-
Notifications
You must be signed in to change notification settings - Fork 82
/
index.html
294 lines (260 loc) · 25.4 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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8" />
<title>Eric Wallace — Home</title>
<link rel="stylesheet" href="master.css" />
<link href="https://fonts.googleapis.com/css?family=Roboto+Slab:300,400,700" rel="stylesheet">
<script>
function unhide(divID) {
var item = document.getElementById(divID);
if (item) {
item.className=(item.className=='hidden')?'unhidden':'hidden';
}
}
</script>
<link rel="icon" href="berkeley_nlp_logo.png" />
<meta name="author" content="Eric Wallace">
<meta name="keywords" content="Eric Wallace Berkeley NLP">
<meta name="robots" content="index,follow">
<meta name="description" content="Homepage of Eric Wallace Berkeley NLP">
</head>
<body>
<div class="wrapper">
<div class="posts-wrapper">
<div class="post">
<img src='full.jpg' style="float:right; width:325px"/>
<h1>Eric Wallace</h1>
<h2>[email protected] | <a style="font-size: 0.95em; font-weight:700" href="https://www.twitter.com/Eric_Wallace_" target="_blank">Twitter</a> | <a style="font-size: 0.95em; font-weight:700" href="https://scholar.google.com/citations?user=SgST3LkAAAAJ" target="_blank"> Scholar </a> | <a style="font-size: 0.95em; font-weight:700" href="https://www.github.com/Eric-Wallace" target="_blank">GitHub</a> |
<a style="font-size: 0.95em; font-weight:700" href="CV.pdf" target="_blank">CV</a></h2>
<br>
<br>
<p>Hi! I am a third-year PhD student at UC Berkeley working on Machine Learning and Natural Language Processing. I am advised by <a href="https://people.eecs.berkeley.edu/~klein/" target="_blank">Dan Klein</a> and <a href="https://people.eecs.berkeley.edu/~dawnsong/" target="_blank">Dawn Song</a>, and I have affiliations with <a href="https://bair.berkeley.edu" target="_blank">BAIR</a>, <a href="http://nlp.cs.berkeley.edu/" target="_blank">Berkeley NLP</a>, and <a href="https://security.cs.berkeley.edu/" target="_blank">Berkeley Security</a>.
<br> <br>
I interned at <a href="https://ai.facebook.com/" target="_blank">FAIR</a> in 2021 with <a href="https://robinjia.github.io/" target="_blank">Robin Jia</a> and <a href="https://douwekiela.github.io/" target="_blank">Douwe Kiela</a>, and also at <a href="https://allenai.org/" target="_blank">AI2</a> in 2019 with <a href="https://matt-gardner.github.io/" target="_blank">Matt Gardner</a> and <a href="http://sameersingh.org/" target="_blank">Sameer Singh</a>. I did my undergrad at the University of Maryland, where I worked with <a href="http://www.umiacs.umd.edu/~jbg/" target="_blank">Jordan Boyd-Graber</a>.
<br><br>
<h3 style="margin-bottom:0.75em;">Current Research Interests</h3>
<p><b>Security & Privacy</b> We study vulnerabilities of NLP systems from various adversarial perspectives, including <a href="https://arxiv.org/abs/2004.15015" target="_blank" style="font-size: 0.95em">stealing</a> model weights, <a href="https://arxiv.org/abs/2012.07805" target="_blank" style="font-size: 0.95em">extracting</a> private training data, <a target="_blank" href="https://arxiv.org/abs/2010.12563" style="font-size: 0.95em">poisoning</a> training sets, and <a href="https://arxiv.org/abs/1908.07125" target="_blank" style="font-size: 0.95em">manipulating</a> test predictions. Our current research develops <a target="_blank" href="https://arxiv.org/abs/2010.12563" style="font-size: 0.95em">defenses</a> <a href="https://arxiv.org/abs/2004.15015" target="_blank" style="font-size: 0.95em">against</a> these vulnerabilities.
