The Heart of Machine Report
Editor: Zenan
Last week, Google Developer Conference reunited offline.
Since Ng , Quoc Le and others published the "Recognition Cat" paper ten years ago, triggering the technological revolution , Google has been leading the trend of machine learning, and the developer conference held by this company will always receive additional attention.
9-14 to 15, the Google Developer Conference was held in Shanghai. In this event, Google introduced cross-platform technology, privacy protection, machine learning, XR and other technologies, as well as the latest implementation results with many partners based on its latest technologies in the field of technology.
In the field of machine learning , TensorFlow has always been well known as the most popular framework. Recently, it has just evolved to version 2.10. At the developer conference, Google introduced the concept of Tensor Projects, an open source machine learning ecosystem.
Machine learning is a process of using known data to train inference models. The trained models can make effective predictions on unprecedented data and complete various tasks ranging from image recognition, speech to natural language processing. The process of building a leading machine learning framework is to do just that. Google hopes to bring more inspiration and inspiration to the world through the capabilities provided by the TensorFlow framework.
In 2016, Google opened the source of TensorFlow, and the industry situation at that time was very different from the current ones - at that time, there were only a few thousand people who could truly master machine learning technology, and today the number has grown to millions, and there will be more in the future.
Among Tensor Projects users, there are those who focus on the field of technology and those who use machine learning as an auxiliary tool to overcome other fields of research. However, in many cases, these audiences have similar needs. As the number of machine learning talents increases, development tools also need to move forward.
"There are 24.3 million software developers worldwide in 2021, and this number is expected to reach 45 million in 2030," said Wei Wei, a technology promotion engineer at Google TensorFlow. "Most of them do not work full-time on machine learning. About 1/4 will use machine learning during development. Some of the core tasks include training, deploying and managing machine learning models. They will create many practical features that are unimaginable now, such as the ability to intelligently erase unwanted content on photos on mobile phones."
Machine learning is also constantly evolving, and cutting-edge research has recently brought us major breakthroughs such as AlphaFold. To promote these goals, Google has been working on building machine learning frameworks, libraries, infrastructure, counterexamples and tutorials over the years, and has also mastered a wealth of experience in building machine learning applications from scratch.
Nowadays, TensorFlow technology itself has included the entire process of data, modeling, deployment to maintenance. Under this, there is a dedicated acceleration infrastructure that can optimize the entire life cycle , realizing a free open source product ecosystem that is suitable for both hobbyists and researchers.
Tensor Projects As the ecosystem of Google machine learning, it contains different technical products. For example, when building a model, Kerash can help you build a model in a concise way, and it is considered a development tool more suitable for beginners. Google says that people now use TensorFlow Light Model Maker to solve many of the complex tasks faced when creating mobile models.
"When you make an app or website, you need to integrate models. We have created a task library to help people use models. Model Maker and Task libraries currently support large-scale near-neighbor searches on the end side. You can find similar pictures, text or audio in millions of data in milliseconds. Everything can happen on your phone." Wei Wei said.
JAX is a high-performance machine learning library developed by the Google Research team. It has an API interface similar to NumPy and uses an XLA compiler to accelerate the model.Many researchers have high hopes for it, hoping that it can replace TensorFlow as a new generation of popular deep learning frameworks.
In Google, JAX is defined as a machine learning framework used in cutting-edge research, and TensorFlow is the framework used in application. According to reports, DeepMind used JAX to develop AlphaFold to solve the protein folding problem and accurately predict the protein structure.
JAX has specially optimized mathematical calculations for Google's infrastructure. At present, this infrastructure has been opened to all developers for use. Developers can simply modify a few lines of code on TensorFlow to connect to the computing power of Google TPU.
Google has also developed tools for algorithm deployment. Now when deploying models, you can use TensorFlow Extended (TFX) to deploy models to all locations: from the cloud to web servers, browsers, embedded systems, and more. But at the same time, we need to fix bugs, process new data, and ensure that the model outputs responsible results. TensorFlow Extended (TFX) makes continuous training of models possible: it can help you understand model performance more deeply. You can use TFX to train multi-end models and access Colab at any time.
Recently, people have paid more and more attention to trustworthy machine learning. When you create responsible machine learning models, model cards can provide transparency, and now TFX can also automatically generate model cards.
TFX A popular component on the previous one is TensorFlow Serving, which can help deploy the model to the server and then call it remotely. Google has recently released four new Learning Pathways to help everyone learn how to call TensorFlow's interface from Android, Flutter, and Web to complete model inference.
Google has also released two new Colabs to help developers learn TensorFlow Recommenders and TensorFlow Agents to create a full-stack cross-platform app based on machine learning.
At present, the TensorFlow Lite runtime library has been integrated into Google Play Service, which means that users can always use the latest version of TensorFlow Lite. Many applications are currently using TensorFlow Lite in Google Service, with more than 400 million users per month and 20 billion inferences are completed.
Google is also helping more developers and lower the application threshold for machine learning.
Google proposed MediaPipe, hoping to provide developers with highly customized device-side machine learning solutions. Its technology has been applied in the background of Google Meet online meetings, Nest's personnel movement detection, package delivery notifications, packing recognition, and YouTube's virtual makeup trials. Supports multi-model, multi-hardware acceleration, cross-platform machine learning tasks.
At the conference, Google showed us MediaPipe's capabilities in computer vision tasks such as skeletal binding, movement and gesture recognition. The algorithm can detect various actions of people in front of you in real time with just a local GPU and camera.
MediaPipe encapsulates complex machine learning pipelines into Tasks, making it easier for developers to customize models in the easiest way. In the future, calls to machine learning technology on device side will be simplified to require only a few lines of code, or even no code. Google's next plan is to expand MediaPipe from vision to speech recognition and natural language processing.
In the machine learning community, Google's official machine learning tutorials have always been popular. After launching the "TensorFlow Introductory Practical Course", the TensorFlow team recently joined hands with NetEase Youdao to launch a special course on the Mooc platform for developers interested in the deployment field: "TensorFlow Introductory Course - Deployment Chapter", and also launched the Chinese version of "Developer Online Course".
Google provides a wealth of choices for people. Whether it is scientific research or production, when doing the next machine learning project, you will definitely try how to do it with TensorFlow first.
