Ai Aws Sagemaker | AI Final Year Project Report

Complete Technical Paper Content — Ai Aws Sagemaker

1. Executive Abstract, Problem Formulation & Background

Amazon SageMaker AI is a cloud-based machine-learning platform that allows the creation, training, and deployment by developers of machine-learning (ML) models on the cloud. It can be used to deploy ML models on embedded systems and edge-devices. The platform was launched in November 2017.

2017-11-29: SageMaker is launched at the AWS re:Invent conference. 2018-02-27: Managed TensorFlow and MXNet deep neural network training and inference are now supported within SageMaker. 2018-02-28: SageMaker automatically scales model inference to multiple server instances. 2018-07-13: Support is added for recurrent neural network training, word2vec training, multi-class linear learner training, and distributed deep neural network training in Chainer with Layer-wise Adaptive Rate Scaling (LARS). 2018-07-17: AWS Batch Transform enables high-throughput non-real-time machine learning inference in SageMaker. 2018-11-08: Support for training and inference of Object2Vec word embeddings. 2018-11-27: SageMaker Ground Truth "makes it much easier for developers to label their data using human annotators through Mechanical Turk, third-party vendors, or their own employees." 2018-11-28: SageMaker Reinforcement Learning (RL) "enables developers and data scientists to quickly and easily develop reinforcement learning models at scale." 2018-11-28: SageMaker Neo enables deep neural network models to be deployed from SageMaker to edge-devices such as smartphones and smart cameras. 2018-11-29: The AWS Marketplace for SageMaker is launched. The AWS Marketplace enables 3rd-party developers to buy and sell machine learning models that can be trained and deployed in SageMaker. 2019-01-27: SageMaker Neo is released as open-source software.

Amazon Web Services, Inc. (AWS) is a subsidiary of Amazon that provides on-demand cloud computing platforms and APIs to individuals, companies, and governments, on a metered, pay-as-you-go basis. Clients often use this in combination with autoscaling (a process that allows a client to use more computing in times of high application usage, and then scale down to reduce costs when there is less traffic). These cloud computing web services provide various services related to networking, compute, storage, middleware, IoT and other processing capacity, as well as software tools via AWS server farms. This frees clients from managing, scaling, and patching hardware and operating systems. One of the foundational services is Amazon Elastic Compute Cloud (EC2), which allows users to have at their disposal a virtual cluster of computers, with extremely high availability, which can be interacted with over the Internet via REST APIs, a CLI or the AWS console. AWS's virtual computers emulate most of the attributes of a real computer, including hardware central processing units (CPUs) and graphics processing units (GPUs) for processing; local/RAM memory; hard-disk (HDD)/SSD storage; a choice of operating systems; networking; and pre-loaded application software such as web servers, databases, and customer relationship management (CRM). AWS services are delivered to customers via a network of AWS server farms located throughout the world. Fees are based on a combination of usage (known as a "Pay-as-you-go" model), hardware, operating system, software, and networking features chosen by the subscriber requiring various degrees of availability, redundancy, security, and service options. Subscribers can pay for a single virtual AWS computer, a dedicated physical computer, or clusters of either. Amazon provides select portions of security for subscribers (e.g. physical security of the data centers) while other aspects of security are the responsibility of the subscriber (e.g. account management, vulnerability scanning, patching). AWS operates from many global geographical regions, including nine in North America. Amazon markets AWS to subscribers as a way of obtaining large-scale computing capacity more quickly and cheaply than building an actual physical server farm. All services are billed based on usage, but each service measures usage in varying ways. As of 2023 Q1, AWS has 31% market share for cloud infrastructure while the next two competitors Microsoft Azure and Google Cloud have 25%, and 11% respectively, according to Synergy Research Group. AWS was the first cloud provider to build infrastructure specifically for U.S. government with respect to security and compliance requirements (2011; launched AWS GovCloud), accredited to support classified workloads (2014; launched AWS Top Secret), and achieve accreditation across all U.S. government data classifications (2017; launched AWS Secret).

