boosting the use of ML by enterprises without the need for highly skilled employees. On top of that, SageMaker Studio offers more interfaces for test inference, adding models, and enabling auto scaling policies on the deployed endpoints.

For more details check Package and deploy models faster with new tools and guided workflows in Amazon SageMaker

3. Accelerated Inference Capabilities:

You can deploy several foundation models (FMs) on a single SageMaker endpoint and manage the number of accelerators and memory given to each FM by utilizing the newly available inference capabilities. When ML models are executed in production, AWS guarantees improved performance and cost-effectiveness. Its powerful capabilities enable real-time decision-making, accelerate inferencing activities for businesses, and improve operational efficiency for a variety of applications, such as computer vision and natural language processing.

For more details Amazon SageMaker adds new inference capabilities to help reduce foundation model deployment costs and latency

4. Launching Amazon Q, a new assistant powered by generative AI :

Amazon Q, a new generative AI-powered personal assistant that can be customized for your business, has been launched by AWS. Through access to the code, data, enterprise systems, and information repositories of your organization, Amazon Q enables you to engage in conversations, solve issues, produce content, acquire insights, and act. With Amazon Q's user-based plans, you may customize the product's features, cost, and options to suit your needs. Based on the organization's current identities, positions, and permissions, Amazon Q can customize its interactions for every single user.

For more details Introducing Amazon Q, a new generative AI-powered assistant (preview)

5. Build generative AI applications with Amazon Bedrock :

Two newly optimized integration between Amazon Bedrock and AWS Step Functions were announced by AWS. With the use of Step Functions, a visual workflow tool, developers can build distributed applications, automate workflows, integrate microservices, and build pipelines for data and machine learning (ML).
AWS released Amazon Bedrock earlier this year, which is the simplest approach for developing and expanding generative AI systems using foundation models (FMs). Bedrock offers a wide range of capabilities that customers require to develop generative AI applications while upholding privacy and security, offering a selection of foundation models from top suppliers including AI21 Labs, Anthropic, Cohere, Stability AI, and Amazon.

For more details Build generative AI apps using AWS Step Functions and Amazon Bedrock

The ML and generative AI announcements released at AWS Reinvent 2023 show how devoted AWS is to providing businesses with state-of-the-art capabilities. AWS is well-positioned for leading the next wave of innovation in machine learning and generative artificial intelligence applications with the release of SageMaker Studio for Generative AI, improved inference capabilities for faster inference, development generative AI apps with Amazon Bedrock, and Amazon Q.

Organizations in a variety of industries will be able to use ML and Generative AI to stimulate corporate growth, improve consumer experiences, and open new opportunities as these new services and tools become more widely available. In the years to come, we can expect interesting improvements and revolutionary uses of ML and generative AI, with AWS Reinvent 2023 laying the foundation for what's to come.

The mentioned announcements are some of the main announcements for ML and generative AI to see all announcement related to Generative AI / Machine Learning, check #Generative_AI for more announcements and their details

Study Plan to pass exam AWS Machine Learning Specialty exam with tips and advice

Marwa Talaat — Thu, 03 Nov 2022 08:58:36 +0000

The AWS Machine Learning Specialty exam is a challenging task that requires a lot of commitment, dedication, and covering theory and practice. The main objective of the exam is to validate your expertise in machine learning and your ability to apply it in the AWS cloud. It requires a lot of knowledge and experience to pass.

Whether or Not You Should follow This Guide?

Before we get started, let's evaluate if this tutorial is right for you. The target audience for this learning guide is:

Understand at least one high-level programming language in-depth and practical development experience (at least 2 years as recommended by AWS)
Basic understanding of Machine Learning. if you have a solid knowledge of data science and machine learning. You won’t need to go through all these resources.
You have 3 hours a day, 5 times a week excluding weekends 😃

These learning resources are offered based on the real requirement in each plan stage.

Stage one: Testing the waters:

Learning Resource	Duration	Outcomes
What is AWS Cloud?	60 minutes	You will learn about the foundations of getting started in the AWS Cloud.
AWS Cloud Practitioner Essentials	6 hours	This course is intended for anyone who, regardless of their technical positions, wants a general grasp of the Amazon Web Services (AWS) Cloud. You will gain an understanding of AWS Cloud principles, AWS services, security, architecture, pricing, and support. You may use this course to be prepared for the AWS Certified Cloud Practitioner test as well.
Demystifying AI/ML/DL	45 minutes	You will learn the relationship between artificial intelligence (AI), machine learning (ML), and deep learning after completing this collection of courses (DL).
AWS Foundations: Machine Learning Basics	30 minutes	You learn about the concepts, terminology, and processes of the exciting topic of machine learning!
Learn Python on AWS Workshop	3 Hours	You will learn the fundamentals of Python programming in this class utilizing Amazon Web Services (AWS).It is intended for beginners who have never coded in Python before, and it employs similar ways of introducing the fundamentals as previous books and tutorials on the Python programming language.

Stage Two: Self-based Learning Resources

These learning resources are crucial components of the suggested learning strategy and will aid in your acquisition of new AWS machine learning capabilities and services.

Week- 1: Average 11 hours to 13 hours

During this week you get to know the fundamentals of AWS cloud and AWS machine learning.

