Machine learning can be understood from multiple aspects. Tom Mitchell, a global machine learning scientist, provided a widely quoted definition: "A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E." These definitions are simple and abstract. However, as we deepen our understanding of machine learning, we will find that the connotation and extension of machine learning are changing over time. Because a variety of fields and applications are involved and machine learning develops rapidly, it is not easy to define machine learning simply and clearly.

In general knowledge, processing systems and algorithms of machine learning (ML) are an identification mode that performs prediction by finding a hidden mode in data. ML is an important subfield of AI, which also intersects with data mining (DM) and knowledge discovery in database (KDD), for better understanding and distinguishing of artificial intelligence, machine learning, data mining, pattern recognition, statistics, and neural computing.

<!-- Slide number: 10 -->

Relationship of AI, Machine Learning, and Deep Learning


Notes:

<!-- Slide number: 11 -->

Relationship of AI, Machine Learning and Deep Learning

AI: A new technical science that focuses on the research and development of theories, methods, techniques, and application systems for simulating and extending human intelligence.

Machine learning: A core research field of AI. It focuses on the study of how computers can obtain new knowledge or skills by simulating or performing learning behavior of human beings, and reorganize existing knowledge architecture to improve its performance. It is one of the core research fields of AI.

Deep learning: A new field of machine learning. The concept of deep learning originates from the research on artificial neural networks. The multi-layer perceptron (MLP) is a type a deep learning architecture. Deep learning aims to simulate the human brain to interpret data such as images, sounds, and texts.

Notes:

<!-- Slide number: 12 -->

Three Major Schools of Thought: Symbolism

Basic thoughts

The cognitive process of human beings is the process of inference and operation of various symbols.

A human being is a physical symbol system, and so is a computer. Computers, therefore, can be used to simulate intelligent behavior of human beings.

The core of AI lies in knowledge representation, knowledge inference, and knowledge application. Knowledge and concepts can be represented with symbols. Cognition is the process of symbol processing while inference refers to the process of solving problems by using heuristic knowledge and search.

Representative of symbolism: inference, including symbolic inference and machine inference

Notes:

Symbolism (logicism, psychology, computer science): Symbols represent knowledge, and inference produces cognitive results.

In symbolism, the concept of apple can be represented by the apple node or a group of nodes that represent its attributes. The focus of symbolism is theoretical logic inference. Connectionism, being weak in logic, is result-oriented.

<!-- Slide number: 13 -->

Three Major Schools of Thought: Connectionism

Basic thoughts

The basis of thinking is neurons rather than the process of symbol processing.

Human brains vary from computers. A computer working mode based on connectionism is proposed to replace the computer working mode based on symbolic operation.


Representative of connectionism: neural networks and deep learning

Notes:

Connectionism is derived from bionics, especially studies of the human brain model. In connectionism, a concept is represented by a set of numbers, vectors, matrices, or tensors in a specific activation mode of the entire network. Each node, without specific meaning, plays its role in the representation of the concept. For example, in symbolism, the concept of a cat may be represented by a "cat node" or a set of nodes representing the cat's attributes, such as "two eyes", "four legs", and "fluffy". However, in connectionism, each node does not represent a particular concept, and it is impossible to find a cat node or an eye neuron.

<!-- Slide number: 14 -->

Three Major Schools of Thought: Behaviorism

Basic thoughts:

Intelligence depends on perception and action. The perception-action mode of intelligent behavior is proposed.

Intelligence requires no knowledge, representation, or inference. AI can evolve like human intelligence. Intelligent behavior can only be demonstrated in the real world through the constant interaction with the surrounding environment.

Representative of behaviorism: behavior control, adaptation, and evolutionary computing

Notes:

Behaviorism concerns more about application practices and how to learn from the environment continuously to make corrections.

<!-- Slide number: 15 -->

Brief Development History of AI

Notes:

In August 1956, a group of scientists gathered at Dartmouth College in the US to discuss an otherworldly topic: how to make computers simulate learning behavior and other intelligent actions of human beings. Many well-known scientists attended the Dartmouth Conference, including John McCarthy (founder of LISP), Marvin Minsky (expert in AI and cognition), and Claude Shannon (founder of the information theory), Allen Newell (computer scientist), and Herbert Simon (Nobel laureate in economics).