<p><b>Large Language Models</b> We use large language models for few-shot learning by "prompting" them with training examples. We've shown that few-shot learning can be highly <a href="https://arxiv.org/abs/2102.09690" target="_blank" style="font-size: 0.95em">sensitive</a> to the choice of the prompt, and we've mitigated this sensitivity and improved model accuracy by <a href="https://arxiv.org/abs/2010.15980" target="_blank" style="font-size: 0.95em">automatic</a> prompt design and <a href="https://arxiv.org/abs/2102.09690" target="_blank" style="font-size: 0.95em">calibration</a>. Our current research focuses on making few-shot finetuning <a href="https://arxiv.org/abs/2106.13353" target="_blank" style="font-size: 0.95em">simple and efficient</a>.</p>
<p><b>Robustness & Generalization</b> We analyze the robustness of models to test-time distribution shift. We have shown models are brittle to <a href="https://arxiv.org/abs/2004.06100" target="_blank" style="font-size: 0.95em">natural</a>, <a href="https://arxiv.org/abs/2004.02709" target="_blank" style="font-size: 0.95em">expert</a>-<a href="https://arxiv.org/abs/1809.02701" target="_blank" style="font-size: 0.95em">designed</a>, and <a href="https://arxiv.org/abs/1908.07125" target="_blank" style="font-size: 0.95em">adversarial</a> shifts. We attribute many of these failures to issues in the training data, e.g., spurious correlations in <a href="https://arxiv.org/abs/1908.07125" target="_blank" style="font-size: 0.95em">classification</a> and <a href="https://arxiv.org/abs/1906.02900" target="_blank" style="font-size: 0.95em">question answering</a> datasets. Our recent work develops new methods for <a href="https://arxiv.org/abs/2110.08514" target="_blank" style="font-size: 0.95em">training data collection</a>.
</p>
<br>
</div>
</div>
<div class="posts-wrapper" style="clear:both">
<h3 style="margin-bottom:0.75em;">Publications</h3>
</i>
<p>
<ul class="pubs">
<li>
<a href="https://arxiv.org/abs/2102.09690" target="_blank" style="color:black;font-size:1.0em">
Calibrate Before Use: Improving Few-shot Performance of Language Models</a><br>
Tony Z. Zhao*, Eric Wallace*, Shi Feng, Dan Klein, Sameer Singh<br>
<i>ICML 2021</i><br>
<a href="javascript:unhide('calibration21tldr');">TLDR</a> | <a href="https://twitter.com/Eric_Wallace_/status/1410627135899906060" target="_blank">Twitter</a> <a href="https://twitter.com/arankomatsuzaki/status/1363666486682783744" target="_blank">Discussions</a> | <a href="https://arxiv.org/abs/2102.09690" target="_blank">Paper</a> | <a href="https://github.com/tonyzhaozh/few-shot-learning/" target="_blank">Code</a> | <a href="slides_and_posters/calibration_slides.pdf" target="_blank">Slides</a> | <a href="javascript:unhide('calibration21');">Citation</a>
<div id="calibration21tldr" class="hidden"><b>TLDR:</b> We show that GPT-3's few-shot accuracy has high variance across different choices of the prompt. We propose a calibration procedure that reduces this variance and substantially improves average accuracy.<br></div>
<div id="calibration21" class="hidden">
<pre>@inproceedings{Zhao2021Calibrate,
Title = {Calibrate Before Use: Improving Few-shot Performance of Language Models},
Author = {Tony Z. Zhao and Eric Wallace and Shi Feng and Dan Klein and Sameer Singh},
booktitle={International Conference on Machine Learning},
Year = {2021}}
</pre>
</div>
</li>
<li>
<a href="https://arxiv.org/abs/2012.07805" target="_blank" style="color:black;font-size:1.0em">
Extracting Training Data From Large Language Models</a><br>
Nicholas Carlini, Florian Tramèr, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Úlfar Erlingsson, Alina Oprea, and Colin Raffel<br>
<i>USENIX Security Symposium 2021</i><br>