Learning Resource	Duration	Outcomes
What is Cloud Computing with AWS?	10 minutes	You will get an overview of the AWS cloud, functionalities, regions, etc.
AWS glossary	30 minutes	AWS Glossary to know definitions of key terms and concepts
Job Roles in the Cloud	30 minutes	You will learn about the typical job roles applicable to an enterprise-level AWS Cloud environment
AWS Pricing	15 minutes	You will learn a broad view of How does AWS pricing work? How do you pay for AWS? And pricing for AWS products
How AWS Pricing Works: AWS Pricing Overview	60 minutes	You'll discover that the flexibility to adjust expenses to match your needs, even as those needs change over time, is one of the key advantages of cloud services
AWS Support Plans	10 minutes	You will discover how AWS Support plans are created to provide you with the ideal combination of tools and access to knowledge so that you can succeed with AWS while maximizing performance, minimizing risk, and keeping costs in check.
AWS Shared Responsibility Model Shared Responsibility Model	15 minutes	You will get the AWS Shared Responsibility Model introduction. This course clarifies the separation of those obligations between AWS and the client, who both share responsibility for security and compliance.
AWS Cloud Security	5 minutes	You will discover how AWS enables you to take back control of your organization and instils the confidence you need to operate it safely in the most adaptable and secure cloud computing environment currently available.
Machine Learning for Leaders	70 minutes	This course will teach you how as a business leader, machine learning may help your teams in maximizing project performance and gaining crucial insight into your company's or your customers' demands.
Machine Learning for Business Challenges	60 minutes	You will learn how can you use machine learning (ML) to address business challenges in ways that weren't previously feasible, but you must think broadly.
Machine Learning Terminology and Process	60 minutes	You will understand the fundamentals of machine learning in this course, as well as how machines analyze data. We thoroughly examine each stage of the machine learning process and define some of the words and methods that are frequently used in a particular stage of an ML project.
Process Model: CRISP-DM on the AWS Stack	50 minutes	You learn about data science as a circular process through the CRISP-DM paradigm. With the help of Jake Chen, an AWS data scientist consultant, we'll go over the CRISP-DM methodology and framework before putting its six stages to use in your day-to-day job as a data scientist.
Machine Learning on AWS	30 minutes	You will get an overview of the machine learning capabilities that AWS has.
Exploring the Machine Learning Toolset	20 minutes	You will learn that anyone can apply machine learning no matter what your background or experience is. you'll go through a few of the AWS machine learning services you can use to create models and give app intelligence in this lecture.
Data Analytics Fundamentals	210 minutes	You will learn about the process of planning data analysis solutions, as well as the numerous data analytic techniques involved.

Week- 2: Average 16 hours to 19 hours

This week is kind of tough you will go back to basics in math to understand vectors and matrices, linear algebra, probability theorems, univariate calculus, and multivariate calculus as well as elements of data science.

Learning Resource	Duration	Outcomes
Math for Machine Learning	480 minutes	Modern machine learning requires an understanding of vectors and matrices, linear algebra, probability theorems, univariate calculus, and multivariate calculus.
The Elements of Data Science	480 minutes	You will learn how to develop and constantly improve machine learning models.

Week- 3: Average 15 hours to 17 hours

Concepts and best practices for AWS Machine Learning in the cloud.

Learning Resource	Duration	Outcomes
Planning a Machine Learning Project	30 minutes	You learn how to assess the data, time, and production requirements for a successful ML project.
Building a Machine Learning Ready Organization	30 minutes	This course outlines the components required for successful machine learning adoption in organizations.
Introduction to Amazon SageMaker	13 minutes
AWS Foundations: How Amazon SageMaker Can Help	30 minutes	You will explore how Amazon SageMaker mitigates the primary problems associated with creating a machine learning pipeline.
Developing Machine Learning Applications	240 minutes	You'll look at Amazon SageMaker, a fully managed machine learning platform. We'll talk about how to train and fine-tune models; how certain algorithms are built-in, etc.
Amazon SageMaker: Build an Object Detection Model Using Images Labeled with Ground Truth	70 minutes	You will learn how to implement a machine learning pipeline using Amazon SageMaker and Amazon SageMaker Ground Truth.
Machine Learning Security	120 minutes	You will how to control and maintain permissions, as well as approve traffic, which are all part of developing highly secure applications and environments on the AWS platform.
Amazon SageMaker Technical Deep Dive Series	275 minutes	Deep dive videos series provided by AWS.
Deep Learning on AWS	120 minutes	A technical guide to running deep learning in AWS cloud.

Week- 4: Average 18 hours to 20 hours

Learning Resource	Duration	Outcomes
Analyze Datasets and Train ML Models using AutoML	18 hours	Prepare data, detect statistical data biases, and perform feature engineering at scale to train models with pre-built algorithms.

Week- 5: Average 13 hours to 15 hours

Learning Resource	Duration	Outcomes
Build, Train, and Deploy ML Pipelines using BERT	13 hours	Store and manage machine learning features using a feature store. Debug, profile, tune and evaluate models while tracking data lineage and model artifacts.

Week- 6: Average 14 hours to 16 hours

Learning Resource	Duration	Outcomes
Optimize ML Models and Deploy Human-in-the-Loop Pipelines	14 hours	You will learn how to build, train, and deploy scalable, end-to-end ML pipelines - both automated and human-in-the-loop - in the AWS cloud.

Stage Three: Real live environment practice

Week- 7 & 8 & 9: Average 36 hours to 40 hours

It's time to get your hands dirty by solving some ML Use Cases of your own from AWS SageMaker Use Cases repo.

Learning Resource	Duration	Outcomes
Risk Bucketing	6 hours	One of the most common use cases for machine learning in financial services is estimating the probability of default on a loan.
Churn Prediction for Music Streaming	6 hours	Use case to detect whether the customer tends to leave and stop paying for a service.
Payment Classification	6 hours	Use case to classify payment transactions.
Fleet Predictive Maintenance	6 hours	Use case to demonstrate a Predictive Maintenance (PrM) solution for automible fleet maintenance via Amazon SageMaker Studio.
Amazon SageMaker Pipelines	6 hours	In this use case we use SageMaker Pipelines to train and deploy a text classification model to predict e-commerce product ratings based on customers’ product reviews.
Computer Vision	6 hours	Use case about computer Vision for Medical Imaging - Pipeline Mode

Stage Four: Prepare for and take the AWS Certified Machine Learning - Specialty certification exam

Additional recommended resources

Till this stage you should be about 50% ready for the exam. Though AWS provides an exact passing score of 750 for this test, it's safe to infer that a score of 75% to 80% is necessary to pass. With an extra month or two of preparation, I feel the certification is well within reach.

Official documentations are important preparation resource. Its always recommended to review the following:

These tips for exam preparation:

You should be familiar with AWS's AI services’ Rekognition, AWS Polly, Comprehend, Textract, Lex, etc. are some examples. You will be asked to select a service based on a high-level use case when they are present on the exam, and you may be asked about service requirements to use it.
Use Amazon SageMaker Built-in Algorithms or Pre-trained Models
For review and practice, the following courses are highly recommended:
o AWS Certified Machine Learning Specialty 2022 - Hands On! and AWS Certified Machine Learning Specialty Full Practice Exam
Study notes directly from AWS Training course
You can practice more for free on examtopics , braindumps, certification-questions, and many more . Although if you’re willing to pay tutorialsdojo has several nice sets of questions available.
It is preferable to verify and answer the questions accurately than to guess. Obviously, as you progress through the practice tests, you should refer to the materials less frequently.
Save the questions you're having problems answering, so you can easily go back and revise them.
You also need to bear in mind, you need to know what is new in AWS from AI services.
Non-native English speakers can request a 30-minute exam extension when taking an English exam. The accommodation "ESL +30" only must be requested once, before enrolling for a test.