The meeting lasted for two months. Although no consensus was reached, a name was given to the discussed object: artificial intelligence. The year of 1956, therefore, is the birth year of AI.

<!-- Slide number: 16 -->

Overview of AI Technologies

AI technologies are multi-layered, covering the application, algorithm mechanism, toolchain, device, chip, process, and material layers.

Application


Algorithm

Device

Chip

Process

Notes:

On one hand, the rapid development of applications and algorithms, especially deep learning and convolutional neural networks, raises performance optimization requirements for AI chips by two to three orders of magnitude, which has triggered the upsurge of AI chip R&D in recent years. On the other hand, the rapid development of new materials, processes, and components, such as 3D stacked memory and process evolution, also makes significant improvements in performance and power consumption of AI chips possible. This driving force comes from breakthroughs in basic research. In general, the above driving forces empower rapid advancement of AI chip technologies in recent years.

Application layer:

Voice: speech recognition, speech synthesis, voice wakeup, voiceprint recognition, music generation, smart speaker, smart navigation...

Text: text analysis, language translation, man-machine dialog, reading comprehension, recommendation system...

Control: autonomous driving, drone, robot, industrial automation...

<!-- Slide number: 17 -->

Types of AI

Strong AI

The strong AI view holds that it is possible to create intelligent machines that can really reason and solve problems. Such machines are considered to be conscious and self-aware, can independently think about problems and work out optimal solutions to problems, have their own system of values and world views, and have all the same instincts as living things, such as survival and security needs. It can be regarded as a new civilization in a certain sense.

Weak AI

The weak AI view holds that intelligent machines cannot really reason and solve problems. These machines only look intelligent, but do not have real intelligence or self-awareness.

Notes:

A key counter of AI is to achieve a superhuman level in challenging fields through self-learning without any prior knowledge.

Strong AI can compete with humans in all aspects. Therefore, it aims to enable robots to implement human-like capabilities in all aspects rather than a specific field. Strong AI can think, make plans, solve problems, perform abstract thinking, understand complex concepts, quickly learn, and learn from experience. Currently, it is believed that if we can simulate the human brain and copy all its neurons and synapses on the same scale, strong AI will naturally occur.

Now we are in the weak AI phase. The emergence of weak AI alleviates human intellectual labor, similar to advanced bionics. Both AlphaGo and robots that can write press releases and novels fall in the weak AI phase because they are better than humans only in some ways. The roles of data and computing power are self-evident in the era of weak AI, and promote the commercialization of AI. In the era of strong AI, these two factors are still critical. At the same time, the research on quantum computing by technology giants like Google and IBM also provides powerful support for humans to enter the era of strong AI.

<!-- Slide number: 18 -->

Classification of Intelligent Robots

Currently, there is no unified definition of AI research. Intelligent robots are generally classified into the following four types:

"Thinking like human beings": weak AI, such as Watson and AlphaGo

"Acting like human beings": weak AI, such as humanoid robot, iRobot, and Atlas of Boston Dynamics

"Thinking rationally": strong AI (Currently, no intelligent robots of this type have been created due to the bottleneck in brain science.)

"Acting rationally": strong AI

Notes:

Generally, AI can be divided into four categories: "thinking like a human", "acting like a human", "thinking rationally", and "acting rationally". The "acting" here should be understood in a broad sense as taking actions or making decisions, rather than physical movements. Mainstream research focuses on weak AI, and it is generally believed that this research field has achieved considerable achievements.

Boston Dynamics 10 Years challenge on Robots：https://www.youtube.com/watch?v=WjE1_XEUoGE

<!-- Slide number: 19 -->

AI Industry Ecosystem

The four elements of AI are data, algorithm, computing power, and scenario. To meet requirements of these four elements, we need to combine AI with cloud computing, big data, and IoT to build an intelligent society.


