-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
519 lines (443 loc) · 27.3 KB
/
index.html
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
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
---
layout: default
title: Home
---
<!-- ========== History ========== -->
<div class="docs-section" id="bio">
<h4>Bio</h4>
<p>
I am a postdoctoral associate at the graduate school of artificial intelligence in Ulsan National Institute of Science and Technology (UNIST). In February 2023, I received my Ph.D. degree and won the best researcher award for my thesis titled <i>"Tackling Three Problems in Controlled Sequence Generations: Bridging Reinforcement Learning with Language Models."</i>.
</p>
<p>
My research interest lies in the development of large language models (LLMs) and reinforcement learning (RL) with their application in AI4Healthcare & AI4Science. In particular, leveraging RL to align LLMs with human preferences is of the most interest that I am currently focusing on. In this regard, my research topics range from fundamental to applied LLMs+RL research, including but not limited to:
</p>
<dl style="PADDING-LEFT: 40px">
<dt>
<li>Developing fundamental frameworks, algorithms, and building blocks for controllable LLMs.</li>
<li>Designing alignment mechanisms to enhance the interaction between LLMs and humans.</li>
<li>Deploying an LLM-based service to solve real-world problems in healthcare and science (e.g., nutrition coaching, drug discovery, etc.).</li>
</dt>
</dl>
<!--
<dt>
<li><b>Optimality</b></li>
</dt>
<dd>
<p>
<b>Control over AI must be considered with guaranteed results.</b> We need to understand the optimality of the objective function or training algorithm and thereby provide guarantees about the results.
</p>
<!--
<p>To build controllable AI, the generative model must be optimized for the likelihood and control objectives simultaneously (jointly). Given this multi-objective nature, studying controllability from the perspective of Pareto optimality can provide the theoretical foundations of controllable AI.</p>
-->
<!--
<li>
Pareto Optimal
</li>
<li>
Objective Mismatch
</li>
</dd>
<dt>
<li>Robustness to Changing Environment and Adeversarial Attack</li>
<li>Exposure Bias and Data Perturbation</li>
<li><b>Maximum Mutual Information</b></li>
<li><b>Dynamics</b></li>
</dt>
<dd>
<p>
<b>Control over AI must be considered in dynamic situations.</b> We need to recognize and reflect that AI is subject to many uncertainties (e.g., interaction with users or other AIs, long-term distribution shifts in training datasets, adversarial attacks on AI from the outside, etc.).
</p>
</p>
<p>In terms of Pareto optimality, the more the Pareto improvement, the higher the controllability. Since mutual information represents the joint probability of two random variables, mapping both likelihood and control objectives onto the probability space together and maximizing their mutual information will derive the Pareto improvement.</p>
<li>
Generalization under Out-of-Distribution
</li>
<li>
Robustness from Adversarial Attack
</li>
<li>
Domain Adaptation with Changing Environment
</li>
</dd>
<dt>
<li>Applicability</li>
<li>Latent Hole and Continuous Representation</li>
<li><b>Approximation Methods</b></li>
<li><b>Equilibrium</b></li>
</dt>
<dd>
<p>
<b>Control over AI must be considered at the equilibrium.</b> We need to seek an equilibrium of AI control that achieves long-term optimality under any dynamic situation.
</p>
<p>Computing mutual information is expensive, so it is necessary to "approximate" the maximum mutual information without estimating the exact amount of mutual information. Therefore, in my research, variational inference or perturbation techniques are the most preferred methodology and a research topic in itself.</p>
Increases the usecases of AI in industrial fields such as A, B, and C.
<li>
Definition of Domain-driven Problems
</li>
</dd>
<dt>
<li><b>Gamification, Simulation & Mechanism Design</b></li>
</dt>
<dd>
<p>
<b>Control over AI must be tested from the perspective of whether what was intended has been followed.</b> To test all possible intended scenarios, it is best for researchers to be able to arbitrarily design and manipulate experimental settings at will.
</p>
</dd>
-->
<!--
<p>Please check out my latest work at <b><a href="https://arxiv.org/abs/2310.04483"><font color="violet">arXiv</font></a></b>.</p>
<p>
I received my Ph.D. degree in February 2023 and won the best researcher award for my thesis titled <i>"Tackling Three Problems in Controlled Sequence Generations: Bridging Reinforcement Learning with Language Models."</i>. Currently, I am a post-doc associate at the graduate school of artificial intelligence at UNIST, conducting <a href="assets/publications/2023_review/paper.pdf" target="_blank">research</a> to analyze controllability and improve the control performance of LLMs.