Last but not least, always remember that there is no need to know all the answers to ace the exam, so don't put yourself under unnecessary stress!

Best of luck with the exam!

Why do we love AWS Step Functions vs AWS Lambda in AI/ML?

Marwa Talaat — Mon, 03 Oct 2022 18:22:52 +0000

Interested to know if you need to use AWS step functions in your application? This is a complete guide on what are AWS Step functions, how AWS Step Functions works, their benefits, and their strengths over AWS Lambda and weaknesses in different situations.

What are AWS Step Functions?

AWS Step Functions is a low-code visual workflow application that allows you to create, deploy, and automate business processes as well as data and machine learning pipelines using AWS services. Workflows handle exceptions, retries, parallelization, service integration, and monitoring, allowing developers to concentrate on higher-value business logic.
In other words, step functions are orchestrators to help to design and implement complex workflows such as batch processing. Step Functions coordinates between multiple tasks that need orchestration, making it simple to build multi-step systems.

You could develop interactive and complicated systems that use all the features mentioned in addition to full orchestration and ease of transparency with AWS Step Functions to manage and shape these interactions. Let's discuss it right away.

AWS Step Function Benefits

Build and deploy rapidly:
Workflow Studio offers a straightforward drag-and-drop user interface that makes getting started quickly. Step Functions allow you to quickly connect services, systems, or people by using low-code, event-driven workflows to describe complex business logic.
Write less integration code:
Build robust business workflows, data pipelines, or apps using AWS resources from more than 200 services, such as Lambda, ECS, Fargate, Batch, DynamoDB, SNS, SQS, SageMaker, EventBridge, or EMR.
Build fault-tolerant and stateful workflows:
Step Functions keeps track of managing state, checkpoints, and restarts so that your workflows proceed as planned. Based on your predefined business logic, automatic error and exception handling are provided through built-in try/catch, retry, and rollback capabilities.
Designed for reliability and scale:
Depending on your particular use case, you can choose between the Standard or Express workflow types that Step Functions offers. Long-running workloads are managed using standard workflows. Workloads for high-volume event processing are supported by Express Workflows.
Parallelism:
Declarative parallelism is possible for the work. A state of a step machine may invoke different states simultaneously. The workflow will proceed more quickly as a result.
High Execution Time:
If some of the tasks in the workflow require a lot of time (exceeding 10 minutes), they can be executed on ECS, EC2, or as an Activity hosted outside of AWS because Step Functions have a maximum execution time of one year.

How did AWS Step Functions build?

AWS Step Functions consist of the following main components:

State Machine

The term "state machine" is used to describe an application workflow in AWS Step Functions, which is based on this very idea. The Amazon States Language (ASL ) allows programmers to create state machines in Step Functions using JSON files.

Processes that take a while to complete or require human involvement can be defined as a regular workflow. while Express workflows are ideal for quick, high-volume procedures that complete in under five minutes.

State machine takes data in 3 main forms; input in the initial state, data which passed between states, and output in the final state.

For more information, see State Machine Data

State

Individual states could act, make decisions based on their input, and send output to other states. In your state machine, states are elements. A state is identified by its name, which can be any string but needs to be unique across the entire state machine.

A state represents a step in your workflow. States can perform a variety of functions:

• Do some work in your state machine (a Task state)
• Choose between branches of execution (a Choice state)
• Stop execution with a failure or success (a Fail or Succeed state)
• Simply pass its input to its output or inject some fixed data (a Pass state)
• Provide a delay for a certain amount of time or until a specified time/date (a Wait state)
• Begin parallel branches of execution (a Parallel state)
• Dynamically iterate steps (a Map state)

Example of state definition of task type using ASL

"States": {
"FirstState": {
"Type": "Task",
"Resource": "$",
"Next": "My Next state"
}

The states that you decide to include in your state machine and the relationships between your states form the core of your Step Functions workflow.

For more information, see Concepts States

Tasks state and Activities

Tasks: A Task also referred to as a task state is a single unit of work used in the state machine. A task can be used in invoking the Lambda function or calling the API of other services.

Activities: Activity is used to perform a task. It lets you connect your step function with a batch of code that is running elsewhere which is known as an activity worker.

You can see and verify your state machine as a set of steps through your AWS console. Step Functions record the execution time, input, output, number of retries, and errors for each step as it is carried out. Engineering teams may quickly identify which step or steps may have caused a workflow to fail and which steps may have caused that failure with the use of this information.

Drawbacks of Step functions

Vendor lock: since it's used on AWS. if you decide to migrate your application to a different cloud vendor., you will need to remodel your application or replace it with an alternative service from a new vendor.
Complex syntax: Your application code may become more difficult to comprehend for other members of your team who might need to edit or upgrade it because of separating business logic from workflow logic. The Amazon Statements language used to configure step functions is very complex. The syntax of this language is based on JSON. In other words, the language is ideal for machine readability, not for humans. This language can be difficult to learn and is only available for AWS Step Functions as it is an AWS proprietary language.
The maximum limit for keeping execution history logs is 90 days.
The missing trigger for some events like in DynamoDB and Kinesis.
Each Execution name for a state machine must be unique and not used in the last 90 days.

Step Functions limitation

Maximum item execution is 25,000 per workflow; so, you will need to divide workflow into multiple workflows to not exceed the limit.
Request made to Step functions size should not exceed 1MB.
Maximum 50 tags per each resource in step functions

AWS Step Functions vs AWS Lambda: what are the differences?

Overall comparison

AWS Step Functions	AWS Lambda
Using visual workflow, it builds distributed applications. Simplify coordinating components of distributed applications and microservices through visual workflows. You can expand and update apps efficiently by creating them from separate components that each perform a distinct function	Runs code automatically to respond to object modifications. Create your back-end services that use AWS scalability, performance, and security while extending existing AWS services with custom logic.
Cloud Task Management category	Serverless category
Integration with other services	No infrastructure
State machine, tasks execution unit	Lambda function execution unit
Supported Runtimes: Java,.NET, Ruby, PHP, Python (Boto 3),JavaScript, Go, C++	Supported Runtimes: Nodejs 12, 14, 16, Python 3.6, 3.7, 3.8, Java 8, 11, .NET 5,6, Core 3.1, GO 1.x, Ruby 2.7
AWS free tier offers 4,000 Step Functions state transitions per month Beyond the free tier the price is depending on region, but the price is calculated per 1,000 state transitions	free tier offers 400,000 GB-seconds of compute time per month. Beyond free tier—$0.00001667 per every GB-second

Check AWS Price Calculator to calculate your AWS services and architecture cost estimation.

