第13章ハイブリッドマルチクラウドのデジタルソリューション開発

Avatars are a popular way to represent ourselves graphically online.

Can avatars in the metaverse’s virtual-reality environment make use of big cloud PaaS capabilities to operate semi-autonomously? If they’re semi-autonomous, they’ll be able to act independently, with limited control from users. What ethical consequences could this pose?

Vendors provide various services to facilitate streaming big data analytics in the cloud, such as Spark Structured Streaming, Amazon Kinesis Data Firehose and Kinesis, Azure Stream Analytics, and Google Dataflow (based on Apache Beam). Spark Structured Streaming also provides another high-level concept called the DStream, which represents a continuous stream of data as a sequence of RDD fragments with windowed computations. Spark Structured Streaming leverages Spark Core and its fast scheduling engine to perform streaming analytics.

Amazon Firehose, which is designed for extreme scale, can load data directly into Amazon S3 or other Amazon services. Kinesis Data Analytics for SQL applications provides SQL-based tools for real-time analysis of streaming data from Kinesis Data Streams or Firehose. Kinesis Data Streams provides ordered, replayable, real-time streaming data. Various open-source frameworks, such as Kafka, Storm, RisingWave, Apache Spark, Apache Flume, Apache Beam, and Apache Flink, are also available to process streaming data on local machines. Table 13.2 compares Firehose to Kinesis Data Streams. The two primary components of the Azure Stream Analytics are the Azure Event Hubs service and Stream Analytics engine.

Apache Beam is the open-source release of the Google Cloud Dataflow system. Beam treats the batch and streaming cases uniformly and supports pipelines to encapsulate computations, as well as PCollections, which represent data as they move through a pipeline. Beam enables computational transformations that operate on PCollections, produces PCollections, and relies on sources and sinks, from which data are read and to which data are written, respectively.

All the big clouds today provide integrated machine learning services that use various techniques to create models that leverage prior experience and make it possible to improve the ability of machines to perform tasks. These services may be used as part of big data analytics and streaming big data analytics techniques, as explained in the previous subsections. Deep learning (DL) is another technique that leverages artificial neural networks (e.g., RNNs, CNNs, GANs) to create models that perform predictive tasks requiring special training and the ability to relate to a vast combination of labels/patterns (e.g., image recognition, speech recognition, language translation). There are various types of deep learning improvements that include reinforcement learning and transfer learning among others. Artificial intelligence (AI) is also a technique that leverages ML to enable computers to mimic human intelligence. Generative AI (GenAI) is a subset of AI that uses techniques such as deep learning and transformers to generate new content (e.g., create images, text, or audio) that matches a request. Transformers are special types of ML model architectures that are suited for solving problems that contain sequences such as text or time-series data.

Transformers have been used recently to solve natural language processing problems (e.g., translation, text generation, question answering, text summarization), and various transformer implementations have been quite successful, including the bidirectional encoder representations from transformers (BERT) and the generative pre-trained transformers (GPTs). ML models that support GenAI are referred to as large language models (LLMs) and/or foundation models (FMs). LLMs are typically tuned toward specific conversational applications and require more parameters and data-intensive training. Foundation models (FMs) are more general-purpose than LLMs and less data-intensive. Examples of LLMs include OpenAI’s ChatGPT, Google Gemini, Meta M2M-100 and LlaMA, IBM’s Granite model, Anthropic’s Claude models, Mistral AI’s models, and many others. LLMs (and FMs) can be augmented using retrieval augmented generation (RAG) AI frameworks that supplement the FMs internal representation of information to improve the quality of the LLM-generated responses.

Cloud vendors provide DL services in the form of programming frameworks that help implement deep learning applications using differentiable programming. GenAI services provided by the big clouds and other vendors are prompt interfaces that are specially engineered to allow users to get the most out of an LLM by including a sufficient amount of information in the prompts they create.

The following provides a nonexhaustive list of DL services provided by some of the big clouds:

• Amazon AWS deep learning services:

◦ Amazon Deep Learning AMIs (DLAMIs). Amazon DLAMIs are customized machine images that may be used for deep learning in the cloud. They can be deployed on various types of Amazon VMs (i.e., EC2 instances), including CPU-only instances or the latest high-powered multi-GPU instances. DLAMIs come preconfigured with NVIDIA CUDA and NVIDIA cuDNN and the latest releases of the most popular deep learning frameworks. Amazon released the Amazon Deep Learning AMI with Conda, which uses Conda virtual environments to isolate each framework, allowing you to switch between them at will without their dependencies conflicting. The full list of supported frameworks by Amazon Deep Learning AMI with Conda includes PyTorch, TensorFlow 2, and Apache MXNet (now retired but can still be accessed and used). Starting with the v18 release, Amazon Deep Learning AMI with Conda no longer includes the CNTK, Caffe, Caffe2, Theano, Chainer, or Keras Conda environments. Configuring the Amazon DLAMIs to use Jupyter is easy; go to the Amazon Marketplace on the EC2 portal and search for “deep learning.” You will find the DLAMIs, then select the server type you would like to use. If you simply want to experiment, it works well with a no-GPU option; when the VM comes up, log in with ssh, and configure Jupyter for remote access.

