July 2020 AWS highlights of AI content


July 2020 AWS highlights of AI content

If you want to stay in the air-conditioned room without going anywhere, you may as well take this opportunity to learn technology. Maybe it can help you solve the difficult problems in your work. In July 2020, we continued to share with you a large number of technical articles on Amazon Web services (AWS), such as AI, machine learning, and deep learning. It’s time to review them briefly.

Amazon SageMaker

In April 2020, Amazon sagemaker has been officially opened in AWS China (Beijing) region operated by halo new network and AWS China (Ningxia) region operated by Xiyun data. As a machine learning platform service, Amazon sagemaker makes it easy for data scientists, Algorithm Engineers and business developers to master machine learning by constantly enriching functional components. How powerful is it? What are the successful cases of Amazon sagemaker? Welcome to read: with the advent of machine learning era, they use Amazon sagemaker to do this!

Machine learning workflow is an iterative process, which needs to make many decisions, and all decisions will affect the results of the learning system. In addition, we also need to combine with the actual needs, choose the most suitable learning system, supervised learning or unsupervised learning? Or semi supervised learning? Intensive learning? The whole process involves too many technologies and decisions, which often confuses the novice. May as well through an example, to fully understand the machine learning workflow in the end how to build it. Welcome to read: which restaurants are rated the most? Learn about the construction of machine learning workflow.

machine learning

Ernie (enhanced representation through knowledge integration) is a semantic understanding framework, which aims to capture semantic patterns with pure text, and to provide richer structural representation by knowledge enhancement and knowledge relations in semantics. The framework has won five World Championships in semeval 2020, the largest semantic evaluation competition in the world recently, including key text fragment mining in visual media, multilingual aggressive language detection and sentiment analysis of mixed languages! Even better, Amazon sagemaker has provided full support for enrie 2.0! Further understanding: Ernie | “the best” semantic understanding framework is fully supported by AWS

Creating a reliable and efficient machine learning reasoning service requires a lot of investment, such as creating applications, packaging and deployment, loading models, running terminal nodes Many of them need complex development and operation and maintenance tasks. AWS recently released Amazon sagemaker operator for kubernetes, which can help developers enhance the existing kubernetes cluster with terminal nodes hosted by sagemaker, and greatly simplify the burden and difficulty of related operation and maintenance work. Welcome to read: machine learning reasoning on kubernetes can be further enhanced in this way.

Create one or several ml models according to the business requirements, and then obtain the decision-making basis required by the business according to the reasoning of the models. This method of building models based on groups or market segments must be familiar to all of us. However, in some scenarios, we may need to build models based on the data of individual users in order to obtain more accurate results, but this practice often leads to the cost of deploying the models greatly, and in the production environment, so many models become difficult to manage. Amazon sagemaker multi model terminal node function solves this problem. Welcome to read: what kind of experience is it to create a unique ML model for each user?

Users who have used the AWS EC2 instance may know that EC2 has a spot instance, which can provide users with EC2 capacity not used in AWS temporarily. Compared with the on-demand instance, EC2 can enjoy a price discount of up to 90%. Similarly, Amazon sagemaker recently started offering managed spot training. This is a new feature based on Amazon EC2 spot instance, which can save up to 90% ml training cost compared with using on-demand instance in Amazon sagemaker. Further understanding: using spot instance to train ML model, who said that the training cost can not be reduced!

Part of the reason why the process of machine learning is so difficult is that there are many best practices in the field known to experienced practitioners. For novices in the field of data science, they may spend a lot of time practicing a method that they think is right. What if experts could compile their best practices into an easy-to-use package for all developers? This is the idea behind automatic machine learning (automl) and the design concept of autogluon automl library, through which we can train the most advanced machine learning models for image classification, object detection, text classification and tabular data prediction without any experience in machine learning. Learn more: machine learning | Na, you are already a mature machine. Please solve the learning problems yourself!

Building and training ML model is the combination of science and technology. From collecting and preparing data sets to experimenting with different algorithms to find out the best training parameters, ML practitioners need to remove many obstacles to provide high-performance models. However, in the process of training, many people often encounter some thorny problems, resulting in the model can not be correctly extracted or difficult to learn the model in the data set, and ultimately found that most of these problems are caused by improper parameter initialization, bad super parameter combination, design problems in our own code, etc. The recently released Amazon sagemaker debugger can automatically identify complex problems in machine learning (ML) training assignments and help us create models more quickly and effectively. Welcome to read: use the debugger to get your ML model out of Murphy’s law.

Deep learning

In the field of deep learning, reasoning is a process of prediction using training models. For deep learning applications using pytorch and other frameworks, reasoning costs account for 90% of the computing costs. Because the deep learning model needs different amount of GPU, CPU and memory resources, it is difficult to select the appropriate instance for reasoning. Amazon elastic inference solves this problem successfully by attaching appropriate amount of GPU support reasoning to any Amazon sagemaker or EC2 instance or Amazon ECS task, which helps elastic inference users reduce costs and significantly improve the delay of pytorch model on Amazon sagemaker. Further understanding: pytorch model reasoning, speed up and reduce costs, how to achieve this?

artificial intelligence

With the popularity of machine learning technology, people find that using GPU to train ML model is a better way, faster and better performance. However, after all, the original intention of GPU is to process computer images, not for AI. Is there a chip for machine learning that can achieve faster speed and lower cost than GPU training? To solve this problem, AWS has designed and built inferentia chip, which has the characteristics of high performance, low delay, high ease of use, and can provide higher performance machine learning reasoning ability at lower cost. To learn more: AWS has developed a dedicated chip for machine learning and reasoning. Don’t you want to experience it?

July 2020 AWS highlights of AI content