Cost factors comparison

Prices are based on US East Ohio Region

	Lambda	Step Functions-Standard workflows
Invocation	$0.20 per 1M requests	NA
Consuming GB per seconds	$0.0000166667 for every GB-second	NA
State transitioning	NA	$0.025 per 1K

You may observe that standard workflows have different pricing. Their charge is primarily based on the number of state changes.

Flexibility Comparison

Lambda is not well known for its flexibility and its ability to perform complex and long-running operations. For instance, Lambda is limited to up to 5 minutes, however, you can extend it to 15 minutes. This may not be practical for large complex scripts.

For example, if you build a simple app using Lambda that you expect some common steps that people expect like retrying the connection until service will be available to move to the next step or run operations in parallel. Additionally, users are unable to run failed code again. Unfortunately, Lambda does not come with these functionalities by default.

This should not discourage you from using Lambda at all, as AWS Step Functions can win in this situation.

You can develop interactive and complicated systems that make use of all the elements we just mentioned and more with complete orchestration and ease of transparency by using AWS Step Functions to manage and shape these interactions.

Development & Integrations Comparison

To perform these longer activities, it is possible to add containers using a solution like Amazon Elastic Container Service (EC2), but what happens when the containers are also working on other tasks? You need to include the state.

AWS Lambda's inability to handle state effectively is the issue. State management requires developers to add code into their systems, which complicates management and extends processing times. As a result, users are forced to decide between using resource-intensive apps and ensuring there is enough state for peak usage.

Fortunately, these obstacles are easily overcome thanks to AWS Step Functions. Because less code is needed to do tasks, AWS Step Functions are genuinely unique. Engineers can build standardized procedures for dealing with errors, retries, and parallelization that free up developer resources for other higher-value tasks. By employing multi-service management, they may get rid of complicated state handling and streamline laborious application development procedures.

For more information on both services integration; check Using AWS Lambda with other services, and Using AWS Step Functions with other services

Benchmarking Comparison

If you are interested in benchmarking AWS Lambda and AWS Step functions, you could check the following blog posts:

Conclusion

To sum up, developers can quickly add, move, switch, and reorganize Lambda functions using AWS Step Functions' visual workflow interface without having to alter the business logic that is already in place. This abstraction makes it relatively simple to increase application performance without writing additional code.

Additionally, AWS Step Functions easily integrates with other AWS services. The task can be configured to execute concurrently with other activities, wait for external processes, or wait for the completion of other internal tasks by developers. With AWS SageMaker, AWS Batch, serverless ETL with AWS Glue, and many more tools, users may integrate machine learning algorithms into their applications.

All these benefits make AWS Step Functions an appealing solution for AI/ML apps that are using the AWS cloud platform.

References:

AWS Step Functions - Lumigo. (2022). From https://lumigo.io/aws-serverless-ecosystem/aws-step-functions-limits-use-cases-best-practices/
AWS Step Functions | Serverless Visual Workflows | Amazon Web Services. (2022).From https://aws.amazon.com/step-functions/
Hombergs, T. (2022). Getting Started with AWS Step Functions. From https://reflectoring.io/getting-started-with-aws-step-functions-tutorial/
Lambda runtimes. (2022). From https://docs.aws.amazon.com/lambda/latest/dg/lambda-runtimes.html
What is AWS Step Functions? How it Works & Use Cases | Datadog. (2022). From https://www.datadoghq.com/knowledge-center/aws-step-functions/
What is State Machine? - Definition from Techopedia. (2022). From https://www.techopedia.com/definition/16447/state-machine
AWS Lambda vs AWS Step Functions | What are the differences? (2022).From https://stackshare.io/stackups/aws-lambda-vs-aws-step-functions
Hirsch, R. (2022). Ask the Expert: AWS Lambda vs Lambda@Edge vs Step Functions.From https://www.jeffersonfrank.com/insights/aws-lambda-vs-edge-vs-step-functions
Top 3 Benefits of Using AWS Step Functions. (2022).From https://blog.clearscale.com/top-3-benefits-of-aws-step-functions/

Hybrid Machine Learning | AWS Whitepaper Summary

Marwa Talaat — Sat, 13 Nov 2021 10:09:32 +0000

This article aims to discuss the outline, known considerations, design patterns, and solutions that customers should know when considering integrations between local compute and the AWS cloud across the machine learning lifecycle.

AWS purpose hybrid ML patterns as an intermediate step in their cloud and ML journey. The patterns involve a minimum of two compute environments, typically local compute resources such as personal laptops or corporate data centers, and the cloud.

This article is intended for individuals who already have a baseline understanding of machine learning, in addition to Amazon SageMaker.

Stage One: Basics

Developing technology that applies machine learning is challenging since it depends on data. Datasets vary from bytes to petabytes, objects to file systems, text to vision, tables to logs. Software frameworks supporting machine learning models evolve rapidly, undergoing potentially major changes multiple times a year, if not quarter or month. Nowadays, data science projects require different skill levels in team, from business stakeholders, quality and availability of datasets and models, and customer adoption

Companies who adopt a cloud-native approach realize its value of compute capacity with the needs of their business, technical resources to focus on building features, rather than taking on the burden of managing and maintaining their own underlying infrastructure. But for those companies born before the cloud and even for newer companies founded more recently, potentially those that made an informed decision to build on-premises, how can they realize the value of newly launched cloud services when the early requirements that were once infeasible on the cloud are now within reach?

For customers who want to integrate the cloud with existing on-premises ML, AWS propose a series of tenets to guide our discussion: -

Seamless management experience: Customers need end-to-end ML across multiple compute environments, limiting the burden of administrative while successfully operate complex tasks.
Tunable latency: The customers enjoy using applications that respond within the timelines of their moment-to-moment expectations, and designers of these applications understand the criticality of time-bound SLA’s. Engineers want to work with ML models that can respond to an app’s request within in milliseconds, regardless of the hosted. While not every customer requires response times at low latency levels, but faster is better.
Fast time-to-value: Customers expect solutions to be easy to use, with simple interfaces, and not requiring significant amounts of platform-specific engineering to execute a task.
Flexible: Customers need compute paradigms that provide the flexibility their business demands. ML applications may need to serve real-time responses to billions of users worldwide. Service providers should anticipate for deploying all environments.
Low Cost: Customers want transparency in their cost structure, they need to see a clear economic advantage to computing in the cloud relative to developing locally. Service providers need to anticipate this and compete on cost with respect to local compute options
End in the cloud: If there was any doubt that cloud computing is the way of the future, the global pandemic of 2020 put that doubt to rest. AWS also call out the final state of that design, helping customers understand which cloud technologies to leverage in the long runleverage in the long run.