Notes:

<!-- Slide number: 20 -->

Sub-fields of AI


AI Development Report 2020

Notes:

<!-- Slide number: 21 -->

AI Overview

Technical Fields and Application Fields of AI

Huawei's AI Development Strategy

AI Disputes

Future Prospects of AI

Notes:

<!-- Slide number: 22 -->

Technical Fields and Application Fields of AI


Global AI Development White Paper 2020

Notes:

<!-- Slide number: 23 -->

Distribution of AI Application Technologies in Enterprises Inside and Outside China


At present, application directions of AI technologies mainly include:

Computer vision: a science of how to make computers "see"

Speech processing: a general term for various processing technologies used to research the voicing process, statistical features of speech signals, speech recognition, machine-based speech synthesis, and speech perception

Natural language processing (NLP): a subject that use computer technologies to understand and use natural language

Distribution of AI application technologies in enterprises inside and outside China

China AI Development Report 2018

Notes:

<!-- Slide number: 24 -->

Voice Processing Application Scenario (1)

The main topics of voice processing research include voice recognition, voice synthesis, voice wakeup, voiceprint recognition, and audio-based incident detection. Among them, the most mature technology is voice recognition. As for near field recognition in a quite indoor environment, the recognition accuracy can reach 96%.

Application scenarios:

Voice navigation

Question Answering Bot (QABot)


Notes:

<!-- Slide number: 25 -->

Voice Processing Application Scenario (2)



Return Visit

Real-time conference records

Other applications:

Spoken language evaluation

Diagnostic robot

Voiceprint recognition

Smart sound box

...

Notes:

Telephone follow-up: converts telephone follow-up content into human voices to communicate with customers, improving the overall convenience of calling customers.

Intelligent education: converts textbooks into life-like voices to simulate classroom teaching, helping teachers find new, innovative ways to better the education of their students.

Conference records: quickly identifies audio files of conference records and converts them into text for easy recording.

Real-time conference records: converts the audio files of a video conference or a conference call into text in real time and allows users to check, modify, and retrieve the conference content, improving the conferencing efficiency.

<!-- Slide number: 26 -->

NLP Application Scenario (1)

The main topics of NLP research include machine translation, text mining, and sentiment analysis. NLP imposes high requirements on technologies but confronts low technology maturity. Due to high complexity of semantics, it is hard to reach the human understanding level using parallel computing based on big data and parallel computing only.

In future, NLP will achieve more growth: understanding of shallow semantics → automatic extraction of features and understanding of deep semantics; single-purpose intelligence (ML) → hybrid intelligence (ML, DL, and RL)

Application scenarios:


Trend analysis

Theme mining

Emotional analysis

Hotspot event

Information distribution

Public opinion analysis


Public opinion analysis

Evaluation analysis

Notes:

Public opinion analysis: uses algorithm models such as keyword extraction, text clustering, and theme mining to mine emergencies and public opinion orientation, discover topics and trends, and analyze public opinions. It analyzes emotions, hot topics, trends, and prorogation channels from multiple dimensions to master public opinion trends in a timely manner.

Evaluation analysis: uses approaches such as emotion analysis and opinion extraction to extract emotional tendency and key opinion information from a large number of user comments.

<!-- Slide number: 27 -->

NLP Application Scenario (2)



Text classification

Machine translation

Other applications:

Knowledge graph

Intelligent copywriting

Video subtitle

...

Notes:

Translation center: builds a machine translation system that meets specific requirements and efficiently and accurately translates emails, papers, and news.

Text classification: classifies a large number of files based on the preset categories. Typical application scenarios include category labeling, content retrieval, and personalized content recommendation.

Intelligent copywriting: automatically generates new copywriting by learning the writing mode and structure of existing documents and using machine learning algorithms.

Video subtitles: uses machine translation and speech recognition to implement real-time translation and provide bilingual subtitles, helping people quickly understand each other.

Automobile knowledge graph drives technological transformation: The knowledge graph can be well applied in the automobile field with various attributes. It is able to provide consumers with a comprehensive shopping guide map by integrating the automotive information of different models and brands. In the pre-sales scenario, knowledge graph helps answer consumer questions about the product price, maintenance, configurations, and comparisons of price and performance. In addition, it also recommends customers specified models with outstanding features, such as technology and security. By building a knowledge graph for the automotive field, we establish a multi-round dialog system. This system analyzes the entities and relations in the user discourse, query in the knowledge graph, and select the optimal dialog strategy.

<!-- Slide number: 28 -->

AI Application Field - Intelligent Healthcare

Notes:

With AI technologies, we can enable AI to "learn" professional medical knowledge, "remember" numerous historical medical cases, and identify medical images with computer vision technologies to provide reliable and efficient assistance for doctors.

