Why You Need to Understand Amazon SageMaker Before Discussing AI Platforms

[Key Points]

1. Amazon SageMaker provides a well-equipped “central kitchen” for machine learning developers, allowing them to start cooking (training models) without having to build the kitchen themselves, as long as they bring their own ingredients (training data).

2. Through Amazon SageMaker, one can glimpse several development trends in machine learning platforms: the increasing importance of openness in machine learning platforms, the rise of MLOps, and machine learning security.

3. In the chaotic landscape of AI platforms in China, Amazon SageMaker’s entry into the Chinese market will lead to win-win cooperation with AI players in the entire industry.

The smallest combat unit of the US military evolved from battalions during World War II to small squads of several people or a dozen people during the Afghan War, thanks to a powerful missile command system in the rear. This may be the most straightforward analogy for “big middle platform, small front platform”.

The concept of a middle platform was proposed by domestic tech companies, while Gartner’s “Packaged Business Capability” refers to encapsulated business capabilities, which can be seen as more concrete suggestions. As AI applications gradually increase in enterprise R&D, AI middle platforms have become exceptionally popular in China, with many internet giants successively launching their own AI middle platforms.

The AI middle platform is the infrastructure for building large-scale intelligent services, a complete artificial intelligence model lifecycle management platform and service system that provides capabilities such as model design training, model/algorithm libraries, reuse labeling management, and model monitoring services. The AI middle platform aims to enable enterprise business fronts to engage in quick, small-scale battles and reduce trial-and-error costs, making it easier to leverage AI to accelerate business innovation.

In fact, what the AI middle platform aims to do has a mature precedent in the industry—Amazon SageMaker. On April 30, 2020, Amazon SageMaker officially entered the Chinese market, opening in the AWS China (Beijing) region operated by Guanghuan Xinnet and the AWS China (Ningxia) region operated by Western Cloud Data. Recently, Dr. Zhang Xia, Chief Cloud Computing Enterprise Strategy Consultant at AWS, had an in-depth communication with domestic media.

There has been much discussion in the industry regarding the basic functionalities of Amazon SageMaker. “Intelligent Evolution Theory” focuses on what makes Amazon SageMaker exceptional. In the chaotic landscape of AI middle platforms in China, is Amazon SageMaker’s entry into the Chinese market a case of ‘the wolf is coming’ or will it accelerate the industry’s landing?

Time Zone Differences Yielding Machine Learning Platform Services

In 2015, AWS Vice President Swami Sivasubramanian, who was responsible for artificial intelligence and machine learning, celebrated his 10th anniversary at Amazon with a special four-week vacation. When Swami returned to his hometown in India from Seattle for vacation, he suffered from insomnia due to time zone differences.

During those sleepless nights, Swami studied the development of artificial intelligence and realized the huge gap between machine learning (ML) and enterprise applications. “Machine learning is not something every enterprise or institution can master; building algorithm models is difficult, and we want to make it easier for enterprises to use machine learning,” Swami said in an interview.

Setting aside concepts like AI and middle platforms, AWS is more inclined toward the implementation of technical solutions. Amazon SageMaker is positioned as AWS’s machine learning platform service. In Gartner’s 2020 Magic Quadrant for Cloud AI Developer Services, AWS was rated as a leader, thanks in large part to Amazon SageMaker. Currently, 85% of TensorFlow workloads globally run on AWS. Since its release in November 2017, Amazon SageMaker has been chosen by tens of thousands of enterprises worldwide to run machine learning workloads, with the number of machine learning customers on AWS exceeding twice that of all other cloud vendors combined.

Amazon SageMaker is a fully managed service, with its core users being algorithm engineers and data scientists with certain machine learning development capabilities. Amazon SageMaker can help them quickly build, train, and deploy models in the cloud.

Within the overall framework of AWS machine learning services, Amazon SageMaker is positioned in the middle layer. The underlying layer provides the widest range of infrastructure, including mainstream machine learning frameworks, open-source artificial neural network libraries like Keras, Linux images like Amazon Linux AMI, and various computing capabilities. The upper layer features artificial intelligence services tailored for specific business domains, all of which have been successfully validated by Amazon’s commercial system. For example, the core of the smart speaker Alexa—the Amazon human-machine dialogue engine Lex, the intelligent recommendation capabilities outputting products for Amazon’s e-commerce business—Personalize, and the financial transaction anti-fraud tool Amazon Fraud Detector.