There are two very different approaches to hosting – one type of pattern trains in the cloud with the intention of hosting the model itself on-premises, while another hosts the model in the cloud to applications deployed on-premises. Finally, a key pillar in applying these patterns is security.

What is hybrid?

At AWS, we look at “hybrid” capabilities as those that touch the cloud in some capacity, while also touching local compute resources. Local compute such as laptops hosting Jupyter notebooks and Python scripts, HDFS clusters storing terabytes of data, AWS Outposts stored on-premises, or AWS Outposts stored on-premises.

The hybrid architectures are having a minimum of two compute environments, what we will call “primary” and “secondary” environments. The primary environment as where the workload begins, and secondary environment is where the workload ends.

Depending on your case for instance if you are packaging up a model locally to deploy to the cloud, you might call your local laptop “primary” and your cloud environment “secondary.”. However, if you are training on cloud and want to deploy locally, you might call your cloud environment “primary” and your local environment “secondary.”

What hybrid is not?

There are some container-specific tools that provide a “run anywhere” experience, such as EKS and ECS. In those contexts, we will lean into prescriptive guidance for building, training, and deploying machine learning models with these services.

Stage Two: Hybrid patterns for development

Development refers to the phase in machine learning when customers are iteratively building models. This may or may not include exploratory data analysis, deep learning model development and compilation, software package installation and management, Jupyter kernels, visualization, Docker image building, and Python-driven data manipulation.

There are two major options for hybrid development that customer can apply one or both. Laptop and desktop personal computers. Self-managed local servers utilizing specialized GPUs, colocations, self-managed racks, or corporate data centers.

Develop on personal computers, to train and host in the cloud

Customers can use local development environments, such as PyCharm or Jupyter installations on their laptops or personal computer, and then connect to the cloud via AWS Identity and Access Management (IAM) permissions and interface with AWS service APIs through the AWS CLI or an AWS SDK (ex boto3). Having connected to the cloud, customers can execute training jobs and/or deploy resources.

Develop on local servers, to train and host in the cloud

Stage Three: Training

Hybrid patterns for training

Hybrid pattern for training comes down to one of two paths. Either you train locally and deploy on the cloud. Or the data sitting on local resources and want to select from that to move into the cloud to train.

Training locally, to deploy in the cloud

During enterprise migrations, training locally may be advantageous as a first step to develop a model. There are two key actions here. First, if you are training locally then you will need to acquire the compute capacity to train a model and think about size of dataset, and models that you want to use. When you are training on-premises, you need to plan for that well in advance and acquire the compute resources ahead of time.

After your model is trained, there are two common approaches for packaging and hosting it in the cloud. One Simple path is Docker, using a Docker file you can build your own custom image that hosts your inference script, model artifact, and packages. Register this image in the Elastic Container Registry (ECR) and point to it from your SageMaker estimator.

Another option is using pre-built containers within the SageMaker Python SDK. Bring your inference script and custom packages, upload your model artifact to Amazon S3, and import an estimator for your framework of choice.

In the following diagram, we outline how to do this from your laptop. The pattern is similar for doing the same from an enterprise data center with servers, as outlined above.

How to monitor your model in the cloud?

A key feature for hosting is model monitor, or the ability to detect data, bias, feature, and model quality drift. It is ability to capture data hitting your real-time endpoint, and programmatically compare this to your training data.

Enabling model monitor is easy in SageMaker. Upload your training data to an Amazon S3 bucket and use our pre-built image to learn the upper and lower bounds on your training data. This job uses Amazon Deequ to perform “unit testing for data,” and you will receive a JSON file with the upper and lower statistically recommended bounds for each feature. You can modify these thresholds. After confirming your thresholds, schedule monitoring jobs in your production environment. The jobs run automatically, comparing your captured inference requests in Amazon S3 with your thresholds.

CloudWatch will alert you when your inference data is outside of your pre-determined thresholds, and you can use those alerts to trigger a re-training pipeline.

How to handle retraining / retuning?

SageMaker makes train and tuning jobs easy to manage, because all you need to bring is your training script and dataset. Ensure your new dataset is loaded into an Amazon S3 bucket, or other supported data source.

Once you have defined a training estimator, it is trivial to extend this to support hyperparameter tuning. Define your tuning job configuration using tuning best practices and execute. Having defined a tuning job, you can automate this in a variety of ways. While AWS Lambda may seem compelling upfront, to use the SageMaker Python SDK (and not boto3) with Lambda, sadly you need to create an executable layer to upload within your function.

You may consider SageMaker Pipelines, an MLOps framework that uses your SageMaker Python SDK job constructs as argument and creates a step-driven framework to execute your entire pipeline.

How to serve thousands of models in the cloud at low cost?

You may consider Multi-model endpoints give you the ability to serve thousands of models from a single endpoint, invoking the name of the model when calling predict.
Create the multi-model endpoint, pointing to Amazon S3, and load your model artifacts into the bucket. Invoke the endpoint from your client application, (eg. with AWS Lambda), and dynamically select the right model in your application. It allows to host up to 5 containers on a single SageMaker endpoint, invoking the endpoint with the name of the model you want to use.

Storing data locally, to train and deploy in the cloud

when and how do I move my data to the cloud?

Schedule data transfer jobs with AWS DataSync

It is a data transfer service that simplifies, automates, and accelerates moving data between on-premises storage systems and AWS storage services, as well as between AWS storage services.

It can be easily moved petabytes of data from your local on-premises servers up to the AWS cloud.
It can be deployed can easily move petabytes of data from your local on-premises servers up to the AWS cloud.
It can be connected to your local NFS resources and deploy directly into Amazon S3 buckets or EFS volumes, or both.

Migrating from Local HDFS

As customers explore migrating data stored in local HDFS clusters, typically they find themselves somewhere between two extremes.

On the other, you might wholly embrace HDFS as your center and move towards hosting it within a managed service, Amazon Elastic Map Reduce (EMR).

Best practices

• Use Amazon S3 intelligent tiering for objects over 128 KB
• Use multiple AWS accounts, and connect them with Organizations
• Set billing alerts
• Enable SSO with your current Active Directory provider
• Turn on Studio!

