Cutting-edge AI applications
0 Huawei Confidential
Objectives
Upon completion of this course, you will be able to:
Understand the concepts and applications of reinforcement learning;
Understand the concepts and applications related to the GAN;
Understand the concepts and applications related to knowledge graph;
Understand the concepts and applications of intelligent driving.
1 Huawei Confidential
Contents
1. Reinforcement Learning
2. GAN
3. Knowledge Graph
4. Intelligent Driving
2 Huawei Confidential
Reinforcement Learning
Reinforcement learning (RL) is a branch of
machine learning that emphasizes how to act
AI
based on the environment in order to maximize
the expected benefits.
Machine Learning
Reinforcement
Learning
3 Huawei Confidential
Environment Sensing
State
–6 points
Reward
Action Agent
Environment
4 Huawei Confidential
Basic Concepts in Reinforcement Learning
Two objects:
Agent
Environment
Main elements:
Action: performed at the current moment based on the status.
Policy: makes decisions based on the status and controls the agent to perform actions.
Reward: provided by the environment based on the current action.
Return: sum of rewards at all moments.
5 Huawei Confidential
Classification of Reinforcement Learning Algorithms
The reinforcement learning algorithms can be divided into two directions based on the process of finding
the optimal policy:
Direct solution: The optimal policy function is optimized during the interaction with the environment.
Indirect solution: This type of algorithms are the most common algorithms. They indirectly calculate other
indicators and deduce the optimal policy based on the results of these indicators.
6 Huawei Confidential
Reinforcement Learning Model - Go
AlphaGO
Status
Agent
Reward
What is the situation? Action
Who is winning? Where to place the go?
Reward
Environment
Status
7 Huawei Confidential
Contents
1. Reinforcement Learning
2. GAN
3. Knowledge Graph
4. Intelligent Driving
8 Huawei Confidential
Generative Adversarial Network (GAN)
A GAN is a class of frameworks that trains generator G and discriminator D to compete in a
game. The game between the two makes the discriminator D unable to distinguish whether the
sample is a fake sample or a real sample.
The generator G generates "fake" images that look like the images for training.
The discriminator D determines whether the images output by the generator are real images or fake
images.
GANs are used in scenarios such as image generation, text generation, speech enhancement,
and image super-resolution.
9 Huawei Confidential
GAN Training Process
10 Huawei Confidential
GAN Application
Image dataset generation
Image-to-image conversion
Resolution enhancement, making photos clearer
Text-to-image conversion
11 Huawei Confidential
GAN - Photo Repair
If an area in a photo is faulty (for example, colored or erased), GAN can repair the area and
restore it to its original state.
12 Huawei Confidential
Contents
1. Reinforcement Learning
2. GAN
3. Knowledge Graph
4. Intelligent Driving
13 Huawei Confidential
Basic Concepts
Knowledge graph is essentially a knowledge base of the semantic network. It describes various concepts,
entities, and relationships in the real world in a structured manner to form a huge semantic network.
Nodes in the network are entities, and each edge represents a property of entities or the relationship
between entities.
• Entity: an object that exists in the real world and
can be distinguished from other objects.
• Property: The nature and relationship of a specific
object are called the properties of the object.
• Concept: a set of entities with the same features.
• Ontology: a set of abstract concepts used to describe
the common features of all the things in a domain
and the relationships between them.
14 Huawei Confidential
History of Knowledge Graph
Knowledge graph is a relatively mature application "fruit" of knowledge engineering in the current stage of AI development.
Five phases of knowledge engineering:
Pre-knowledge engineering period (1950s–1970s)
Expert system period (1970s–1990s)
World Wide Web 1.0 period (1990s–early 21st century)
Swarm intelligence period (2000–2006)
Knowledge graph period (2006–present)
In the knowledge graph period, the vigorous development of large-scale structured encyclopedia websites and the continuous
progress of text information technology provide conditions for obtaining large-scale knowledge.