2. Related Work, Literature Survey & Algorithmic Foundations

As of 2025 AWS comprises over 200 products and services including computing, storage, networking, database, analytics, application services, deployment, management, machine learning, mobile, developer tools, RobOps and tools for the Internet of Things. The most popular include Amazon Elastic Compute Cloud (EC2), Amazon Simple Storage Service (Amazon S3), Amazon Connect, and AWS Lambda (a serverless function that can perform arbitrary code written in any language that can be configured to be triggered by hundreds of events, including HTTP calls). Services expose functionality through APIs for clients to use in their applications. These APIs are accessed over HTTP, using the REST architectural style and SOAP protocol for older APIs and exclusively JSON for newer ones. Clients can interact with these APIs in various ways, including from the AWS console (a website), by using SDKs written in various languages (such as Python, Java, and JavaScript), or by making direct REST calls.

The genesis of AWS came in the early 2000s. After building Merchant.com, Amazon's e-commerce-as-a-service platform that offers third-party retailers a way to build their own web-stores, Amazon pursued service-oriented architecture as a means to scale its engineering operations, led by then CTO Allan Vermeulen. Around the same time frame, Amazon was frustrated with the speed of its software engineering, and sought to implement various recommendations put forth by Matt Round, an engineering leader at the time, including maximization of autonomy for engineering teams, adoption of REST, standardization of infrastructure, removal of gate-keeping decision-makers (bureaucracy), and continuous deployment. He also called for increasing the percentage of the time engineers spent building the software rather than doing other tasks. Amazon created "a shared IT platform" so its engineering organizations, which were spending 70% of their time on "undifferentiated heavy-lifting" such as IT and infrastructure problems, could focus on customer-facing innovation instead. Besides, in dealing with unusual peak traffic patterns, especially during the holiday season, by migrating services to commodity Linux hardware and relying on open source software, Amazon's Infrastructure team, led by Tom Killalea, Amazon's first CISO, had already run its data centers and associated services in a "fast, reliable, cheap" way. In July 2002 Amazon.com Web Services, managed by Colin Bryar, launched its first web services, opening up the Amazon.com platform to all developers. Over one hundred applications were built on top of it by 2004. This unexpected developer interest took Amazon by surprise and convinced them that developers were "hungry for more". By the summer of 2003, Andy Jassy had taken over Bryar's portfolio at Rick Dalzell's behest, after Vermeulen, who was Bezos' first pick, declined the offer. Jassy subsequently mapped out the vision for an "Internet OS" made up of foundational infrastructure primitives that alleviated key impediments to shipping software applications faster. By fall 2003, databases, storage, and compute were identified as the first set of infrastructure pieces that Amazon should launch. Jeff Barr, an early AWS employee, credits himself, Vermeulen, Jassy, Bezos, and a few others for coming up with the idea that would evolve into EC2, S3, and RDS; Jassy recalls the idea was the result of brainstorming for about a week with "ten of the best technology minds and ten of the best product management minds" on about ten different Internet applications and the most primitive building blocks required to build them. Werner Vogels cites Amazon's desire to make the process of "invent, launch, reinvent, relaunch, start over, rinse, repeat" as fast as it could was leading them to break down organizational structures with "two-pizza teams" and application structures with distributed systems; and that these changes ultimately paved way for the formation of AWS and its mission "to expose all of the atomic-level pieces of the Amazon.com platform". According to Brewster Kahle, co-founder of Alexa Internet, which was acquired by Amazon in 1999, his start-up's compute infrastructure helped Amazon solve its big data problems and later informed the innovations that underpinned AWS. Jassy assembled a founding team of 57 employees from a mix of engineering and business backgrounds to kick-start these initiatives, with a majority of the hires coming from outside the company. They included Jeff Lawson, the Twilio CEO; Adam Selipsky, the Tableau CEO; and Mikhail Seregine, a co-founder at Outschool. In late 2003, the concept for compute, which would later launch as EC2, was reformulated when Chris Pinkham and Benjamin Black presented a paper internally describing a vision for Amazon's retail computing infrastructure that was completely standardized, completely automated, and would rely extensively on web services for services such as storage and would draw on internal work already underway. Near the end of their paper, they mentioned the possibility of selling access to virtual servers as a service, proposing the company could generate revenue from the new infrastructure investment. Thereafter Pinkham, Willem van Biljon, and lead developer Christopher Brown developed the Amazon EC2 service, with a team in Cape Town, South Africa. In October 2004, Amazon released Alexa Web Information Service (AWIS) as beta. In November 2004, AWS launched its first infrastructure service, as beta: Simple Queue Service (SQS)., and officially launched in July 2006.

