Repository navigation
Expand file tree
/
Copy pathexperience.html
More file actions
255 lines (255 loc) · 19.2 KB
/
Copy pathexperience.html
File metadata and controls
255 lines (255 loc) · 19.2 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.1//EN"
"http://www.w3.org/TR/xhtml11/DTD/xhtml11.dtd">
<html xmlns="http://www.w3.org/1999/xhtml" xml:lang="en">
<head>
<meta name="generator" content="jemdoc, see http://jemdoc.jaboc.net/"/>
<meta http-equiv="Content-Type" content="text/html;charset=utf-8"/>
<link rel="stylesheet" href="experience_css.css" type="text/css"/>
<title>Duo Wu (吴铎)</title>
</head>
<body>
<table summary="Table for page layout." id="tlayout">
<tr valign="top">
<td id="layout-menu">
<div class="menu-category">Duo Wu</div>
<div class="menu-item"><a href="index.html">Home</a></div>
<div class="menu-item"><a href="awards.html">Awards</a></div>
<div class="menu-item"><a href="experience.html" class="current">Experience</a></div>
<div class="menu-category">Life</div>
<div class="menu-item"><a href="moments.html">Moments</a></div>
<div class="menu-item"><a href="travel.html">Travel</a></div>
</td>
<td id="layout-content">
<!-- <div id="toptitle">-->
<!-- <h1>Experience</h1>-->
<!-- </div>-->
<h2>Research Experience</h2>
<h3>MANSY: Generalizing Neural Adaptive Immersive Video Streaming With Ensemble and Representation
Learning[2023.2 ~ 2023.11]</h3>
<div class="image">
<img src="picture/experience/system_framework.png" alt="system_framework.png" width="60%">
</div>
<ul>
<li><p><b>Abstract:</b> The popularity of immersive videos has prompted extensive research into neural
adaptive tile-based streaming to optimize video transmission over networks with limited bandwidth.
However, the diversity of users' viewing patterns and Quality of Experience (QoE) preferences has
not been fully addressed yet by existing neural adaptive approaches for viewport prediction and
bitrate selection. Their performance can significantly deteriorate when users' actual viewing
patterns and QoE preferences differ considerably from those observed during the training phase,
resulting in poor generalization. In this paper, we propose MANSY, a novel streaming system that
embraces user diversity to improve generalization. Specifically, to accommodate users' diverse
viewing patterns, we design a Transformer-based viewport prediction model with an efficient
multi-viewport trajectory input output architecture based on implicit ensemble learning. Besides,
we for the first time combine the advanced representation learning and deep reinforcement learning
to train the bitrate selection model to maximize diverse QoE objectives, enabling the model to
generalize across users with diverse preferences. Extensive experiments demonstrate that MANSY
outperforms state-of-the-art approaches in viewport prediction accuracy and QoE improvement on both
trained and unseen viewing patterns and QoE preferences, achieving better generalization.
</p></li>
<li><p><b>Key words:</b> tile-based neural adaptive immersive video streaming, generalization, ensemble
learning, representation learning</p></li>
<li><p><b>Role:</b> First author.</p></li>
<li><b>Status:</b> Submitted to IEEE Transactions on Mobile Computing (CCF-A).
</li>
</ul>
<h3>ILCAS: Imitation Learning-Based Configuration-Adaptive Streaming for Live Video Analytics with Cross-Camera
Collaboration [2022.8 ~ 2023.1]</h3>
<div class="image">
<img src="picture/experience/System_Overview.png" alt="System_Overview.png" width="60%">
</div>
<ul>
<li><p><b>Abstract:</b> The high-accuracy and resource-intensive deep neural networks (DNNs) have been
widely adopted by live video analytics (VA), where camera videos are streamed over the network to
resource-rich edge/cloud servers for DNN inference. Common video encoding configurations (e.g.,
resolution and frame rate) have been identified with significant impacts on striking the balance
between bandwidth consumption and inference accuracy and therefore their adaption scheme has been a
focus of optimization. However, previous profiling-based solutions suffer from high profiling cost,
while existing deep reinforcement learning (DRL) based solutions may achieve poor performance due to
the usage of fixed reward function for training the agent, which fails to craft the application
goals in various scenarios. In this paper, we propose ILCAS, the first imitation learning (IL) based
configuration-adaptive VA streaming system. Unlike DRL-based solutions, ILCAS trains the agent with
demonstrations collected from the expert which is designed as an offline optimal policy that solves
the configuration adaption problem through dynamic programming. To tackle the challenge of video
content dynamics, ILCAS derives motion feature maps based on motion vectors which allow ILCAS to
visually “perceive” video content changes. Moreover, ILCAS incorporates a cross-camera collaboration
scheme to exploit the spatio-temporal correlations of cameras for more proper configuration
selection. Extensive experiments confirm the superiority of ILCAS compared with state-of-the-art
solutions, with 2-20.9% improvement of mean accuracy and 19.9-85.3% reduction of chunk upload lag.
