How do you view important issues in the direction of artificial intelligence?

Artificial intelligence algorithm

At present, AI (artificial intelligence) has become a major trend in the technology industry. Every industry is closely related to "AI". The fields related to AI are shown in the figure below, including fields strongly related to AI and fields indirectly empowered by AI. So what exactly is artificial intelligence, the application of artificial intelligence, and the artificial intelligence system will be introduced one by one later.

In the 80 years of artificial intelligence development, it has achieved from the Turing test to the face-changing of the whole people. Machines use artificial intelligence to answer questions, create, or calculate and analyze like humans. In some fields, computers have been able to do as well as humans. For example, the "face-changing of all people" on the Internet in 2019 is the result of the extensive application of technologies such as deep learning and neural networks in artificial intelligence.

At present, there are many applications of "AI" technology in people's life and industrial production to replace human work. For example, the more popular "ELON MUSK'S" can simulate the work of the human brain. However, with the rapid development of artificial intelligence, there have also been some reflections on artificial intelligence and some "fake AI".

Artificial intelligence AI has faced a series of incidents in the development process, among which there is a relatively serious incident of defrauding 200 million financing by counterfeit and shoddy AI. So what exactly is artificial intelligence and what are its main uses are the issues that will be discussed next.

In academia, the definition of artificial intelligence also varies. Artificial intelligence is to accept the input information, through the arrangement and judgment of the information, make a series of rational behaviors and decisions on the input information like human beings. Its main characteristic is "rational action".

In this feedback from "perception" to "decision-making", how to perceive the information of the external world becomes the key to whether artificial intelligence can act. Since it is to simulate the human brain, the process of human perception is actually a process of understanding and learning. That is the problem to be solved by "deep learning" in artificial intelligence.

deep learning

Only by converting external information (video, text, passwords, etc.) into machine language can it be accepted and responded to by artificial intelligence. The thinking on this issue has been considered and researched by scientists as early as the early days of artificial intelligence.

After that, people began to discuss how to complete the input of information through visual perception, and did a lot of research. In 2012, Hinton, the winner of the ImageNet competition at the University of Toronto, Canada, and his student Alex Krizhevsky designed it. Also after that year, more and deeper neural networks were proposed, such as the excellent vgg, GoogLeNet. This is already quite good for traditional machine learning classification algorithms.

The deep learning journey started by AlexNet

In layman's terms, it is to accurately identify the objects needed in our instructions among a large number of objects. The application of this model has made rapid development in the field of image recognition and has been widely used.


The hierarchical learning model of neural networks is the same as our human brain. As we continue to learn, neural networks become more and more complex. Suppose you want to find the labeled information "cat" in millions of image information, and then train the edited visual network model in a very large data set. More complex training is achieved through iteration of the model.

At present, the more commonly used "RestNet model" has a depth of more than 100 layers, and some latest scientific research results have been added. For example, the quick link in the arch bridge part in the bottom picture can effectively and quickly train such a deep network. Ultimately solve the "perception" problem in the visual field.

Alibaba Cloud: Smart Aviation Ramp Management
Use artificial intelligence to identify aircraft types, boarding gates, airport vehicles, combine them with actual maps, and understand the trajectory of aircraft during flight, etc., all of which can be used as input information to be managed by artificial intelligence, making Airport operations are faster and more efficient.

As mentioned above, deep learning is an important form and method of perception. The main components of the deep learning algorithm are:

Data annotation
Algorithmic model development
High Performance Distributed Training
Model tuning
model deployment
After "perception", another thing that artificial intelligence needs to do is "decision-making". Deep learning is a black-box operation. It can learn and perceive external information well, but it cannot give feedback and how to explain the reason for the problem it perceives. That requires "decision making" for analysis and feedback.

The role models of traditional machine learning are decision tree algorithms and logistic regression. For example, the process of a bank granting loans is a decision-making process after weighing various factors. In the form of a decision tree, the judgment of "Yes" or "No" can be used to finally decide whether to issue a loan. Logistic regression refers to the relationship between two types of data, which can be solved accurately through mathematical methods.

In fact, deep learning and machine learning are complementary states. Deep learning solves perception problems very well (computer vision, speech, etc.), and can use the neural network architecture to solve many "perception" problems, but it needs to explain these perception things. Traditional machine learning does not have such a humanized perception function, but its model is relatively small, and we can directly explain it (such as finance, risk control, etc.).

Artificial intelligence has long been applied in the field of advertising. As early as the Song Dynasty, there were advertisements, which were used to help solicit business.