On March 14, 2006, AWS launched Amazon S3 cloud storage, and that was the first AWS' IaaS that was generally available, followed by EC2 in August 2006. Pi Corporation, a startup Paul Maritz co-founded, was the first beta-user of EC2 outside of Amazon, while Microsoft was among EC2's first enterprise customers. Later that year SmugMug, one of the early AWS adopters, attributed savings of around US$400,000 in storage costs to S3. According to Vogels, S3 was built with 8 microservices when it launched in 2006 and had over 300 microservices by 2022. In September 2007, AWS announced its annual Start-up Challenge, a contest with prizes worth $100,000 for entrepreneurs and software developers in the US using AWS services such as S3 and EC2 to build their businesses. The first edition saw participation from Justin.tv, which Amazon later acquired in 2014. Ooyala, an online media company, was the eventual winner. AWS offers, as of June 16, 2022, two block-storage options: the EC2 Instance Store and the Elastic Block Store (EBS). Some Amazon EBS features that help with data management, backups, and performance tuning include:

3. Dataset Ingestion, Preprocessing & Stochastic Augmentations

EBS volume tagging to allow the user to find and filter EBS resources on the Amazon Console and CLI. Software-level RAID arrays to enable creation of groups of EBS volumes with high performance network throughput between them, using the standard RAID protocol. Additional AWS services from this period include SimpleDB, Mechanical Turk, Elastic Beanstalk, Relational Database Service, DynamoDB, CloudWatch, Simple Workflow, CloudFront, and Availability Zones.

In November 2010, it was reported that all of Amazon.com's retail sites had migrated to AWS. Prior to 2012, AWS was considered a part of Amazon.com and so its revenue was not delineated in Amazon financial statements. In that year industry watchers for the first time estimated AWS revenue to be over $1.5 billion. On November 27, 2012, AWS hosted its first major annual conference, re:Invent with a focus on AWS's partners and ecosystem, with over 150 sessions. The three-day event was held in Las Vegas because of its relatively cheaper connectivity with locations across the United States and the rest of the world. Andy Jassy and Werner Vogels presented keynotes, with Jeff Bezos joining Vogels for a fireside chat. AWS opened early registrations at US$1,099 per head for their customers from over 190 countries. On stage with Andy Jassy at the event which saw around 6000 attendees, Reed Hastings, CEO at Netflix, announced plans to migrate 100% of Netflix's infrastructure to AWS. To support industry-wide training and skills standardization, AWS began offering a certification program for computer engineers, on April 30, 2013, to highlight expertise in cloud computing. Later that year, in October, AWS launched Activate, a program for start-ups worldwide to leverage AWS credits, third-party integrations, and free access to AWS experts to help build their business. In 2014, AWS launched its partner network, AWS Partner Network (APN), which is focused on helping AWS-based companies grow and scale the success of their business with close collaboration and best practices. In October 2014, Amazon Aurora was launched. In January 2015, Amazon Web Services acquired Annapurna Labs, an Israel-based microelectronics company for a reported US$350–370M. In April 2015, Amazon.com reported AWS was profitable, with sales of $1.57 billion in the first quarter of the year and $265 million of operating income. Founder Jeff Bezos described it as a fast-growing $5 billion business; analysts described it as "surprisingly more profitable than forecast". In October, Amazon.com said in its Q3 earnings report that AWS's operating income was $521 million, with operating margins at 25 percent. AWS's 2015 Q3 revenue was $2.1 billion, a 78% increase from 2014's Q3 revenue of $1.17 billion. 2015 Q4 revenue for the AWS segment increased 69.5% y/y to $2.4 billion with a 28.5% operating margin, giving AWS a $9.6 billion run rate. In 2015, Gartner estimated that AWS customers are deploying 10x more infrastructure on AWS than the combined adoption of the next 14 providers.