</p></li>
<li><p><b>Key words:</b> live video analytics, configuration adaption, imitation learning,
cross-camera collaboration</p></li>
<li><p><b>Role:</b> First author.</p></li>
<!-- <li><p><b>My contributions:</b></p>-->
<!-- <ol>-->
<!-- <li><p>Propose to use the advanced imitation learning framework Generative Adversarial Imitation-->
<!-- Learning (GAIL) for optimizing live video analytics streaming, and design the core expert-->
<!-- policy based on dynamic programming to guide the training of agent.</p></li>-->
<!-- <li><p>Design the methodology of deriving motion feature map from motion vectors to enable-->
<!-- ILCAS to adaptively select configurations according to video dynamics.</p></li>-->
<!-- <li><p>Implement most parts of experiments and finish most of the work of paper writing.</p></li>-->
<!-- </ol>-->
<!-- </li>-->
<li><b>Status:</b> Accepted to appear in IEEE Transactions on Mobile Computing
(CCF-A) <i><a href="https://doi.org/10.1109/TMC.2023.3327097">[pdf]</a></i>.
</li>
</ul>
<h3>A Comprehensive Survey on Segment Routing Traffic Engineering [2019.11 ~ 2020.11]</h3>
<div class="image">
<img src="picture/experience/Survey_Overview.png" alt="Survey_Overview.png" width="45%">
</div>
<ul>
<li><p><b>Abstract:</b> Traffic Engineering (TE) enables management of traffic in a manner that
optimizes utilization of network resources in an efficient and balanced manner. However,
existing TE
solutions face issues relating to scalability and complexity. In recent years, Segment Routing
(SR)
has emerged as a promising source routing paradigm. As one of the most important applications of
SR,
Segment Routing Traffic Engineering (SR-TE), which enables a <i>headend</i> to steer traffic
along
specific paths represented as ordered lists of instructions called <i>segment lists</i>, has the
capability to overcome the above challenges due to its flexibility and scalability. In this
paper,
we conduct a comprehensive survey on SR-TE. A thorough review of SR-TE architecture is provided
in
the first place, reviewing the core components and implementation of SR-TE such as SR Policy,
Flexible
Algorithm and SR-native algorithm. Strengths of SR-TE are also discussed, as well as its major
challenges. Next, we dwell on the recent SR-TE researches on routing optimization with various
intents, e.g., optimization on link utilization, throughput, QoE (Quality of Experience) and
energy
consumption. Afterwards, node management for SR-TE are investigated, including SR node
deployment
and candidate node selection. Finally, we discuss the existing challenges of current research
activities and propose several research directions worth of future exploration.
</p></li>
<li><p><b>Key words:</b> segment routing, traffic engineering, SR policy, routing optimization,
segment
list computation</p></li>
<li><p><b>Role:</b> First author.</p></li>
<!-- <li><p><b>My contributions:</b></p>-->
<!-- <ol>-->
<!-- <li><p>Extensively collected and read relevant literature, wrote reports for research-->
<!-- papers.</p></li>-->
<!-- <li><p>Organized paper structure and finished paper writing.</p></li>-->
<!-- <li><p>Brainstormed several in-depth future research directions based on our survey work and-->
<!-- personal insights.</p></li>-->
<!-- </ol>-->
<!-- </li>-->
<li><b>Status:</b> Accepted to appear in Digital Communications and Networks
(Q1), 2022.<i><a href="https://doi.org/10.1016/j.dcan.2022.02.006">[PDF]</a></i></li>
</ul>
<!-- <h3>Research on Micro-loop in Heterogeneous Segment Routing Networks [2020.12 ~ 2021.05]</h3>-->
<!-- <div class="image">-->
<!-- <img src="picture/experience/microloop.png" alt="microloop.png" width="60%">-->
<!-- </div>-->
<!-- <ul>-->
<!-- <li><p><b>Abstract:</b> Micro-loop refers to a looping phenomenon entailed by the unsynchronized network-->
<!-- convergence, which can cause serious negative effects (e.g. packet loss and bandwidth exhaustion) to-->
<!-- networks. Segment Routing (SR) provides a complete and simple solution for micro-loop avoidance.-->
<!-- However, such solution can not effectively avoid micro-loops in heterogeneous networks where there-->
<!-- are routers not supporting the SR functions. This research proposes a distributed solution which-->
<!-- efficiently avoids micro-loops in heterogeneous SR networks by carefully crafting ‘lies’ (OSPF type-->
<!-- 5 LSAs with specific forwarding information) to deceive routers to change their routing tables.-->
<!-- Performance evaluation results show that the proposal can reduce almost 60% of the average duration-->
<!-- of micro-loops without causing too much networking overhead.</p></li>-->
<!-- <li><p><b>Key words:</b> micro-loop avoidance; heterogeneous segment routing networks; OSPF type 5 LSA-->
<!-- </p></li>-->