Google took the lead in applying knowledge graphs to search engines in 2012, successfully improving users' search quality and
experience.
15 Huawei Confidential
Process for Constructing a Knowledge Graph
The general knowledge graph construction process is as follows:
Determining Knowledge
Data acquisition mapping and
the domain
• Structured data fusion
• Determine the
target domain • Semi-structured • Knowledge
of the data mapping Knowledge
knowledge • Unstructured data • Knowledge graph
graph. fusion application
Understanding the Information
Knowledge
processing
domain extraction and reasoning
• Fully understand the
target domain of the • Property Extraction
knowledge graph. • Relationship
• Knowledge system extraction
construction • Entity extraction
16 Huawei Confidential
Knowledge Graph Application Scenario (1)
The common application scenarios are as follows:
recommendation
Precise
Music apps accurately recommend songs that you like to listen to.
Intelligent Q&A
Retrieval
Semantic search
Huawei Xiaoyi technology
Apple Siri Knowledge
Microsoft Cortana graph
Baidu Xiaodu Deep learning
The graph technology applies to the intelligent Q&A robots
Exact keyword matching by the Petal search engine
of major vendors.
17 Huawei Confidential
Knowledge Graph Application Scenario (2)
Oil and Gas Knowledge Computing
Straight
well Pattern layer
Block Well (Ontology)
Oil well Item Value
Concept
Measured
horizontal 3.81980698645996E7
coordinate
Name Pool 155
Location - Located in
Entity Long 6 Borehole
2710.0
block depth
Measured
Pool 21 vertical 5883120.49477415
coordinate
Management - Belong to
Relationship Kelly bushing
1571.779
Long elevation
layer 9 Measured
Layer - Drilling catching Pool 11 borehole 2710.0
Data layer depth
Layer - Drilling catching
18 Huawei Confidential
Contents
1. Reinforcement Learning
2. GAN
3. Knowledge Graph
4. Intelligent Driving
19 Huawei Confidential
History of Intelligent Driving
1995 – Navlab, Carnegie Mellon University 2004 - The first DARPA Grand Challenge
The Rapidly Adapting Lateral Position Handler (RALPH) vision system Autonomous driving of 240 km route across the Mojave Desert, USA.
Autonomous driving of 4,496 km (98.1% of the whole trip) during the long- Carnegie Mellon University's Sandstorm traveled the farthest distance,
distance driving experiment completing 11.78 km of the course.
Test environments included morning, night, and extreme weather conditions Some vehicles were able to avoid obstacles but required large, expensive
such as rainstorm. sensing systems.
1995 - Mercedes-Benz VITA developed by Ernst Dickmanns 2005 - The second DARPA Grand Challenge
Pioneered the 4D spatiotemporal dynamic vision model The Mojave Desert 212 km off-road route was used again, but had rougher
Could drive at 130 km/h on highways conditions.
Competitors were much more successful than 2004, with Stanford University's
1996- ARGO autonomous, University of Parma Stanley traveling the farthest in the shortest time. This challenge observed the
Built on a general-purpose chip and low-cost cameras first prototypes of intelligent driving vehicles.
Autonomous driving for 94% of the 2,000 km trip, at speeds up to 112 km/h
2007 - The third DARPA Grand Challenge
1996-2000 - ATB-2 co-developed by NJUST, BIT, THU, ZJU, and NUDT The urban challenge of 96 km, with vehicles ranked on obeying all traffic
Developed on Mercedes-Benz Sprinter 414 regulations and avoiding collisions.
Oriented to structured roads and off-road environments Carnegie Mellon University's car Ross ranked first.
Maximum speed of 74 km/h on structured roads DARPA races foster the development of intelligent vehicles.
2009 - The first Intelligent Vehicle Future Challenge of China
The race involved obstacles, traffic lights, and hairpin turns.
Teams from HNU, BIT, SJTU, XJTU, and University of Parma competed in the
challenge.
20 Huawei Confidential
Intelligent Driving Will Lift Society to New Heights
About 1.2 million people die from traffic
accidents every year.