<!-- Slide number: 29 -->

AI Application Field - Smart Home

Based on IoT technologies, a smart home ecosystem is formed with hardware, software, and cloud platforms, providing users personalized life services and making home life more convenient, comfortable, and safe.


Okay, the temperature's set.

Set the temperature to 26 degrees.

Notes:

<!-- Slide number: 30 -->

AI Application Field - Retail

AI will bring revolutionary changes to the retail industry. A typical symptom is unmanned supermarkets. For example, Amazon Go, unmanned supermarket of Amazon, uses sensors, cameras, computer vision, and deep learning algorithms to completely cancel the checkout process, allowing customers to pick up goods and "just walk out".

One of the biggest challenges for unmanned supermarket is how to charge the right fees to the right customers. So far, Amazon Go is the only successful business case and even this case involves many controlled factors. For example, only Prime members can enter Amazon Go. Other enterprises, to follow the example of Amazon, have to build their membership system first.


Notes:

<!-- Slide number: 31 -->

AI Application Field - Autonomous Driving

The Society of Automotive Engineers (SAE) in the U.S. defines 6 levels of driving automation ranging from 0 (fully manual) to 5 (fully autonomous). L0 indicates that the driving of a vehicle completely depends on the driver's operation. The system above L3 can implement the driver's hand-off operation in specific cases, L5 depends on the system when vehicles are driving in all scenarios.

Currently, only some commercial passenger vehicle models, such as Audi A8, Tesla, and Cadillac, support L2 and L3 Advanced driver-assistance systems (ADAS). It is estimated that by 2020, more L3 vehicle models will emerge with the further improvement of sensors and vehicle-mounted processors. L4 and L5 autonomous driving is expected to be first implemented on commercial vehicles in closed campuses. A wider range of passenger vehicles require advanced autonomous driving, which requires further improvement of technologies, policies, and infrastructure. It is estimated that L4 and L5 autonomous driving will be supported by common roads in 2025–2030.

Notes:

Video material: Why is it so difficult to implement autonomous driving? PaperClip X NIO https://www.youtube.com/watch?v=MYPZbNEI77o

Understand the autonomous driving in 5 minutes! |Science Big Bang 2-EP.25: https://www.youtube.com/watch?v=09tRb1G5afU

Can not tell the difference between autonomous driving and unmanned driving? After watching this audio, you will be able to share the up-to-date car technologies with friends: https://www.youtube.com/watch?v=EaeV526oqWE

Currently, the electronic control system uses the distributed ECU architecture. The information, entertainment, vehicle body, vehicle movement, and power assembly systems and their subdivision functions are separately controlled by different ECUs. Some high-end models have more than 100 ECUs. In the future, as vehicles enter the autonomous driving era of L3 or higher, more vehicle-mounted sensors will be used, and the amount of data generated by the sensors will increase sharply. The distributed electronic system cannot meet requirements for efficient convergence processing on a large amount of diversified sensor data and vehicle control decision making based on all sensor data. To meet the preceding requirements, the electronic control system needs to be centralized towards the domain control unit (DCU) and multi-domain control unit (MDC). In the future, the electronic control system will further be centralized and platform-based and support software and hardware decoupling. A unified supercomputing platform will be used to process, converge, and make decisions on sensor data to achieve high-level autonomous driving.

<!-- Slide number: 32 -->

AI Overview

Technical Fields and Application Fields of AI

Huawei's AI Development Strategy

AI Disputes

Future Prospects of AI

Notes:

<!-- Slide number: 33 -->

Huawei's Full-Stack, All-Scenario AI Portfolio

AI Applications

Application Enablement

HiAI Engine

ModelArts

Framework

TensorFlow

PyTorch

PaddlePaddle

MindSpore

Chip Enablement

CANN

Full Stack

IP & Chip

IP and Chip

Ascend

Ascend-Nano

Ascend-Tiny

Ascend-Lite

Ascend-Max

Ascend-Mini






Atlas











All Scenarios

Edge Computing

Consumer Device

Public Cloud

Private Cloud

Industrial IoT Device

Application enablement: provides end-to-end services (ModelArts), layered APIs, and pre-integrated solutions.