</p>
-->
</div>
<!-- ========== PUBLICATIONS ========== -->
<div class="docs-section" id="publications">
<h4>Publications</h4>
<p>Most recent publications on <a href="{{ site.data.main_info.google_scholar }}" target="_blank">Google Scholar</a>.<br/>
<sup>‡</sup> indicates equal contribution.
</p>
<ul class="tab-nav">
<li><div class="button active" data-ref="#papers-selected">Selected</div></li>
<li><div class="button" data-ref="#papers-application">Application</div></li>
<li><div class="button" data-ref="#papers-research">Research</div></li>
<li><div class="button" data-ref="#papers-all">All</div></li>
</ul>
<div class="tab-content">
<div class="tab-pane active" id="papers-selected">
{% assign selected_papers = site.data.publications.papers | where: "selected", "y" %}
{% for paper in selected_papers %}
<div class="paper">
<p class="title"><b>{{ paper.title }}</b></p>
<p>{{ paper.authors }}</p>
<p><i>{{ paper.venue }}</i></p>
<div class="paper-buttons">
{% if paper.paper_pdf %}
<a class="button" href="{{ paper.paper_pdf | prepend: site.baseurl }}" target="_blank">Paper</a>
{% endif %}
{% if paper.slides %}
<a class="button" href="{{ paper.slides | prepend: site.baseurl }}" target="_blank">Slides</a>
{% endif %}
{% if paper.poster %}
<a class="button" href="{{ paper.poster | prepend: site.baseurl }}" target="_blank">Poster</a>
{% endif %}
{% if paper.video %}
<a class="button" href="{{ paper.video }}" target="_blank">Video</a>
{% endif %}
{% if paper.code %}
<a class="button" href="{{ paper.code }}" target="_blank">Code</a>
{% endif %}
{% if paper.data %}
<a class="button" href="{{ paper.data }}" target="_blank">Data</a>
{% endif %}
</div>
</div>
{% endfor %}
</div>
<div class="tab-pane" id="papers-application">
{% assign application_papers = site.data.publications.papers | where: "application", "y" %}
{% for paper in application_papers %}
<div class="paper">
<p class="title"><b>{{ paper.title }}</b></p>
<p>{{ paper.authors }}</p>
<p><i>{{ paper.venue }}</i></p>
<div class="paper-buttons">
{% if paper.paper_pdf %}
<a class="button" href="{{ paper.paper_pdf | prepend: site.baseurl }}" target="_blank">Paper</a>
{% endif %}
{% if paper.slides %}
<a class="button" href="{{ paper.slides | prepend: site.baseurl }}" target="_blank">Slides</a>
{% endif %}
{% if paper.poster %}
<a class="button" href="{{ paper.poster | prepend: site.baseurl }}" target="_blank">Poster</a>
{% endif %}
{% if paper.video %}
<a class="button" href="{{ paper.video }}" target="_blank">Video</a>
{% endif %}
{% if paper.code %}
<a class="button" href="{{ paper.code }}" target="_blank">Code</a>
{% endif %}
{% if paper.data %}
<a class="button" href="{{ paper.data }}" target="_blank">Data</a>
{% endif %}
</div>
</div>
{% endfor %}
</div>
<div class="tab-pane" id="papers-research">
{% assign research_papers = site.data.publications.papers | where: "research", "y" %}
{% for paper in research_papers %}
<div class="paper">
<p class="title"><b>{{ paper.title }}</b></p>
<p>{{ paper.authors }}</p>
<p><i>{{ paper.venue }}</i></p>
<div class="paper-buttons">
{% if paper.paper_pdf %}
<a class="button" href="{{ paper.paper_pdf | prepend: site.baseurl }}" target="_blank">Paper</a>
{% endif %}
{% if paper.slides %}
<a class="button" href="{{ paper.slides | prepend: site.baseurl }}" target="_blank">Slides</a>
{% endif %}
{% if paper.poster %}
<a class="button" href="{{ paper.poster | prepend: site.baseurl }}" target="_blank">Poster</a>
{% endif %}
{% if paper.video %}
<a class="button" href="{{ paper.video }}" target="_blank">Video</a>
{% endif %}
{% if paper.code %}
<a class="button" href="{{ paper.code }}" target="_blank">Code</a>
{% endif %}
{% if paper.data %}
<a class="button" href="{{ paper.data }}" target="_blank">Data</a>
{% endif %}
</div>
</div>
{% endfor %}
</div>
<div class="tab-pane" id="papers-all">
{% for paper in site.data.publications.papers %}
<div class="paper">
<p class="title"><b>{{ paper.title }}</b></p>