<a href="javascript:unhide('extracting20tldr');">TLDR</a> | <a href="https://bair.berkeley.edu/blog/2020/12/20/lmmem/" target="_blank">Blog</a> | <a href="https://twitter.com/colinraffel/status/1339012222811598848" target="_blank">Twitter</a> <a href="https://twitter.com/Eric_Wallace_/status/1341221479426400256" target="_blank">Discussions</a> | <a href="https://arxiv.org/abs/2012.07805" target="_blank">Paper</a> | <a href="https://github.com/ftramer/LM_Memorization" target="_blank">Code</a> | <a href="javascript:unhide('extracting20');">Citation</a>
<div id="extracting20tldr" class="hidden"><b>TLDR:</b> We create a black-box method for extracting verbatim training examples from a language model.<br></div>
<div id="extracting20" class="hidden">
<pre>@inproceedings{carlini2020extracting,
title={Extracting Training Data from Large Language Models},
author={Nicholas Carlini and Florian Tram\`er and Eric Wallace and Matthew Jagielski
and Ariel Herbert-Voss and Katherine Lee and Adam Roberts and Tom Brown
and Dawn Song and \'Ulfar Erlingsson and Alina Oprea and Colin Raffel},
booktitle={USENIX Security Symposium},
year={2021}}
</pre>
</div>
</li>
<li>
<a href="https://arxiv.org/abs/2010.12563" target="_blank" style="color:black;font-size:1.0em">
Concealed Data Poisoning Attacks on NLP Models</a><br>
Eric Wallace*, Tony Z. Zhao*, Shi Feng, and Sameer Singh<br>
<i>NAACL 2021</i><br>
<a href="javascript:unhide('poisoning20tldr');">TLDR</a> | <a href="http://ericswallace.com/poisoning" target="_blank">Blog</a> | <a href="https://twitter.com/Eric_Wallace_/status/1319650623705370624" target="_blank">Twitter</a> | <a href="https://arxiv.org/abs/2010.12563" target="_blank">Paper</a> | <a href="https://github.com/Eric-Wallace/data-poisoning" target="_blank">Code</a> | <a href="slides_and_posters/Poisoning-NAACL-June'21.pdf" target="_blank">Slides</a> | <a href="javascript:unhide('poisoning20');">Citation</a>
<div id="poisoning20tldr" class="hidden"><b>TLDR:</b> We develop a new training data poisoning attack that allows an adversary to control model predictions whenever a desired phrase is present in the input.<br></div>
<div id="poisoning20" class="hidden">
<pre>@InProceedings{wallace2021poisoning,
title={Concealed Data Poisoning Attacks on {NLP} Models},
author={Eric Wallace and Tony Z. Zhao and Shi Feng and Sameer Singh},
booktitle={North American Chapter of the Association for Computational Linguistics},
year={2021}}
</pre>
</div>
</li>
<li>
<a href="https://arxiv.org/abs/2010.15980" target="_blank" style="color:black;font-size:1.0em">
AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts</a><br>
Taylor Shin*, Yasaman Razeghi*, Robert L Logan IV*, Eric Wallace, and Sameer Singh<br>
<i>EMNLP 2020</i><br>
<a href="javascript:unhide('auto20tldr');">TLDR</a> | <a href="https://twitter.com/rloganiv/status/1321992351649202177" target="_blank">Twitter</a> | <a href="https://arxiv.org/abs/2010.15980" target="_blank">Paper</a> | <a href="https://github.com/ucinlp/autoprompt" target="_blank">Code</a> | <a href="javascript:unhide('auto20');">Citation</a>
<div id="auto20tldr" class="hidden"><b>TLDR:</b> We propose a method for automatically designing prompts for large language models.<br></div>
<div id="auto20" class="hidden">
<pre>@inproceedings{Shin2020Autoprompt,
Author = {Taylor Shin and Yasaman Razeghi and Robert L. Logan IV and Eric Wallace and Sameer Singh},
BookTitle={Empirical Methods in Natural Language Processing},
Year = {2020},
Title = {AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts}}
</pre>
</div>
</li>
<li>
<a href="https://arxiv.org/abs/2004.15015" target="_blank" style="color:black;font-size:1.0em">
Imitation Attacks and Defenses for Black-box Machine Translation Systems</a><br>
Eric Wallace, Mitchell Stern, and Dawn Song<br>
<i>EMNLP 2020</i><br>
<a href="javascript:unhide('stealing20tldr');">TLDR</a> | <a href="http://ericswallace.com/imitation" target="_blank">Blog</a> | <a href="https://twitter.com/Eric_Wallace_/status/1256227702056595456" target="_blank">Twitter</a> | <a href="https://arxiv.org/abs/2004.15015" target="_blank">Paper</a> | <a href="slides_and_posters/stealing_slides.pdf" target="_blank">Slides</a> | <a href="https://github.com/Eric-Wallace/adversarial-mt" target="_blank">Code</a> | <a href="javascript:unhide('stealing20');">Citation</a>