Develop in the cloud while connecting to data hosted on-premises

Customers who see the value of outsourcing management and upkeep of their enterprise ML development platforms, i.e. through using managed services like Amazon SageMaker, can still connect in to their on-premises data store at the beginning and middle phases of their enterprise migration.

Data Wrangler & Snowflake

Data Wrangler enables customers to browse and access data stores across Amazon S3, Amazon 3rd Athena, Amazon Redshift, and party data warehouses like Snowflake. This hybrid ML patten provides customers the ability to develop in the cloud while accessing data stored on premises, as organizations develop their migration plans.

Train in the cloud, to deploy ML models on-premises

You can download whatever type of model artifact you need, but if you are deploying on-premises, you need to develop and host your own local web server.

This scenario builds on your previous experience developing and training in the cloud, with the key difference of exporting your model artifact to deploy locally. AWS recommends using dev and/or test endpoints in the cloud to give your teams the maximum potential to develop the best models they can.

If you are using a managed deep learning container, also known as “script mode,” for training and tuning, but you still want to deploy that model locally, plan on building your own image with your preferred software version, scanning, maintaining, and patching this over time. If you are using your own image, you will need to own updating that image as the software version, such as TensorFlow. Note that the best practice is to decouple hosting your ML model from hosting your application.

As models grow and shrink in size, hitting potentially billions of parameters and hundreds of gigs in byte size, or shrinking down to hundreds of parameters and staying under a few MB in size, you want the elasticity of the cloud to seamlessly map the state-of-the-art model to an efficient hardware choice.

Stage Four: Deployment

Monitor ML models deployed on-premises with SageMaker Edge Manager

Customers can train ML models in the cloud, deploy these on-premises, and monitor and update them in the cloud using SageMaker Edge Manager. SageMaker Edge Manage makes it easy for customers to manage ML models deployed on Windows, Linux, or ARM-based compute environments.

While customers do still need to provision, manage, procure, and physically secure the local compute environments in this pattern, Edge Manage simplifies the monitoring and updating of these models by bringing the control plane up to the cloud. However, you can bring your own monitoring algorithm to the service and trigger retraining pipelines as necessary, using the service the redeploy that model back down to the local device. This is particularly common for technology companies developing models for personal computers, such as laptops and desktops.

Hybrid patterns for deployment

Hybrid ML patterns around deployment can be interesting and complex. Choosing the “best” local deployment option has a lot of variety. You want to think about where your customers sit geographically, then you want to get your solution as close to them as you can. You want to balance speed with cost, cutting-edge solutions with ease of deployment and managing.

In this section will discuss the architecture for hosting an ML model via SageMaker in an AWS region, serving responses to requests from applications hosted on-premises. After that we’ll look at additional patterns for hosting ML models via Lambda at the Edge, Outposts, Local Zones, and Wavelength.

Serve models in the cloud to applications hosted on-premises

The most common use case for a hybrid pattern like this is enterprise migrations. You might have a data science team with tens of models, if not more than one hundred, ready to deploy via the cloud, while your application team is still refactoring their code to host on cloud-native services.

Host ML Models with Lambda at Edge to applications on-premises

This pattern takes advantage of a key capability of the AWS global network – the content delivery network known as Amazon CloudFront. Deploying content to CloudFront is easy, customers can package up code via AWS Lambda and set it to trigger from their CloudFront distribution.

What’s elegant about this approach is that CloudFront manages which of the 230+ points of presence will execute your function. Once you’ve set your Lambda function to trigger off CloudFront, you’re telling the service to replicate that function across all available regions and points of presence. This can take up to 8 minutes to replicate and become available.

This is a huge value-add for global companies looking at improving their digital customer experience worldwide.

AWS Local Zones

Local Zones are a way to extend your cloud resources to physical locations that are geographically closer to your customers. You can deploy ML models via ECS or EKS to serve inference with ultra- low latency near your customers, using AWS Local Zones.

AWS Wavelength

Wavelength is ideal when you are solving applications around mobile 5G devices, either anticipating network drop-offs or serving uses real-time model inference results. Wavelength provides ultra-low latency to 5G devices, and you can deploy ML models to this service via ECS or EKS. Wavelength embeds storage and compute inside the telecom providers, which is the actual 5G network.

Training with a 3rd party SaaS provider to host in the cloud

There are a lot of great SaaS providers for ML out there in the market today, like H20, DataRobot, Databricks, SAS, and others. 3rd Hosting a model in Amazon SageMaker that was trained from a party SAAS provider is easy. Ensure your provider allows export of proprietary software frameworks, such as with jars, bundles, images, etc. Follow steps to create a Docker file usingthat software framework, port into the Elastic Container Registry, and host on SageMaker.

Keep in mind that providers will have different ways of handling software, in particular images and image versions.

Control plane patterns for hybrid ML

AWS uses the concept of a “control plane,” or set of features dedicated to operations and management, while keeping this distinct from the “data plane,” or the datasets, containers, software, and compute environments.

Customers’ need it for operationalizing ML workloads are as varied and diverse as the businesses they exist within. Today it is not feasible for a single workflow orchestration tool to solve every problem, so most customers standardize on one workflow paradigm while keeping options open for others that may better solve given use cases. One such common control plane is Kubeflow in conjunction with EKS Anywhere.

SageMaker offers a native approach for workflow orchestration, known as SageMaker Pipelines.. It is ideal for advanced SageMaker users, especially those who are already onboarded to the IDE SageMaker Studio. The Studio also offers a UI to visual workflows built with SageMaker Pipelines. Apache Airflow is also a compelling option for ML workflow orchestration.

Orchestrate Hybrid ML Workloads with Kubeflow and EKS Anywhere

In this example we’re demonstrating training within local on-premises resources and orchestrating it using Kubeflow.

Stage Five: Other Services

Auxiliary services for hybrid ML patterns

AWS Outposts

Outposts is a keyway to enable hybrid experiences within your own data center. Order AWS Outposts, and Amazon will ship, install, and manage these resources for you. You can connect into these resources however you prefer and manage them from the cloud.

Outposts helps solve cases where customers want to build applications in countries where there is not currently an AWS Region, or for regulations that have strict data residency requirements, like online gambling and sports betting.

AWS Inferentia

A compelling reason to consider deploying your ML models in the cloud is the ease of accessing custom hardware for ML inferencing, specifically AWS Inferentia. You can use SageMaker’s managed deep learning containers to train your ML models, compile them for Inferentia with Neo, host on the cloud, and develop retrain and tune pipeline as usual. Using AWS Inferentia, Alexa was able to reduce their cost of hosting by 25%.