MindSpore: supports the unified training and inference framework that is independent of the device, edge, and cloud.

CANN: a chip operator library and highly automated operator development tool.

Ascend: provides a series of NPU IPs and chips based on a unified, scalable architecture.

Atlas: enables an all-scenario AI infrastructure solution that is oriented to the device, edge, and cloud based on the Ascend series AI processors and various product forms.

Huawei's "all AI scenarios" indicate different deployment scenarios for AI, including public clouds, private clouds, edge computing in all forms, industrial IoT devices, and consumer devices.

Notes:

Huawei announced that it will open source the server OS on December 31, 2020, the standalone GaussDB OLTP database in June 2020, and the MindSpore all-scenario AI computing framework in the first quarter of 2020.

"Full-stack" refers to its technical function. Huawei's full-stack portfolio includes chips, chip enablement, a training and inference framework, and application enablement.

By "all scenarios", Huawei means different deployment scenarios for AI, including public clouds, private clouds, edge computing in all forms, industrial IoT devices, and consumer devices.

As the cornerstone of Huawei full-stack AI solution, Atlas provides modules, cards, and servers based on the Ascend AI processor to meet customers' computing requirements in all scenarios.

<!-- Slide number: 34 -->

Full Stack - ModelArts Full-Cycle AI Workflow

EI Intelligent Twins

EI Cognition Service AI Service

AI data framework

Algorithm development

Training

Deployment

Market

Efficient filtering and semi-automated labeling, data preprocessing

Efficiency improved by 100 times

Out-of-the-box development environment compatible with mainstream frameworks

MoXing library, simplifying model development

Built-in model algorithms, improving development efficiency

Distributed training, shortening training period from weeks to minutes

Wizard-based AutoLearning, code-free development, enabling model training from scratch

One-click deployment on device, edge, and cloud

All-scenario deployment

Inference on the Ascend AI processor

AI sharing platform helps enterprises build internal and external AI ecosystems


Data

Visualized Workflow Management


Version management, traceable and worry-free development

ModelArts





AI data framework

accelerates data processing by 100 folds.

Visualized workflow management

makes development worry-free.

Distributed training

shortens training from weeks to minutes.

One-click deployment on device, edge, and cloud

supports various deployment scenarios.

Automatic learning

enables you to start from scratch.

AI sharing platform

builds internal and external AI ecosystems for enterprises.

AI applications

Notes:

ModelArts is a one-stop development platform for AI developers. With data preprocessing, semi-automated data labeling, distributed training, automated model building, and model deployment on the device, edge, and cloud, ModelArts helps AI developers build models quickly and manage the lifecycle of AI development.

Automatic learning: can automate model design, parameter adjustment, and model training, compression, and deployment with the labeled data. The process is code-free and requires no model development experience.

Device-edge-cloud: indicates devices, Huawei intelligent edge devices, and HUAWEI CLOUD, respectively.

Online inference: a web service that synchronously provides the inference result for each inference request.

Batch inference: a job that processes batch data for inference.

<!-- Slide number: 35 -->

Full Stack — MindSpore (Huawei AI Computing Framework)

MindSpore provides automatic parallel capabilities. With MindSpore, senior algorithm engineers and data scientists who focus on data modeling and problem solving can run algorithms on dozens or even thousands of AI computing nodes with only a few lines of description.

The MindSpore framework supports both large-scale and small-scale deployment, adapting to independent deployment in all scenarios. In addition to the Ascend AI processors, MindSpore also supports other processors such as GPUs and CPUs.

AI application ecosystem for all scenarios

MindSpore

Unified APIs for all scenarios

MindSpore intermediate representation (IR) for computational graph

On-demand collaborative distributed architecture across device-edge-cloud (deployment, scheduling, and communications)

Processors: Ascend, GPU, and CPU

Notes:

In the intelligent era, AI applications in device-edge-cloud scenarios are booming. However, AI still faces huge challenges. Technical barriers, high development cost, and long deployment period hinder the development of the AI developer ecosystem in the entire industry. The all-scenario AI computing framework MindSpore is developed based on the principles of friendly development, efficient operation, and flexible deployment.