<p>{{ paper.authors }}</p>
<p><i>{{ paper.venue }}</i></p>
<div class="paper-buttons">
{% if paper.paper_pdf %}
<a class="button" href="{{ paper.paper_pdf | prepend: site.baseurl }}" target="_blank">Paper</a>
{% endif %}
{% if paper.slides %}
<a class="button" href="{{ paper.slides | prepend: site.baseurl }}" target="_blank">Slides</a>
{% endif %}
{% if paper.poster %}
<a class="button" href="{{ paper.poster | prepend: site.baseurl }}" target="_blank">Poster</a>
{% endif %}
{% if paper.video %}
<a class="button" href="{{ paper.video }}" target="_blank">Video</a>
{% endif %}
{% if paper.code %}
<a class="button" href="{{ paper.code }}" target="_blank">Code</a>
{% endif %}
{% if paper.data %}
<a class="button" href="{{ paper.data }}" target="_blank">Data</a>
{% endif %}
</div>
</div>
{% endfor %}
</div>
</div>
</div>
<!-- ========== RESEARCH INTEREST ========== -->
<!--
<div class="docs-section" id="research-interest">
<h4>Research Interest</h4>
<div class="image-container">
<img src='{{ site.baseurl }}{{ site.data.images.pareto_image }}' class="u-max-full-width"></a>
</div>
<div class="project-caption" style="text-align:left">
<b>Figure.1 (Multi-objective Nature of CLMs (far-left) and How to achieve Pareto Improvement (far-right)):</b> The Pareto frontier is a boundary within the objective space where the set of optimal policies π* belongs to. Each point represents a combination of objectives achieved by the policy at that point. The <b><span style="color:red">red</span></b> arrow indicates that we cannot increase reward without sacrificing likelihood, while the <b><span style="color:blue">blue</span></b> arrow indicates vice versa. The <b><span style="color:purple">purple</span></b> arrow and <b><span style="color:orange">orange</span></b> area denote potential direction and region, respectively, that improves both objectives simultaneously.
</div>
<br>
<p>
<u>My research interest is to develop controllable language models (CLMs) for controllable sequence generation (CSG).</u> In general, I implement CLMs by treating LMs as agents and leveraging policy-based reinforcement learning (e.g., REINFORCE, TRPO, PPO, etc.,) to fine-tune the LM agents with human-instruction rewards.
<p>
<u>More recently, my research focus has shifted to establishing the theoretical foundations of CLM</u> because I believe that relying on rules-of-thumb (i.e., unprincipled) practices in out-of-control situations can be disastrous in areas that will be increasingly dependent on AI (e.g., finance, medical, security, transportation, etc.). At this time, a theoretical understanding of the controllability will help minimize this risk.
</p>
<p>
<u>Figure 1 describes my recent approach to CLMs from a theoretical perspective.</u> At the core of my approach is the assumption that <u><i>"developing controllable AI is equivalent to solving multi-objective optimization."</i></u> To be specific, controlling an AI model requires it to approximate real data distribution (i.e., likelihood objective) and obey user-intended directions (i.e., reward objective) simulataneously. <u>That is, developing CLMs is a problem of searching for an optimal solution where both likelihood and reward objectives are satisfied at the same time.</u>
</p>
-->
<!--
<p>
Accordingly, <u>I hypothesized that the optimal solutions lie on the Pareto frontier,</u> as is typical for multi-objective optimization problems, <u>and that it is inevitable to sacrifice the likelihood for further increasing the reward.</u> These two hypotheses establish the theoretical foundation for the controllability of AI models, which is a unique aspect that distinguishes my approach from previous studies. You can see my work in recent papers that are <a href="assets/publications/2023_review/paper.pdf" target="_blank">under review</a> and <a href="assets/publications/2023_working/paper.pdf" target="_blank">in progress</a>.