<div id="stealing20tldr" class="hidden"><b>TLDR:</b> We "steal" production NLP systems by training models to imitate their outputs. We then use the imitation models to attack the black-box production systems. We finally propose a defense that mitigates these vulnerabilities.<br></div>
<div id="stealing20" class="hidden">
<pre>@inproceedings{Wallace2020Stealing,
Author = {Eric Wallace and Mitchell Stern and Dawn Song},
BookTitle={Empirical Methods in Natural Language Processing},
Year = {2020},
Title = {Imitation Attacks and Defenses for Black-box Machine Translation Systems}}
</pre>
</div>
</li>
<li>
<a href="https://arxiv.org/abs/2002.11794" target="_blank" style="color:black;font-size:1.0em">
Train Large, Then Compress: Rethinking Model Size for Efficient Training and Inference of Transformers</a><br>
Zhuohan Li*, Eric Wallace*, Sheng Shen*, Kevin Lin*, Kurt Keutzer, Dan Klein, and Joseph E. Gonzalez<br>
<i>ICML 2020</i><br>
<a href="javascript:unhide('efficient20tldr');">TLDR</a> | <a href="https://bair.berkeley.edu/blog/2020/03/05/compress/" target="_blank">Blog</a> | <a href="https://twitter.com/Eric_Wallace_/status/1235616760595791872" target="_blank">Twitter</a> | <a href="https://arxiv.org/abs/2002.11794" target="_blank">Paper</a> | <a href="slides_and_posters/train_large.pdf" target="_blank">Slides</a> | <a href="javascript:unhide('efficient20');">Citation</a>
<div id="efficient20tldr" class="hidden"><b>TLDR:</b> We show that <i>increasing</i> model size actually speeds up training and inference for Transformer models. The key idea is to use a very large model but perform very few epochs and apply heavy compression.<br></div>
<div id="efficient20" class="hidden">
<pre>@inproceedings{Li2020Efficient,
Author = {Zhuohan Li and Eric Wallace and Sheng Shen and Kevin Lin and Kurt Keutzer and Dan Klein and Joseph E. Gonzalez},
Booktitle = {International Conference on Machine Learning},
Year = {2020},
Title = {Train Large, Then Compress: Rethinking Model Size for Efficient Training and Inference of Transformers}}
</pre>
</div>
</li>
<li>
<a href="https://arxiv.org/abs/2004.06100" target="_blank" style="color:black;font-size:1.0em">Pretrained Transformers Improve Out-of-Distribution Robustness</a><br>
Dan Hendrycks*, Xiaoyuan Liu*, Eric Wallace, Adam Dziedzic, Rishabh Krishnan, and Dawn Song<br>
<i>ACL 2020</i><br>
<a href="javascript:unhide('robust20tldr');">TLDR</a> | <a href="https://arxiv.org/abs/2004.06100" target="_blank">Paper</a> | <a href="https://twitter.com/Eric_Wallace_/status/1250507707674578944" target="_blank">Twitter</a> | <a href="https://github.com/camelop/NLP-Robustness" target="_blank">Code</a> | <a href="slides_and_posters/ood_robustness.pdf" target="_blank">Slides</a> | <a href="javascript:unhide('robust20');">Citation</a>
<div id="robust20tldr" class="hidden"><b>TLDR:</b> How does pretraining affect <i>out-of-distribution</i> robustness? We create an OOD benchmark and use it to show that pretraining substantially improves OOD accuracy and detection rates.<br></div>
<div id="robust20" class="hidden">
<pre>@inproceedings{hendrycks2020pretrained,
Author = {Dan Hendrycks and Xiaoyuan Liu and Eric Wallace and Adam Dziedzic and Rishabh Krishnan and Dawn Song},
Booktitle = {Association for Computational Linguistics},
Year = {2020},
Title = {Pretrained Transformers Improve Out-of-Distribution Robustness}}
</pre>
</div>
</li>
<li>
<a href="https://arxiv.org/abs/1908.07125" target="_blank" style="color:black;font-size:1.0em">Universal Adversarial Triggers for Attacking and Analyzing NLP</a><br>
Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh<br>
<i>EMNLP 2019</i><br>