◦ Amazon Lex. Amazon Lex is an IA chatbot that allows users to incorporate voice input and conversational interfaces into applications. It is an extension of Amazon’s Echo product, a networked device with a speaker and microphone that you can ask questions through the Alexa service (e.g., questions about the weather, event scheduling, news, and music). It is possible to associate Echo voice commands with the launching of an Amazon Lambda function that executes a cloud application.

◦ Amazon Polly. Amazon Polly is the opposite of Lex; it turns text into speech for dozens of languages with a variety of voices and uses the Speech Synthesis Markup Language (SSML) to control pronunciation and intonation.

◦ Amazon Rekognition. Amazon Rekognition is at the cutting edge of deep learning applications. It takes an image as input and returns a textual description of the items that it sees in that image. This includes objects, landmarks, dominant colors, activities, and faces. It also performs detailed facial analysis and comparisons and can identify inappropriate content that appears in images.

• Microsoft Azure deep learning services:

◦ Azure Data Science VMs (DSVMs). Azure DSVMs are Azure Virtual Machine images, preinstalled, configured, and tested with several popular tools that are commonly used for data analytics, machine learning, and AI development and training.

◦ Azure Machine Learning (ML). Azure ML is a cloud service that is designed to help accelerate and manage machine learning project life cycles. It can be used to train and deploy ML models and manage machine learning operations (via MLOps). You can create a model in Microsoft ML or use a model built from an open-source platform, such as PyTorch, TensorFlow, or scikit-learn. MLOps tools help you monitor, retrain, and redeploy models. As noted earlier, Azure also provides the ML.NET machine learning framework.

◦ Azure AI services. Azure AI services are APIs/SDKs that can be used to build applications that support natural methods of communication (i.e., see, hear, speak, understand, and interpret user needs). These services include support for vision (e.g., object detection, face recognition, optical character recognition), speech (e.g., speech-to-text, text-to-speech, speaker recognition), languages (e.g., translation, sentiment analysis, key phrase extraction, language understanding), and decision (e.g., anomaly detection, content moderation, reinforcement learning).

• Google deep learning services:

◦ Deep learning VMs. Deep learning VM images are virtual machine images optimized for data science and machine learning tasks. All images include preinstalled ML frameworks and tools and can be used on VM instances with GPUs to accelerate data processing tasks. ML frameworks supported include TensorFlow and PyTorch.

◦ Google machine learning APIs. Google provides various APIs to services that can be used to build applications that support natural methods of communication, including Cloud Vision to understand the content of an image, Cloud Speech-to-Text to transcribe audio to text, Cloud Translation to translate an arbitrary string to any supported language, and Cloud Natural Language to extract information from text.

The following provides more information related to GenAI services provided by the big clouds:

• AWS GenAI services. Amazon AWS provides various GenAI tools, including the Amazon Q AI-powered assistance and the Amazon Bedrock suite of LLMs, FMs, and generative AI tools. Amazon SageMaker may be used to build, train, and deploy FM models at scale.

• Microsoft Azure GenAI services. The Azure OpenAI service and the Azure AI studio can be used to create custom copilot and generative AI applications. Microsoft has partnered with OpenAI, the company that is developing ChatGPT. It also provides the Phi family of small language models (SLMs) that are low-cost and low-latency alternatives to LLMs in some cases.

• Google GCP GenAI services. Vertex AI, Generative AI Studio, and Vertex AI Model Garden are various solutions that Google provides to support the creation of generative AI applications. Google also provides the Gemini family of generative AI models that are capable of processing information from multiple modalities, including images, videos, and text.

• IBM Cloud GenAI services. IBM Watsonx.ai AI studio brings together generative AI capabilities that are powered by FMs and ML. It provides tools to tune and guide models based on enterprise data as well as build and refine prompts. IBM also develops custom Granite AI foundation models that are cost-efficient and enterprise-grade.

• Other GenAI services. In addition to the GenAI services and tools mentioned here, many other vendors focus on the creation of LLMs, SLMs, and FMs. Here are a few of them:

◦ OpenAI’s ChatGPT

◦ Meta LlaMA

◦ Anthropic Claude

◦ Mistral AI