</p>
-->
<!--
<p>
Accordingly, <u>I hypothesized that the optimal solutions lie on the Pareto frontier,</u> as is typical for multi-objective optimization problems, <u>and that it is inevitable to sacrifice the likelihood for further increasing the reward.</u> These two hypotheses establish the theoretical foundation for the controllability of AI models, which is a unique aspect that distinguishes my approach from previous studies. You can see <a href="assets/publications/2023_review/paper.pdf" target="_blank">my recent work under review</a>.
</p>
<p>
For the last two years of Ph.D., <u>I have published several papers proposing novel CLMs that specialize in real-world applications, </u> including diet planning [<a href="assets/publications/2021_diet_planning/paper.pdf" target="_blank">KDD'21</a>, <a href="assets/publications/2021_MIND_dataset/paper.pdf" target="_blank">NeurIPS'21</a>], operational recommendations [<a href="assets/publications/2022_recommendation_in_offline_stores/paper.pdf" target="_blank">KDD'22</a>], and drug repurposing [<a href="assets/publications/2023_working/paper.pdf" target="_blank">in progress</a>]. Figure 2 presents application examples of CSG, which can be addressed by CLMs.
<div class="image-container" style="float:left">
<img src='{{ site.baseurl }}{{ site.data.images.diet_planning }}' class="u-max-full-width" style="width:37.5%; float:left"></a>
<img src='{{ site.baseurl }}{{ site.data.images.drug_repurposing }}' class="u-max-full-width" style="width:60%; float:right"></a>
</div>
<div class="project-caption" style="text-align:left">
<b>Figure.2 (Application Examples of CSG - diet planning (left) and drug repurposing (right))</b>: The data of interest is defined in a sequence format (e.g., diet sequence and compound sequence). The original sequence (i.e., a picture with the caption "before" in diet planning and "base" in drug repurposing examples, respectively) is edited into the desirable sequence whose target property (i.e., a color of diet and a target indication of drug) is controlled as intended.
</div>
</p>
-->
<!--
<p>
This research area is dissected into three elements: 1) sequential model, 2) control theory, and 3) generative framework. Therefore, the three elements must be considered in CSG research, and I usually take them into account according to the three steps as follows:
<ol style="PADDING-LEFT:40px">
<li>Building the langauge models of high capacities to estimate sequential patterns (e.g., Seq2Seq, Transformers, GPT3, etc.,).</li>
<li>Leveraging policy-based RL algorithms to induce controllability (e.g., REINFORCE, PPO, TRPO, etc.,).</li>
<li>Optimizing the language models to be controllable using RL algorithms under the generative framework of unsupervised or self-supervised manners (e.g., AE, VAE, etc.,).</li>
</ol>
</p>
<p>
As mentioned above, my research boundary is grounded on the middle territory between NLG and RL. On this background, my research interest in CSGs is recently extending from real-world applications to theoretical understanding. Specifically, <b>I study the theoretical foundations of how RL and NLP techniques get intertwined and fused on the lens of Bayesian perspectives.</b> There are four reasons why I prefer the Bayesian approach:
<ol style="PADDING-LEFT:40px">
<li>Bayesian methods are one of the most straightforward tools for dealing with conditional distributions.
<ul>
<li>RL is a learning framework that estimates the conditional distribution over actions given states according to the Markov decision process.</li>
<li>Language model describes a sequence by factorizing the joint distribution of words into the product of conditional distributions of tokens given previous tokens.</li>
</ul>
</li>
<li>Bayesian methods provide us with a powerful backdoor, called Variational Bayes, to approximate intractable distributions difficult to estimate directly.
<ul>
<li>Variational inference (VI), a mathematical framework for (approximately) doing Bayesian inference, is even computationally cheap because it can be converted to a dual problem, i.e., maximizing ELBO.</li>
</ul>
</li>
<li>Bayesian methods imply the probabilistic status of unobserved states, and it means implicit data augmentation comes into play in the latent space.