At present, the typical advertising scene is Taobao advertising. The manufacturer first understands the user's preferences through the consumer's personal browsing information, and then uses the intelligent recommendation system to push the relevant products that the consumer searches for. The wide application of such intelligent algorithms makes users' information browsing more efficient and refined.

Both perception and decision-making are related to algorithms.

perception. Related to deep learning algorithms, it involves data labeling, algorithm model development, high-performance distributed training, performance tuning, model deployment, etc.
decision making. Traditional machine learning algorithms and deep learning algorithms are related, involving industry behavior data collection, structured/unstructured data processing, combined modeling of data and algorithms, algorithm development training and tuning, model deployment and real-time training feedback, etc.

artificial intelligence system

Today, with the rapid development of algorithms, the corresponding infrastructure support is also particularly important, which requires the support of artificial intelligence systems. Algorithms and computing power are two indispensable factors for building artificial intelligence or machine learning systems. Behind algorithm innovation is a breakthrough in computing power.

As of 2019, the demand for computing power of artificial intelligence is shown in the figure below. Compared with AlphaGo Zero, AlexNet's demand for computing power has increased by 300,000 times. In this case, solving the problems of algorithm iteration and algorithm implementation puts higher requirements on the system.

The so-called system of AlexNet in 2013 is shown in the figure below. For a simple machine plus GPU, the training cost at that time was about 500 watts per day for seven days, that is, the iteration cycle of the business model was about one week.

Today, with the rapid development of business needs, such as advertising recommendations, a one-week model iteration cycle is far from meeting the needs. Therefore, more and more people are paying attention to how to provide better computing power for artificial intelligence systems through large-scale clusters or chips. MIT made a comparison in 2014. A person can process about 77 pictures in one minute, and a single GPU can process 230 pictures in the same time. Although the processing speed of a single GPU is not much different from that of a human, However, it can achieve larger-scale and faster calculations through GPU clusters. For example, a cluster of 512 GPUs in the figure below can process 60,000 images within one minute.

During the design process of an artificial intelligence system, attention needs to be paid to how to implement high-performance storage, how to realize fast communication between machines, and how to maintain the stability of distributed clusters. Today, Alibaba Cloud has an Eflops platform that can perform 1018 calculations within three minutes and consumes 128 kilowatts per minute. This is a capability that was unimaginable before 2015. The realization of this capability is mainly due to the large-scale cluster and the scalability of the underlying chip of the system.

At present, many domestic companies are committed to the research and development of higher-performance chips, and Ali is no exception. In 2019, Ali released the world's highest-performance AI reasoning chip Hanguang 800, and tested it in the actual test scenarios of urban brain and aviation brain. The peak performance can reach nearly 800,000 pictures per second, which is different from the previous generation's Compared with the chip, it has achieved a performance improvement of about 40 times.

After the system complexity increases, it will bring a series of problems, including software complexity, hardware complexity, resource management complexity, scheduling efficiency complexity, and system-wide optimization complexity. This is a relatively common challenge in the process of system development .

It should be emphasized that AI clusters are not equal to general-purpose clusters. AI requires periodic synchronization of subtasks during training, and high-performance communication between different machines is often based on dedicated GPU or NPU components. Different computing models and different interaction modes present relatively big challenges to AI training.

AI can be used in various business scenarios of Ali, so the platform design can be polished through AI practice, such as the million-category model of Taobao-Pailitao, the voice + NLP of Taobao, and the advertisement recommendation of Ali Mama.

The polished Feitian AI platform is divided into three layers, from the bottom basic hardware, to the middle training and reasoning framework, and then to the development platform. There are three important platforms for AI platforms:

Lightweight AI development platform: helping algorithm and data scientists realize one-click development, debugging and deployment
AI and Big Data Collaborative Development Platform: Helps develop systems for big data-oriented businesses more quickly
AI reasoning service platform: solve the problem of computing resources needed for reasoning, model training, deployment and effect monitoring
The above three platforms support the output of algorithm APIs, vertical domain platforms and brain solutions.

In the field of deep learning, Stanford University launched a test benchmark called DAWNBench. Compared with the previous best results, Alibaba Cloud machine learning has achieved about 10% performance optimization.