<a href="javascript:unhide('triggers19tldr');">TLDR</a> | <a href="https://vimeo.com/396789889" target="_blank">Video</a> | <a href="http://ericswallace.com/triggers" target="_blank">Blog</a> | <a href="https://twitter.com/Eric_Wallace_/status/1168907518623571974" target="_blank">Twitter</a> | <a href="https://arxiv.org/abs/1908.07125" target="_blank">Paper</a> | <a href="https://github.com/Eric-Wallace/universal-triggers" target="_blank">Code</a> | <a href="slides_and_posters/Universal_Adversarial_Triggers.pdf" target="_blank">Slides</a> | <a href="javascript:unhide('triggers19');">Citation</a>
<div id="triggers19tldr" class="hidden"><b>TLDR:</b> We create phrases that cause a model to produce a specific prediction when concatenated to <i>any</i> input. Triggers reveal egregious and insightful errors for text classification, reading comprehension, and text generation.<br> </div>
<div id="triggers19" class="hidden">
<pre>@inproceedings{Wallace2019Triggers,
Author = {Eric Wallace and Shi Feng and Nikhil Kandpal and Matt Gardner and Sameer Singh},
Booktitle = {Empirical Methods in Natural Language Processing},
Year = {2019},
Title = {Universal Adversarial Triggers for Attacking and Analyzing {NLP}}}
</pre>
</div>
</li>
<li>
<a href="https://arxiv.org/abs/1909.07940" target="_blank" style="color:black;font-size:1.0em">Do NLP Models Know Numbers? Probing Numeracy in Embeddings</a><br>
Eric Wallace*, Yizhong Wang*, Sujian Li, Sameer Singh, and Matt Gardner<br>
<i>EMNLP 2019</i><br>
<a href="javascript:unhide('numeracy19tldr');">TLDR</a> | <a href="https://twitter.com/Eric_Wallace_/status/1174360279624192000" target="_blank">Twitter</a> | <a href="https://arxiv.org/abs/1909.07940" target="_blank">Paper</a> | <a href="https://github.com/Eric-Wallace/numeracy" target="_blank">Code</a> | <a href="slides_and_posters/NumeracyPoster.pdf" target="_blank">Poster</a> | <a href="javascript:unhide('numeracy19');">Citation</a>
<div id="numeracy19tldr" class="hidden"><b>TLDR:</b> We show that pre-trained word embeddings (e.g., BERT, word2vec, ELMo, GloVe) capture number magnitude and order, e.g., they know that "74" is smaller than "eighty-two". This facilitates basic numerical reasoning tasks. <br></div>
<div id="numeracy19" class="hidden">
<pre>@inproceedings{Wallace2019Numeracy,
Author = {Eric Wallace and Yizhong Wang and Sujian Li and Sameer Singh and Matt Gardner},
Booktitle = {Empirical Methods in Natural Language Processing},
Year = {2019},
Title = {Do {NLP} Models Know Numbers? Probing Numeracy in Embeddings}}
</pre>
</div>
</li>
<li>
<a href="https://arxiv.org/abs/1909.09251" target="_blank" style="color:black;font-size:1.0em">AllenNLP Interpret: A Framework for Explaining Predictions of NLP Models</a><br>
Eric Wallace, Jens Tuyls, Junlin Wang, Sanjay Subramanian, Matt Gardner, and Sameer Singh<br>
<i>Demo at EMNLP 2019</i> <b><i>Best Demo Award</i></b><br>
<a href="javascript:unhide('interpret19tldr');">TLDR</a> | <a href="https://allennlp.org/interpret" target="_blank">Landing Page</a> | <a href="https://twitter.com/Eric_Wallace_/status/1176886627852898309" target="_blank">Twitter</a> | <a href="https://demo.allennlp.org/reading-comprehension" target="_blank">Demo</a> | <a href="https://arxiv.org/abs/1909.09251" target="_blank">Paper</a> | <a href="slides_and_posters/InterpretPoster.pdf" target="_blank">Poster</a> | <a href="javascript:unhide('interpret19');">Citation</a>
<div id="interpret19tldr" class="hidden"><b>TLDR:</b> An open-source toolkit built on top of AllenNLP that makes it easy to interpret NLP models.<br> </div>
<div id="interpret19" class="hidden">
<pre>@inproceedings{Wallace2019AllenNLP,
Author = {Eric Wallace and Jens Tuyls and Junlin Wang and Sanjay Subramanian and Matt Gardner and Sameer Singh},
Booktitle = {Empirical Methods in Natural Language Processing},
Year = {2019},
Title = {{AllenNLP Interpret}: A Framework for Explaining Predictions of {NLP} Models}}
</pre>
</div>
</li>
<li>
<a href="http://arxiv.org/abs/1906.02900" target="_blank" style="color:black;font-size:1.0em">Compositional Questions Do Not Necessitate Multi-hop Reasoning</a><br>
Sewon Min*, Eric Wallace*, Sameer Singh, Matt Gardner, Hannaneh Hajishirzi, and Luke Zettlemoyer<br>
<i>ACL 2019</i><br>