<ul>
<li>RL requires as many observations as possible to estimate a policy that maximizes the reward.</li>
</ul>
</li>
<li>Bayesian methods are open to introducing tricks such as marginalization so that we can easily derive useful properties such as the amount of information.</li>
</ol>
A paper titled " <a href="assets/publications/2023_levenshtein_agent/paper.pdf" target="_blank">Theoretical Principles of Controllable Generation</a> ," which deals with some of the topics related to the above four contents, was submitted to ICML'23, and is currently in the process of review. Recently, <b>I am thinking about how to introduce the concept of augmentation into CSG research in the causal perspective through the Bayesian method.</b>
</p>
</div>
-->
<!-- ========== FUNDINGS ========== -->
<div class="docs-section" id="funding">
<h4>Fundings</h4>
<p>
As of September 1, 2023, I have a research grant from the National Research Foundation of Korea for the next 2 years, with permission to use this grant to stay abroad for up to 6 months. This means that <b>I can support myself for 6 months and conduct research as a visiting researcher with the status of a self-funded resident fellow before the official postdoctoral associate period begins.</b> Please feel free to contact me at <a href="mailto:[email protected]">[email protected]</a>.
<!-- Since 09/01/2023, I have been granted by the National Research Foundation of South Korea for the next 2yrs. According to the terms of the grant, I can use this grant to stay abroad for up to six months. In other words, <b>I can support myself for up to 6 months and immediately conduct research as a visiting researcher before the official postdoctoral program begins.</b> Please feel free to contact me at <a href="mailto:[email protected]">[email protected]</a>. -->
</p>
</div>
<!-- ========== POSTDOCS ========== -->
<div class="docs-section" id="postdoc">
<h4>Co-work Opportunities</h4>
<p>
I am currently <b>looking for a post-doctoral position related to NLP and RL research,</b> with reponsibilities including:
<ol style="PADDING-LEFT:40px">
<li>studying theoretical foundations of controllable/reliable LLMs, </li>
<li>and applying them to solve real-world problems in AI4Healthcare and AI4Science contexts.</li>
</ol>
Of course, I am also open to any other interesting topics as long as they are relevant to NLP and RL research. If you are interested, please contact me at <a href="mailto:[email protected]">[email protected]</a>.
</p>
</div>
<!--
<div class="docs-section" id="prospective-students">
<h4>Prospective Students</h4>
<p>
<b>I'm recruiting PhD students to start Fall 2023 at the University of Washington! </b>
I'm especially looking for students with a background and/or interest in <i>social network analysis, machine learning, causal inference, or natural language processing</i>, and are passionate about using computational methods to study how online social platforms can be reimagined to <i>enable better conversations, bridge political divides, and reduce the spread of misinformation</i>.
If you are interested, please apply to the UW iSchool <a href="https://ischool.uw.edu/programs/phd/admissions" target="_blank">PhD program</a> and mention me in your application.
If you have any questions, feel free to reach out by email!
</p>
</div>
-->
<!-- ========== PROJECTS ==========
<div class="docs-section" id="projects">
<h4>Projects</h4>
<ul class="tab-nav">
<li><div class="button active" data-ref="#projects-selected">Selected</div></li>
<li><div class="button" data-ref="#projects-all">All</div></li>
</ul>
<div class="tab-content">
<div class="tab-pane active" id="projects-selected">
{% assign selected_projects = site.data.projects.projects | where: "selected", "y" %}
{% for project in selected_projects %}
{% assign index_modulo = forloop.index0 | modulo:3 %}
{% if index_modulo == 0 %}
<div class="row">
{% endif %}
<div class="four columns">
<div class="project-container">
<div class="project-image-container">
<a href="{{ project.url }}">
<img src="{{ project.thumbnail }}" class="u-max-full-width" />
</a>
</div>
<div class="project-caption">
<b>{{ project.title }}</b> <br/>
{{ project.subtitle }}
</div>
</div>
</div>
{% if index_modulo == 2 %}
</div>
{% endif %}
{% endfor %}
</div>
<div class="tab-pane" id="projects-all">