AWS Direct Connect

Ability to establish a private connection between your on-premises resources and your data center. Remember to establish a redundant link, as wires do go south!

Amazon ECS / EKS Anywhere

Both Amazon ECS and Amazon EKS feature “Anywhere” capabilities. This means that you can use the cloud as your control plane, to define, manage, and monitor your deployed applications, while executing tasks both in the Region and on-premises. The customers can use ECS Anywhere to deploy their models both in the cloud and on-premises at the same point in time!

The Final Stage: Use Cases

Hybrid ML Use Cases

Enterprise Migrations

One of the single most common use cases for hybrid patterns is enterprise migrations. For some of the largest and oldest organizations on the planet, without a doubt there is going to be a difference in ability and availability in moving towards the cloud across their teams.

Manufacturing

Applications within agriculture, industrial, and manufacturing are ripe opportunities for hybrid ML. After companies have invested tens, and sometimes hundreds, of thousands of dollars in advanced machinery, it is simply a matter of prudence to develop and monitor ML models to predict the health of that machinery.

Gaming

Customers who build gaming applications may see the value in adopting advanced ML services like Amazon SageMaker to raise the bar on their ML-applications but struggle to realize this if their entire platform was build and is currently hosted on premises. The AWS global delivery network to minimize end-user latency.

Mobile application development

With the introduction of AWS Wavelength, customers can deploy ML models directly inside of the 5G network. To solve applications such as anticipated network drop-off or hosting ML models in the cloud for real-time inferencing with 5G customers, you can use ML models hosted on ECS to deploy and monitor models onto AWS Wavelength. This becomes a hybrid pattern when customers develop and train in a secondary environment, wherever that may be, with the intention to deploy onto AWS Wave.

AI-enhanced media and content creation

Customers can host these billion-plus parameter models via ECS on AWS Local Zones, responding to application requests coming from on-premises data centers, to provide world-class experiences to content creators.

Depending on where customers develop and retrain their models, using Local Zones with SOTA models may or may not be a true hybrid pattern, but used effectively it can enhance content generator’s productivity and ability to create.

Autonomous Vehicles

Customers who develop autonomous machinery, vehicles, or robots in multiple capacity by default require hybrid solutions. This is because while training can happen anywhere, inference must necessarily happen at the edge.

Amazon SageMaker Neo is a compelling option here. Key questions for hybrid ML AV architectures include monitoring at the edge, retraining and retuning pipelines, in addition to efficient and automatic data labelling.

Conclusion

In this document, we explored hybrid ML patterns across the entire ML lifecycle. We looked at developing locally, while training and deploying in the cloud. We discussed patterns for training locally to deploy on the cloud, and even to host ML models in the cloud to serve applications on-premises.

At the end of the day, we want to support customer success in all shapes and forms. We firmly believe that most workloads will end in the cloud in the long run, but because the complexity, magnitude, and length of enterprise migrations may be daunting for some of the oldest organizations in the world, we propose these hybrid ML patterns as an intermediate step on customer’s cloud journey.

References:

If you are interested in learning how to migrate from local HDFS clusters to Amazon EMR, please see this migration guide: https://d1.awsstatic.com/whitepapers/amazon_emr_migration_guide.pdf
Original AWS Whitepaper

Amazon Connect Data Lake Best Practices | AWS Whitepaper Summary

Marwa Talaat — Sat, 24 Jul 2021 12:12:50 +0000

Nowadays, organizations are developing data lake strategies to harness intelligence from diverse and ever-growing data. According to Aberdeen, the survey concludes that today’s organizations manage an average of 33 unique data sources for analytics and experience 50% year-over-year data volume growth. Rapid data volume growth creates challenges in data management and storage capacity.
The traditional on-premises call center. The hardware communication, software, and infrastructure are all stored and operated in business places. Therefore, you need a dedicated IT team responsible for installation, maintenance, software support is controlled internally, and much more. his means heavy recurring fees over the long term for small and midsize businesses.
Therefore, the organizations need to get the most benefit from advanced analytics to robust platform and cost-effective to run a solution like a contact center. Amazon Web Services (AWS) provides customers with a comprehensive set of services and a scalable platform to ensure high availability, security, and resiliency of a data lake in the cloud.

This whitepaper outlines the best practices for architecting a contact center data lake with Amazon Connect.

The figure shows how it involved multiple proprietary systems, resulting in disparate data sources containing data in various formats. Therefore, there are challenges in standardizing and consolidating information that slows down the discovery of new business insights or possible operational issues.

As we know that data plays a crucial role in contact center systems success. So, having a streamlined designed contact center could help customer service agents to deliver a better customer experience by using data lake.

What is Data Lake?
It is more than just a data repository. It is centralized and allows you to store all your structured and unstructured data at any scale. You can store your data as-is, without structuring it, run analytics, big data processing, real-time analytics, and machine learning for predictive decisions.

Is there a difference between Data Lake and Data Warehouses? When to use them?
Yes, it depends on the requirements, a typical organization will require both a data warehouse and a data lake as they serve different needs and use cases. The following table shows some differences in characteristics between them.

Essential elements to consider in building Data Lakes:

1. Data Movement: It allows you to import any amount of data that can be real-time data from multiple sources and moved into the data lake. It also allows you to scale to data of any size, defining structures, schema, and transformations.
2. Securely store, and catalogue data: It allows you to store relational and non-relational. It also allows you to understand data through crawling, cataloguing, and indexing of data. Finally, data must be secured to ensure your data assets are protected.
3. Analytics: It allows various roles like data scientists to access data with their choice of analytic tools and frameworks
4. Machine Learning: It allows organizations to generate insights and doing machine learning models, predictions, and recommendations to achieve the optimal result.

What are the main challenges of Data Lake?

Architecture raw data is stored with no oversight of the contents
Making data usable, it needs to have defined mechanisms to catalogue, and secure data.

Should you host your data lake in the cloud?
Data Lakes are an ideal workload to be deployed in the cloud because the cloud provides performance, scalability, reliability, availability, a diverse set of analytic engines, and massive economies of scale.

Data Lakes in the cloud on AWS

AWS delivers the breadth and depth of services to build a secure, scalable, comprehensive, and cost-effective data lake solution. You can use the AWS services to ingest, store, find, process, and analyze data from a wide variety of sources. The following figure shows a strategic approach for a contact center with Amazon Connect.

Amazon Connect

It is a fully managed cloud-based and artificial intelligence (AI) enabled contact center within minutes, pay-as-you-go model which is easy-to-use and cost-effective and there is no infrastructure to manage or upfront costs.