<a href="javascript:unhide('multihop19tldr');">TLDR</a> | <a href="https://arxiv.org/abs/1906.02900" target="_blank">Paper</a> | <a href="slides_and_posters/Compositional_Slides.pdf" target="_blank">Slides</a> | <a href="https://github.com/shmsw25/single-hop-rc" target="_blank">Code</a> | <a href="javascript:unhide('multihop19');">Citation</a>
<div id="multihop19tldr" class="hidden"><b>TLDR:</b> We argue that constructing multi-hop QA datasets is non-trivial, and that existing datasets are simpler than expected. For instance, single-hop models can solve most of HotpotQA due to weak distractor paragraphs.<br></div>
<div id="multihop19" class="hidden">
<pre>@inproceedings{Min2019Multihop,
Author = {Sewon Min and Eric Wallace and Sameer Singh and Matt Gardner and Hannaneh Hajishirzi and Luke Zettlemoyer},
Booktitle = {Association for Computational Linguistics},
Year = {2019},
Title = {Compositional Questions Do Not Necessitate Multi-hop Reasoning}}
</pre>
</div>
</li>
<li>
<a href="https://arxiv.org/abs/1809.02701" target="_blank" style="color:black;font-size:1.0em">Trick Me If You Can: Human-in-the-loop Generation of Adversarial Examples for Question Answering</a><br>
Eric Wallace, Pedro Rodriguez, Shi Feng, Ikuya Yamada, and Jordan Boyd-Graber<br>
<i>TACL 2019</i><br>
<a href="javascript:unhide('trick19tldr');">TLDR</a> | <a href="https://arxiv.org/abs/1809.02701" target="_blank">Paper</a> | <a href="https://github.com/Eric-Wallace/trickme-interface" target="_blank">Code</a> | <a href="slides_and_posters/TrickMe_Poster.pdf" target="_blank">Poster</a> | <a href="javascript:unhide('trick19');">Citation</a>
<div id="trick19tldr" class="hidden"><b>TLDR:</b> We use a human-in-the-loop approach for generating adversarial examples in NLP. We display model intepretations and predictions in a UI, which enables collaborative + interactive attacks on question answering systems .<br></div>
<div id="trick19" class="hidden">
<pre>@inproceedings{Wallace2019Trick,
Author = {Eric Wallace and Pedro Rodriguez and Shi Feng and Ikuya Yamada and Jordan Boyd-Graber},
Booktitle = {Transactions of the Association for Computational Linguistics},
Year = {2019},
Title = {Trick Me If You Can: Human-in-the-loop Generation of Adversarial Examples for Question Answering}}
</pre>
</div>
</li>
<li>
<a href="https://arxiv.org/abs/1804.07781" target="_blank" style="color:black;font-size:1.0em">Pathologies of Neural Models Make Interpretations Difficult</a><br>
Shi Feng, Eric Wallace, Alvin Grissom II, Mohit Iyyer, Pedro Rodriguez, Jordan Boyd-Graber<br>
<i>EMNLP 2018</i><br>
<a href="javascript:unhide('pathological18tldr');">TLDR</a> | <a href="https://vimeo.com/306158589" target="_blank">Video</a> | <a href="https://arxiv.org/abs/1804.07781" target="_blank">Paper</a> |
<a href="slides_and_posters/pathologies_slides.pdf" target="_blank">Slides</a> | <a href="https://github.com/allenai/allennlp/blob/master/allennlp/interpret/attackers/input_reduction.py" target="_blank">Code</a> | <a href="javascript:unhide('pathological18');">Citation</a>
<div id="pathological18tldr" class="hidden"><b>TLDR:</b> Saliency maps are a popular interpretation technique. We show that certain pathological behavior present in neural models (namely prediction overconfidence) can negatively impact these interpretations.<br> </div>
<div id="pathological18" class="hidden">
<pre>@inproceedings{Feng2018Pathological,
Author = {Shi Feng and Eric Wallace and Alvin Grissom II and Mohit Iyyer and Pedro Rodriguez and Jordan Boyd-Graber},
Booktitle = {Empirical Methods in Natural Language Processing},
Year = {2018},
Title = {Pathologies of Neural Models Make Interpretations Difficult}}
</pre>
</div>
</li>
</ul>
</p>
<!--<p>A recent presentation of my work! Given live at EMNLP 2019.<div class="box"><iframe src="https://player.vimeo.com/video/396789889?byline=0&portrait=0" width="100%" frameborder="0" allow="autoplay; fullscreen" allowfullscreen align="left"></iframe></div></p>
<br><br><br><br><br><br><br>-->
</div>
</div>
</body>
</html>