<!-- <li><p><b>Role:</b> Project leader.</p></li>-->
<!-- <li><p><b>Major work:</b></p>-->
<!-- <ol>-->
<!-- <li><p>Designed the workflow of the proposed solution and completed its details.</p></li>-->
<!-- <li><p>Developed the core algorithm for crafting LSAs based on Dijkstra shortest-path-->
<!-- algorithm.</p>-->
<!-- </li>-->
<!-- <li><p>Completed the experiment design and implemented all experiments.</p></li>-->
<!-- </ol>-->
<!-- </li>-->
<!-- <li><b>Notes:</b> This project was granted by the Innovation and Entrepreneurship Training Program-->
<!-- for College Students, Jinan University, (暨南大学大创项目), No. CX20005.-->
<!-- </li>-->
<!-- </ul>-->
<!-- <h3>Research on Accelerating DNN Training on End Device Based on Cloud and Edge Computing [2021.12 ~-->
<!-- 2022.04]</h3>-->
<!-- <div class="image">-->
<!-- <img src="picture/model_parallelism.png" alt="model_parallelism.png" width="60%">-->
<!-- </div>-->
<!-- <ul>-->
<!-- <li><p><b>Abstract:</b> With the rapid development of deep learning, there is an increasing demand on-->
<!-- deploying deep neural networks (DNNs) on end devices to push AI services to the edge of network.-->
<!-- However, it is impractical to train DNNs on end devices with limited resources. Besides, outsourcing-->
<!-- the training task along with user’s private training dataset to the third party (e.g., cloud and-->
<!-- edge servers) will entail extra network delay and lead to privacy leakage. To solve the above-->
<!-- challenges, this thesis focuses on the cloud-edge-end based DNN training acceleration. The-->
<!-- distributed training scheme based on model parallelism is used to speedup training and reduce device-->
<!-- power consumption, which partitions DNN into several successive sub-models and deploys them on the-->
<!-- end, edge and cloud for collaborative training. This thesis first proposes the system model to-->
<!-- evaluate the relevant metrics in DNN parallel training. Afterwards, based on dynamic programming, an-->
<!-- algorithm called Energy-constrained Partitioning Algorithm (EPA) is designed, which provides the-->
<!-- optimal partitioning solutions in polynomial time. Experimental results show that compared to local-->
<!-- training, EPA can speedup DNN training by 2.51× and reduce power consumption by 57.7%.</p></li>-->
<!-- <li><p><b>Key words:</b> DNN training; cloud and edge computing; model parallelism; model partitioning-->
<!-- algorithm</p></li>-->
<!-- <li><p><b>Type:</b> Thesis.</p></li>-->
<!-- <li><p><b>Major work:</b></p>-->
<!-- <ol>-->
<!-- <li><p>Proposed the system model and formulated the model partitioning problem.</p></li>-->
<!-- <li><p>Developed the Energy-constrained Partitioning Algorithm (EPA) based on dynamic-->
<!-- programming.</p>-->
<!-- </li>-->
<!-- <li><p>Designed the experiments, used PyTorch to implement DNN parallel training and Mininet to-->
<!-- instantiate network topology and virtual machines.</p></li>-->
<!-- </ol>-->
<!-- </li>-->
<!-- </ul>-->
<h2> Internship Experience </h2>
<h3> Summer Internship at Huawei Technology Co., Ltd. [2021.07 ~ 2021.09]</h3>
<table id="tab_intern_img">
<tr valign="top">
<td class="image_grid">
<div class="image">
<img src="picture/experience/huawei_intern1.jpg" alt="huawei_intern1.jpg" width="70%">
</div>
</td>
<td class="image_grid">
<div class="image">
<img src="picture/experience/huawei_intern2.jpg" alt="huawei_intern2.jpg" width="70%">
</div>
</td>
</tr>
</table>
<!-- <div class="image">-->
<!-- <img src="picture/experience/huawei_intern.jpg" alt="huawei_intern.jpg" width="50%">-->
<!-- </div>-->
<ul>
<li><p><b>Duty:</b> Engaged in the development of Java Web backend server based on Spring Boot
framework; responsible for the design and development of FRUD Kit in Heavenly Pond Architecture.
</p>
</li>
<li><p><b>Position:</b> Software engineer.</p></li>
<!-- <li><p><b>Major work:</b></p>-->
<!-- <ol>-->
<!-- <li><p>Implemented the relevant services of FRUD temporary files and completed unit-->
<!-- testing.</p>-->
<!-- </li>-->
<!-- <li><p>Decomposed the module of FRUD file generation with HIPO and Class Diagram, finally-->
<!-- designed 7 classes and 16 sub-modules and implemented the core components.</p>-->
<!-- </li>-->
<!-- <li><p>Used topological-sort and recursive algorithms to generate CDR topology data, with-->
<!-- data-->
<!-- structures such as Queue, Hash Set/Map employed to improve efficiency.</p></li>-->
<!-- </ol>-->
<!-- </li>-->
</ul>
<div id="footer">
<div id="footer-text">
<!-- Page generated 2022-07-07, by <a href="https://duowuyms.github.io/">Duo Wu</a>. -->
<br>Please <a href="mailto:duowu18@outlook.com">email</a> me if you have any constructive
suggestions on my homepage.
</div>
</div>
</td>
</tr>
</table>
</body>
</html>