Huge reductions in In 2019, more than 60,000 people died
from traffic accidents in China.
road accidents 90% of traffic accidents are caused by
human error.
Third living space
Convenient and
efficient travel
"Mobile third space": Intelligent vehicles
automate driving, to give drivers and
passengers a multi-functional space for new
entertainment, social networking, and retail Intelligent driving enables special groups to drive on
scenarios. roads.
With an emphasis on passenger experience, China - 75% of roads in major cities suffer from
consumer demand will concern industries traffic jams during peak hours, with an average
such as media, entertainment, retail, and traffic delay of one hour.
Intelligent driving Vehicle-road collaboration using an intelligent
health. Its reported 63% of people use the
time traveling to socialize, and nearly 25% transportation system (ITS) can greatly improve
want to spend the time on fitness. traffic efficiency.
University of Michigan estimates one shared
autonomous vehicle can replace 9.34 conventional
vehicles.
21 Huawei Confidential
Manual Driving vs. Intelligent Driving
How do I get there?
Where am I? Where am I going? Sensing, positioning, map, and planning
Sensing, positioning, and mapping Navigation and map
and control (PnC)
Sensing & positioning Eyes & ears Cameras & radars
Decision-making & planning Brain Central control unit
Control Hands & feet Steering wheel & pedals
22 Huawei Confidential
Three Subsystems of Intelligent Driving
Sensing Thinking Acting
Ultrasonic radar Sensing Drive system
Millimeter-wave radar Convergence Braking system
Camera Positioning Steering system
Lidar Planning Drive train
5G/Vehicle-to- Prediction Security system
everything (V2X) Decision-making Comfort system
...... Control .......
......
23 Huawei Confidential
Huawei MDC: Leading Intelligent Driving Computing Platform
Personal travel Long-distance Campus Street Freight
RoboTaxi Logistics Third-party scenario-specific
transportation commute cleaning transportation
algorithms
(from partners)
MDC Intelligent driving computing platform
Huawei hardware (CPU/AI processor) + Software (vehicle control OS platform) + Tool
chain
Millimeter-wave/ Sensors/X-by-wire
Camera Lidar T-BOX GPS Chassis-by-wire
ultrasonic radar (from partners)
24 Huawei Confidential
Autonomous Driving Cloud Service: One-Stop Development Platform for
Autonomous Vehicles
Data services:
Data storage management, processing pipeline, overview, and
playback, as well as an annotation platform
Training services:
Training tasks, algorithm and model management, algorithm
iteration, and model evaluation
Simulation services:
Simulated scenarios, online simulations, task management, and
simulation algorithms
Competitive advantages:
Annotation platform
Model training
Parallel simulation
25 Huawei Confidential
26 Huawei Confidential
Summary
This chapter briefly introduces concepts of several cutting-edge AI technologies:
obtaining the optimal policy through reinforcement learning, the game between the
generator and discriminator (GAN), the construction process and applications of
knowledge graphs, and brief introduction to intelligent driving.
27 Huawei Confidential
Quiz
1. (Multiple-answer) Which of the following scenarios can Knowledge Graph be used? ( )
A. Search engine
B. Product keyword search
C. Q&A bot
2. (True or false) In intelligent driving, lidar is the only sensing device. ( )
A. True
B. False
28 Huawei Confidential
Thank you. 把数字世界带入每个人、每个家庭、
每个组织，构建万物互联的智能世界。
Bring digital to every person, home, and
organization for a fully connected,
intelligent world.
Copyright©2023 Huawei Technologies Co., Ltd.
All Rights Reserved.
The information in this document may contain predictive
statements including, without limitation, statements regarding the
future financial and operating results, future product portfolio, new
technology, etc. There are a number of factors that could cause
actual results and developments to differ materially from those
expressed or implied in the predictive statements. Therefore, such
information is provided for reference purpose only and constitutes
neither an offer nor an acceptance. Huawei may change the
information at any time without notice.