In terms of deep learning framework, Huawei MindSpore is the strongest challenger to TensorFlow (Google), MXnet (Amazon), PyTorch (Facebook), and CNTK (Microsoft), which are listed as the four major players.

MindSpore will be open-source on March 30, 2020. It competes with frameworks such as TensorFlow (Google), PyTorch (Facebook), PaddlePaddle (Baidu), and Caffe.

<!-- Slide number: 36 -->

Full Stack — CANN

CANN:

A chip operators library and highly automated operator development toolkit

Optimal development efficiency, in-depth optimization of the common operator library, and abundant APIs

Operator convergence, best matching the performance of the Ascend chip

AI applications

General APIs

Advanced APIs

Pre-integrated Solutions

ModelArts

HiAI Service

Application enablement

CANN

Compute Architecture for Neural Networks

FusionEngine

CCE Operator Library

TBE operator development tool

CCE Compiler

HiAI Engine

Full stack

MindSpore

TensorFlow

PyTorch

PaddlePaddle

…

Framework

CANN

Processor enablement

Ascend

Ascend-Nano

Ascend-Tiny

Ascend-Lite

Ascend-Mini

Ascend-Max

IP and Chip

Consumer device

Public cloud

Private cloud

Edge computing

Industrial devices

All scenarios

Notes:

CANN is a chip enabling layer developed by Huawei for deep neural networks and Ascend chips. It consists of four functional modules:

(1) FusionEngine: FusionEngine is an operator-level fusion engine. It fuses operators, reduces the memory transfer between operators, and improves the performance by 50%.

(2) CCE operator library: The optimized general operator library provided by Huawei can meet the requirements of most mainstream vision and NLP neural networks. (It is estimated that the CCE operator library will open APIs in 2020 Q1.)

Inevitably, customers and partners will have custom operator requirements for timeliness, privacy, and research. In this case, the third functional module is used.

(3) Tensor Boost Engine (TBE): TBE is an efficient and high-performance custom operator development tool. It abstracts hardware resources as APIs, enabling customers to quickly construct required operators. (This functional module is expected to be available in 2020 Q4.)

(4) The last module is the bottom-layer compiler that optimizes performance and supports Ascend processors in all scenarios.

Note: TE-DSL is short for Tensor Engine-Description language.

<!-- Slide number: 37 -->

Full Stack — Ascend 310 AI Processor and Da Vinci Core


Notes:

<!-- Slide number: 38 -->

Ascend AI Processors: Infusing Superior Intelligence for Computing



FLOPS

256T

4

3

125T

2

90T

45T

1

Ascend 910

Ascend 310

Ascend 910

Most powerful AI processor

AI SoC with ultimate energy efficiency

Ascend-Mini

Architecture: Da Vinci

Ascend-Max

Architecture: Da Vinci

Half-precision (FP16): 8 TFLOPS

Integer precision (INT8): 16 TOPS

16-channel full-HD video decoder: H.264/265

Half-precision (FP16): 256 TFLOPS

Integer precision (INT8): 512 TOPS

128-channel full HD video decoder: H.264/265

1-channel full-HD video encoder: H.264/265

Max. power: 8 W

Max. power: 310 W

Notes:

Demands for AI are soaring worldwide. However, with the market being dominated by only a few vendors, AI processors are sold at a very high price. The delivery cycle is long and the local service support is weak. Therefore, the AI requirements of many industries cannot be effectively met.

At HUAWEI CONNECT held in October 2018, Huawei unveiled its Ascend 310 processor for AI inference and Ascend 910 processor for AI training. Built upon the unique Da Vinci 3D Cube architecture, Huawei’s Ascend AI processors boast high computing power, energy efficiency, and scalability.

Ascend 310, an AI SoC with ultimate performance per watt, is designed for edge inference. It provides up to 16 TOPS of computing power, with a power consumption of only 8 watts. This makes it a perfect choice for edge computing.

The Ascend 910 AI processor delivers the industry's highest computing density on a single AI chip. It applies to AI training and delivers 512 TOPS of computing power, with a maximum power consumption of 310 watts.

(In the chart, Google TPU v2, Google TPU v3, NVIDIA Tesla V100, and Huawei Ascend 910 are shown from left to right.)

<!-- Slide number: 39 -->

Atlas AI Computing Platform Portfolio

Internet, security, finance, trans...