Today, AI technology capabilities are of great significance for improving asset utilization and solving the needs of different scenarios. Comprehensive AI technology capabilities mainly involve the following aspects:

Basic hardware: used to provide general-purpose computing power and the computing power required by AI, and provide cloud capabilities through IaaS
AI cloud service: the most basic PaaS layer, providing most users with computing power suitable for AI through an easy-to-launch software and hardware environment
High-performance computing: providing core AI computing engine acceleration
AI system framework: Provides a complete abstraction of AI computing models and cross-architecture modeling iteration and deployment
AI hosting platform: Improve the efficiency of algorithm development, shared deployment and output, and a development platform with user stickiness

Intelligent Computing and Data Computing
AI is intelligent computing, and the field of big data is data computing, and the two are complementary and indispensable.

Data supports AI

The algorithms and computing power just mentioned need the support of a large amount of data, and data is an important part of reflecting the value of algorithms and computing power.

The pictures below show the scenes of the Pope's enthronement in 2005 and 2013, respectively. The current development of the mobile Internet has led to an exponential growth of data, which can also improve the performance of deep learning.

The training data of a small system MNIST in 1998 was only 10MB, ImageNet in 2009 had 200G, WebVision in 2017 had 3TB, and a typical product vision system had 1PB. Massive data helps Ali improve its performance almost linearly.

Give a scene in life to illustrate the effect of data volume on performance improvement. In the field of X-ray medical identification, studies have shown that the effectiveness of doctors in identifying diseases on X-rays is directly proportional to the number of X-rays they have seen. The more you read, the higher your chances of getting it right. In the same way, the current medical engine system can train more data through a large-scale computer system to achieve more accurate medical identification.

AI drives big data towards intelligence

The figure below shows Forum’s summary of trends in the field of big data. At present, the field of big data needs to extract more information, realize real-time calculations, implement AI platforms and online predictions, etc., all of which reflect the trend of big data towards intelligence.

The answer to how different types of data from multiple data sources, such as structured, semi-structured, and unstructured, can play its value after falling into the data warehouse is intelligent computing. Taking the advertising recommendation scenario as an example, the data source is the user’s click, browsing and purchase behavior data on Taobao, which is dropped into the data warehouse through data integration offline or real-time synchronization, offline or real-time ETL, and then through the data warehouse or The data lake solution generates various data models to train the data, and finally outputs the training results through data services. It can be found that the way of understanding and using data in this process begins to become intelligent.

HTAP a few years ago included two parts: OLTP and OLAP. OLAP can be further decomposed into big data analysis, offline and real-time analysis, and different engines can be selected based on the amount of data. At present, data services are becoming more and more important. In some intelligent customer service scenarios, it is necessary to rely on data extraction models for real-time artificial intelligence inference services and applications. Therefore, how to combine analytics and services is also critical. This is also the HSAP that is currently being considered. It uses artificial intelligence to drive offline, real-time data value extraction from data warehouses, and pushes them to users through data services.

Ali has precipitated AI-powered big data methodologies and solutions in its own applications, such as offline computing (batch processing), real-time computing (stream computing), interactive analysis, and graph computing during the Double Eleven promotion. , combined with the Feitian AI platform, provides users with a complete new generation of Feitian big data products powered by AI.

Big data, like AI, is also very performance-oriented. In 2019, Alibaba Cloud's big data platforms MaxCompute and EMR have obvious advantages in computing performance and cost performance on TPC. The specific test results are shown in the figure below.

Ali’s Ali currently provides users with an intelligent voice customer service interaction method, which applies deep learning and intelligent perception AI technology, and needs to be closely connected with the big data business system behind it, such as logistics, user data, etc. In order to achieve the final intelligent effect.



So how should an enterprise embrace AI? To put it simply, artificial intelligence needs to be implemented, and it should start from application requirements and gradually pursue technological innovation, just like Edison invented the electric light. Provide low-cost, high-performance and high-stability infrastructure through the cloud, but the key should be clear what the requirements are.

In the past few years, AI has been doing algorithm innovation and doing demos, but this is far from enough.

AI algorithms are only one part of the system. How to collect data, obtain useful features, how to verify, how to manage processes, resource management, etc., are all issues that companies need to consider when embracing AI.

AI is not omnipotent, but it is absolutely impossible to ignore AI. When enterprises embrace AI, the most important thing is to start from the business. With the increasing amount of data and more and more algorithms, the core is to establish a team of data engineers and algorithm engineers who understand the business, which is the key to the success of current intelligent enterprises. The algorithms, computing power, and data mentioned above can all be realized by using the services and solutions currently provided on the cloud, which can help enterprises realize the implementation of AI more quickly.