Central Kitchen for Machine Learning Developers

The goal of Amazon SageMaker is to let the wisdom, time, and energy of algorithm teams focus on what matters. The workflow of machine learning is exceptionally complex and time-consuming, involving data preparation, selecting and optimizing machine learning frameworks and algorithms, setting up training environments, training and tuning models, and deploying and monitoring models, among others.

If we compare algorithm engineers and data scientists to chefs, then to create a delicious dish (an algorithm model that can be applied in an enterprise environment), they spend over 90% of their time on tasks they are not good at, such as “collecting firewood, building the stove, and casting the iron pot”.

Amazon SageMaker essentially provides developers with a complete “central kitchen” equipped with a rich variety of kitchenware, equipment, and semi-finished ingredients (machine learning services and pre-built algorithms), allowing developers to start cooking (training models) as long as they bring their own ingredients (training data).

It can be said that Amazon SageMaker has become a model in the industry for the completeness and continuity of machine learning services. Amazon SageMaker’s functional components cover the entire workflow of machine learning, significantly reducing the difficulty of model building and training, while accelerating the model training process.

Among them, the integrated development environment (IDE) for machine learning, Amazon SageMaker Studio, provides a unified working interface for all Amazon SageMaker features and the entire machine learning workflow. Amazon SageMaker Notebooks allows developers to easily enable Jupyter Notebook with one click for sharing and collaboration. Amazon SageMaker Autopilot, an automatic model building component, allows users with no machine learning experience to easily generate data-based models, while experienced developers can quickly develop foundational models with it. In addition, Amazon SageMaker provides rich components for experiment management, debugging and analyzing models, detecting and correcting concept drift, etc.

Trends in Machine Learning Platforms Observed Through Amazon SageMaker

Development Trends

Training machine learning models is extremely complex, expensive, and time-consuming, which has become a major obstacle to the large-scale implementation of AI. This has been recognized by top AI companies worldwide.

In 2018, Google’s Fei-Fei Li team launched the Cloud AutoML platform, while Baidu introduced the EasyDL platform. These platforms provide zero-threshold AI development capabilities, addressing some of the most common and basic scenario needs in the field of machine learning. In other words, even those with no coding experience can train an AI model by simply uploading data, effectively democratizing AI open platforms.

However, when it comes to enterprise-level production machine learning models, a certain level of machine learning programming ability is still required. Otherwise, it is difficult to explain why many enterprises find it hard to recruit high-end AI talent even with million-dollar annual salaries. Moreover, the boundaries between algorithm engineers and data scientists are becoming increasingly blurred, with many data experts and DBAs striving to learn Python, indicating that machine learning programming skills are becoming more important.

Despite the continuous emergence of AI open platforms and AI middle platforms in China, Amazon SageMaker remains the most commercially mature machine learning platform service globally. Analyzing the differentiated advantages of Amazon SageMaker also reflects several major development trends in machine learning platforms:

Trend One: Openness of Machine Learning Platforms, including openness to machine learning frameworks, integration of functional components, ecosystems, and open source communities.

Openness to Machine Learning Frameworks. Amazon SageMaker has deeply optimized mainstream frameworks including TensorFlow, PyTorch, and Apache MXNet. For example, under 256 GPUs, the scaling efficiency for TensorFlow can reach 90%, far exceeding the 65% of open-source platforms. Even earlier frameworks like Caffe, or frameworks developed by enterprise clients themselves, can be used in Amazon SageMaker as long as they are brought in through custom containers.

Openness to Integration. Although Amazon SageMaker is powerful and in-depth, its modular design approach allows enterprises to have great flexibility in their choices. Enterprises can use only some features as needed. For example, Company A can choose to train a model in Amazon SageMaker and apply inference at the edge, while Company B can bring a locally trained model to Amazon SageMaker for inference, experimentation, monitoring, etc., to save computing costs.

Openness of the Ecosystem. The vitality of a machine learning platform largely depends on its ecosystem. In foreign markets, AWS Marketplace provides a model and algorithm market for SageMaker, with hundreds of partners having uploaded their self-developed AI models and algorithms for clients to choose from. In the domestic market, AWS has developed leading partners such as Deloitte, Chinasoft International, and Eclode to accelerate the landing of Amazon SageMaker in the enterprise market.