Amazon Connect Benefits in Contact Center

Considerations when designing a data lake:

The way of collecting data, their types and structure.
The Data storage and share petabytes of data, on-demand globally and cost-effective
The scale IT resources to support a high number of concurrent queries against data.
The users view, search, and run queries on multiple data repositories.
The future insights using historical data patterns and past scenarios

Data lake data types

Some of the data types that Amazon Connect manages:
1. Customer profiles: Enables agents to deliver efficient and personalized customer service by importing customer information from various applications into a unified customer profile
2. Contact trace record: Captures transactional metrics such as hold time, wait time, and agent interaction time in JSON format. Amazon Connect aggregates data to create metrics reporting.
3. Contact flow logs: Captures real-time events and metrics about how your customers interact with contact flows
4. Contact Lens output files: Using NLP and speech-to-text analytics, Contact Lens for Amazon Connect provides insights to analyze customer sentiment, identify conversations trends for product feedback, and compliance audits for standard greetings and signoffs.
5. Agent events streams: Captures and stores agent activity in S3 via Amazon Kinesis Data Streams. You can create dashboards for near real-time agent reporting such as agent login, agent logout, agent connects with a contact and agent status change.
6. Voice and chat recordings: Amazon Connect records a conversation only when a customer connects to an agent. When the contact disconnects, the call recordings are available in your S3 bucket, or accessible in the customer's contact trace record.
7. Third-party integration: AWS Partners or other third-party solutions with Amazon Connect to consolidate logs and external data sources in Amazon S3.

Data lake lifecycle

Building a data lake typically involves five stages:
• Setting up storage: As your data lake grows, you need an object storage service like S3 that offers industry-leading scalability, data availability, security, and performance. S3s Storage Lens provides organization-wide visibility into object storage usage and activity trends with actionable recommendations to reduce cost and operational overhead.
• Moving data: AWS provides a comprehensive data transfer services portfolio to move your existing data into a centralized data lake. Amazon Storage Gateway and AWS Direct Connect can address hybrid cloud storage needs.
• Preparing and cataloguing data: Amazon Connect segregates data by AWS account ID and Amazon Connect instance ID to ensure authorized data access at the Amazon Connect instance level.
• Configuring security policies: You can create IAM users, groups, and roles in AWS accounts and associate them with identity-based policies that grant access to S3 resources.
• Making data available for consumption: Once your data lands in the S3 data lake, you can use any purpose-built analytics services such as Amazon Athena and Amazon QuickSight for a wide range of use cases without labour-intensive extract, transform, and load (ETL) job
The following figure is a high-level architecture diagram of an Amazon Connect contact center data lake

Conclusion and further reading

Amazon Connect is a purpose-built omnichannel cloud contact center. You can simplify operations, improve agent efficiency, and lower contact center costs. Amazon S3 is a scalable, durable, and reliable service to build and manage a secure data lake at scale for contact centers. You can store all your contact center data as-is in the S3 data lake without restructuring the data. Your employees and stakeholders can run various analytics on the contact center data lake, including big data processing, real-time dashboards and visualizations, and ML to guide data-driven business decisions. You can harness the power and unleash the intelligence of your contact center data lake to accelerate business growth.

For additional information, see:

• Original AWS Whitepaper
• Data Lake Storage on AWS
• Analytics on AWS
• On-Premises vs. Cloud-Based Call Center Software: How to Make the Call
• Data Lake vs Data Warehouse: What's the Difference?

DEV Community: Marwa Talaat

AWS Reinvent 2023: Unleashing the Power of ML and Generative AI

Study Plan to pass exam AWS Machine Learning Specialty exam with tips and advice

Whether or Not You Should follow This Guide?

Stage one: Testing the waters:

Stage Two: Self-based Learning Resources

Week- 1: Average 11 hours to 13 hours

Week- 2: Average 16 hours to 19 hours

Week- 3: Average 15 hours to 17 hours

Week- 4: Average 18 hours to 20 hours

Week- 5: Average 13 hours to 15 hours

Week- 6: Average 14 hours to 16 hours

Stage Three: Real live environment practice

Week- 7 & 8 & 9: Average 36 hours to 40 hours

Stage Four: Prepare for and take the AWS Certified Machine Learning - Specialty certification exam

Additional recommended resources

These tips for exam preparation:

Why do we love AWS Step Functions vs AWS Lambda in AI/ML?

What are AWS Step Functions?

AWS Step Function Benefits

How did AWS Step Functions build?

Drawbacks of Step functions

Step Functions limitation

AWS Step Functions vs AWS Lambda: what are the differences?

Conclusion

References:

Hybrid Machine Learning | AWS Whitepaper Summary

Stage One: Basics

What is hybrid?

What hybrid is not?

Stage Two: Hybrid patterns for development

Develop on personal computers, to train and host in the cloud

Develop on local servers, to train and host in the cloud

Stage Three: Training

Hybrid patterns for training

Training locally, to deploy in the cloud

How to monitor your model in the cloud?

How to handle retraining / retuning?

How to serve thousands of models in the cloud at low cost?

Storing data locally, to train and deploy in the cloud

Schedule data transfer jobs with AWS DataSync

Migrating from Local HDFS

Best practices

Develop in the cloud while connecting to data hosted on-premises

Data Wrangler & Snowflake

Train in the cloud, to deploy ML models on-premises

Stage Four: Deployment

Monitor ML models deployed on-premises with SageMaker Edge Manager

Hybrid patterns for deployment

Serve models in the cloud to applications hosted on-premises

Host ML Models with Lambda at Edge to applications on-premises

AWS Local Zones

AWS Wavelength

Training with a 3rd party SaaS provider to host in the cloud

Control plane patterns for hybrid ML

Orchestrate Hybrid ML Workloads with Kubeflow and EKS Anywhere

Stage Five: Other Services

Auxiliary services for hybrid ML patterns

AWS Outposts

AWS Inferentia

AWS Direct Connect

Amazon ECS / EKS Anywhere

The Final Stage: Use Cases

*Hybrid ML Use Cases *

Enterprise Migrations

Manufacturing

Gaming

Mobile application development

AI-enhanced media and content creation

Autonomous Vehicles

Conclusion

References:

Amazon Connect Data Lake Best Practices | AWS Whitepaper Summary

Data Lakes in the cloud on AWS

Amazon Connect

Amazon Connect Benefits in Contact Center

Data lake data types

Data lake lifecycle

Conclusion and further reading

For additional information, see:

Hybrid ML Use Cases