On April 20, 2011, AWS suffered a major outage. Parts of the Elastic Block Store service became "stuck" and could not fulfill read/write requests. It took at least two days for the service to be fully restored. On June 29, 2012, several websites that rely on Amazon Web Services were taken offline due to a severe storm in Northern Virginia, where AWS's largest data center cluster is located. On October 22, 2012, a major outage occurred, affecting many sites including Reddit, Foursquare, and Pinterest. The cause was a memory leak bug in an operational data collection agent. On December 24, 2012, AWS suffered another outage causing websites such as Netflix to be unavailable for customers in the Northeastern United States. AWS cited their Elastic Load Balancing service as the cause. On February 28, 2017, AWS experienced a massive outage of S3 services in its Northern Virginia region. A majority of websites that relied on AWS S3 either hung or stalled, and Amazon reported within five hours that AWS was fully online again. No data has been reported to have been lost due to the outage. The outage was caused by a human error made while debugging, that resulted in removing more server capacity than intended, which caused a domino effect of outages. On November 25, 2020, AWS experienced several hours of outage on the Kinesis service in North Virginia (US-East-1) region. Other services relying on Kinesis were also affected. On December 7, 2021, an outage mainly affected the Eastern United States, disrupting delivery service and streaming. On October 20, 2025, AWS suffered a major outage, resulting in platforms including Duolingo, Snapchat, Canva, Reddit, Canvas, Coinbase, Roblox, Fortnite, and Amazon being down. From March 1, 2026, UAE and Bahrain regions are unavailable following attacks by the Islamic Revolutionary Guard Corps.

4. Neural Network Backbone, Layer Specifications & Loss Formulations

AWS also has "pop-up lofts" in different locations around the world. These market AWS to entrepreneurs and startups in different tech industries in a physical location. Visitors can work or relax inside the loft, or learn more about what they can do with AWS. In June 2014, AWS opened their first temporary pop-up loft in San Francisco. In May 2015 they expanded to New York City, and in September 2015 expanded to Berlin. AWS opened its fourth location, in Tel Aviv from March 1, 2016, to March 22, 2016. A pop-up loft was open in London from September 10 to October 29, 2015. The pop-up lofts in New York and San Francisco are indefinitely closed due to the COVID-19 pandemic while Tokyo has remained open in a limited capacity.

The contract for Project Nimbus drew rebuke and condemnation from the companies' shareholders as well as their employees, over concerns that the project would lead to abuses of Palestinians' human rights in the context of the ongoing occupation and the Israeli–Palestinian conflict. Specifically, they voice concern over how the technology will enable further surveillance of Palestinians and unlawful data collection on them as well as facilitate the expansion of Israel's illegal settlements on Palestinian land. A government procurement document featuring 'obligatory customers' of Nimbus, including "two of Israel’s leading state-owned weapons manufacturers" Israel Aerospace Industries and Rafael Advanced Defense Systems, was published in 2021 with periodic updates since (up to Oct 2023).

Some AWS customers have complained about receiving unexpectedly large bills, commonly referred to as "surprise bills." This can occur due to various reasons, including but not limited to misconfigurations, security breaches, complex pricing—especially when multiple AWS services are used together—and unexpected data transfer charges.

5. Training Protocol, Convergence Behavior & Experimental Evaluation

Cloud-computing comparison Comparison of file hosting services James Gosling – Canadian computer scientist (born 1955) Tim Bray – Canadian software developer (born 1955)

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.