Openness to Open Source Communities. Considering that many machine learning workloads are already containerized, Amazon SageMaker provides operators for Kubernetes and maintains openness to numerous components of open-source communities. Additionally, AWS has open-sourced parts of SageMaker’s implementation and teamed up with Facebook to launch TorchServe—the first model server specifically for the PyTorch framework.

Trend Two: MLOps Takes Center Stage

As machine learning evolves and enterprise application scenarios increase, integrating machine learning workloads into enterprise automation workflows has become an urgent issue to address.

The model building of machine learning belongs to development, model training belongs to offline workloads, and real-time inference of models belongs to online businesses. Moreover, the environments related to machine learning differ significantly from ordinary applications or big data programs in terms of computing resources, release iteration processes, and environment monitoring. This entire process is currently in an early stage in many enterprises, even in technology-driven internet companies, gradually becoming a resistance to the landing, scaling, and production of machine learning.

Amazon SageMaker has transformed many manual processes in model building, training tuning, and deployment management into automation, while providing rich interfaces for secondary development. Furthermore, it allows internal operations platform teams, business application development teams, and machine learning algorithm teams to achieve loose coupling in their workflows, significantly accelerating the iteration and landing process of machine learning workloads.

Trend Three: Machine Learning Security

Currently, most enterprises are still learning how to efficiently develop and utilize machine learning, and security issues have not received widespread attention. Some common training operations, such as placing the model training data of different project teams on a single local machine without isolation, pose significant security risks.

Amazon SageMaker provides dedicated resources for each customer as needed for training and offers security protection throughout the entire model training process. For instance, Amazon SageMaker Notebooks can separate programming from training, not only saving foundational technical resources but also achieving data security isolation.

Amazon SageMaker’s Entry into China: The Wolf Is Coming?

“AWS currently has 24 regions globally, with the Beijing and Ningxia regions in China being the fifth and sixth regions where Amazon SageMaker has landed,” Zhang Xia revealed. The importance AWS places on the Chinese market is underpinned by the rapid development of the AI field in China.

Zhang Xia emphasized that AWS’s relationship with many players in China’s AI industry is not competitive but rather a cooperative relationship aimed at mutual benefit. As a technology-empowered machine learning platform, Amazon SageMaker has established extensive collaborations with ecological partners ranging from startups to large enterprises in the AI industry, with the goal of helping more Chinese enterprises apply machine learning in production.

Since Amazon SageMaker is aimed at professional algorithm engineers, for some enterprises that do not have AI teams and capabilities, partners become a key link in bridging the last mile of AI implementation. Based on the Amazon SageMaker technology platform, partners can tailor AI solutions that truly address the actual problems faced by enterprises.

As an AWS core consulting partner (APN Premier Consulting Partner), Eclode’s Vice President in China, Gui Zijie, told the media: “We are already using the Amazon SageMaker platform to accelerate enterprises’ adoption of industry AI solutions, such as labeling, text analysis, semantic understanding, predictive classification, recommendation systems, and fraud detection, creating truly problem-solving end-to-end AI applications tailored to the actual business challenges faced by clients.”

Dayu Infinite is a company specializing in mobile application development, primarily providing mobile short video services for emerging markets such as the Middle East, Southeast Asia, and Latin America. To enhance user experience, Dayu Infinite aimed to implement personalized video content recommendations in their short video app, but their development team faced significant challenges in training AI models. Technical Vice President Liu Kedong stated: “Amazon SageMaker allows us to build infrastructure without the need for complex setups; our algorithm engineers only needed to prepare data, and within three months, we completed the entire system construction and handled actual user traffic, saving 70% of our training costs with Amazon SageMaker.”

Looking at the industry landscape, more and more enterprises are launching AI middle platforms. AWS’s strategic approach with Amazon SageMaker undoubtedly provides valuable reference for current AI middle platform players.

Currently, AWS offers exclusive benefits for machine learning customers,

Scan the QR code below or click Read Original to access the application entry for the China region

(Register to receive a 1000 RMB service discount coupon)

Scan the QR code below to access the application entry for the international region

(Register to receive a 200 USD service discount coupon)

END

This article is an original work of “Intelligent Evolution Theory”,

Welcome to follow.

For communication and cooperation, please add WeChat: abcde363636

Intelligent Evolution Theory

Decoding the Technological Codes Behind Business Evolution