{% for project in site.data.projects.projects %}
{% assign index_modulo = forloop.index0 | modulo:3 %}
{% if index_modulo == 0 %}
<div class="row">
{% endif %}
<div class="four columns">
<div class="project-container">
<div class="project-image-container">
<a href="{{ project.url }}">
<img src="{{ project.thumbnail }}" class="u-max-full-width" />
</a>
</div>
<div class="project-caption">
<b>{{ project.title }}</b> <br/>
{{ project.subtitle }}
</div>
</div>
</div>
{% if index_modulo == 2 %}
</div>
{% endif %}
{% endfor %}
</div>
</div>
</div>
-->
<!-- ========== BIO ========== -->
<div class="docs-section" id="history">
<h4>History</h4>
<p>
<span style="color:red">In February 2015</span>, I got my B.Sc in Economics from Ajou University as an outstanding student in <u>the top 10% of the major GPA.</u>, and I enrolled in the master's program in the Department of Industrial Engineering at Ulsan National Institute of Science and Technology (UNIST) <span style="color:red">in February 2016</span>.<br></br>
For the first one and a half years <span style="color:red">(Jan 2016 - June 2017)</span>, I focused on <u>studying quantitative research methodologies (e.g., advanced statistics) and sequential modeling (e.g., time-series analysis)</u>. Then, I spent half a year as a data analyst intern at Hyundai Mipo Dockyard <span style="color:red">(July 2017 - Dec 2017)</span>, <u>developing language models (LMs) and text-mining algorithms.</u> Completing the internship, I returned to UNIST and transferred to the combined Master-Ph.D program.<br><br>
I spent the first two years of Ph.D. course <span style="color:red">(Jan 2018 - Jan 2020)</span> in <u>analyzing real-world problems with advanced text-mining techniques.</u> For the last three years <span style="color:red">(Feb 2020 - Dec 2022)</span>, I focused on <u>solving real-world problems using controllable language models (CLMs).</u> <br></br>
<span style="color:red">In February 2023</span>, <u>I wrote my Ph.D thesis on controllable sequence generation,</u> titled <i>"Tackling Three Problems in Controlled Sequence Generations: Bridging Reinforcement Learning with Language Models."</i> and <u>won the best researcher award</u> in UNIST. <br></br>
<span style="color:red"> From March 2023</span>, <u> I'm joining the Graduate School of Artificial Intelligence at UNIST as a post-doc associate</u>. Currently, I belong to <a href="https://service.unist.ac.kr/" target="_blank">the Service Intelligence Lab</a> that is led by <a href="https://scholar.google.co.kr/citations?user=NW7F08MAAAAJ&hl=ko&oi=ao" target="_blank">Prof.Chiehyeon Lim</a>, my advisor whom I truly respect and appreciate.
</p>
<p>
I dream of becoming a researcher who is willing to stand up against prejudice and inequality in the world with logical thinking and a scientific attitude.
I am looking forward to a heart-pounding journey of research that will create a better world and promote the future with great colleagues.
I will end this article by introducing a quote from one of my favorite scholars, Alfred Marshall:
</p>
<p style="font-size: large">
<i>"Cool head, but warm heart!"</i>
</p>
</div>
<!-- ========== RESUME ========== -->
<div class="docs-section" id="resume">
<h4>Vitæ</h4>
<!--
<p>You can download my <a href={{ "/assets/cv/CV_Changhun_Lee.pdf" | prepend: site.baseurl }} target="_blank">Resume</a> and <a href={{ "/assets/cv/Cover_Letter_Changhun_Lee.pdf" | prepend: site.baseurl }} target="_blank">Cover Letter</a> in PDF.</p>
-->
<p>You can download my <a href={{ "/assets/cv/CV_Changhun_Lee.pdf" | prepend: site.baseurl }} target="_blank">Curriculum Vitae</a> in PDF file.</p>
<!-- The Timeline -->
<ul class="timeline">
{% for exp in site.data.experience.experiences %}
<li>
{% if exp.category == "work" %}
<div class="direction-l">
{% else %}
<div class="direction-r">
{% endif %}
<div class="flag-wrapper">
<span class="flag">{{ exp.place }}</span>
<span class="time-wrapper"><span class="time">{{ exp.time }}</span></span>
</div>
<div class="desc"><b>{{ exp.title }}</b> <br/> {{ exp.subtitle }}</div>
</div>
</li>
{% endfor %}
</ul>
</div>
<div class="docs-section" id="ack">
<h4>Acknowledgement</h4>
<p>
This website uses the website design and template by <a href="https://github.com/msaveski/www_personal">Martin Saveski</a>
</p>
</div>