arxiv.org.rss.091.xml - sfeed_tests - sfeed tests and RSS and Atom files
HTML git clone git://git.codemadness.org/sfeed_tests
DIR Log
DIR Files
DIR Refs
DIR README
DIR LICENSE
---
arxiv.org.rss.091.xml (806587B)
---
1 <?xml version="1.0" encoding="UTF-8"?>
2
3 <!DOCTYPE rss PUBLIC "-//Netscape Communications//DTD RSS 0.91//EN"
4 "http://www.rssboard.org/rss-0.91.dtd">
5
6 <rss version="0.91">
7
8 <channel>
9 <title>cs updates on arXiv.org</title>
10 <link>http://fr.arxiv.org/</link>
11 <description>Computer Science (cs) updates on the arXiv.org e-print archive</description>
12 <language>en-us</language>
13 <pubDate>Fri, 30 Oct 2020 00:30:00 GMT</pubDate>
14 <lastBuildDate>Fri, 30 Oct 2020 00:30:00 GMT</lastBuildDate>
15 <managingEditor>www-admin@arxiv.org</managingEditor>
16
17 <image>
18 <title>arXiv.org</title>
19 <url>http://fr.arxiv.org/icons/sfx.gif</url>
20 <link>http://fr.arxiv.org/</link>
21 </image>
22 <item>
23 <title>Raw Audio for Depression Detection Can Be More Robust Against Gender Imbalance than Mel-Spectrogram Features. (arXiv:2010.15120v1 [cs.SD])</title>
24 <link>http://fr.arxiv.org/abs/2010.15120</link>
25 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Bailey_A/0/1/0/all/0/1">Andrew Bailey</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Plumbley_M/0/1/0/all/0/1">Mark D. Plumbley</a></p>
26
27 <p>Depression is a large-scale mental health problem and a challenging area for
28 machine learning researchers in terms of the detection of depression. Datasets
29 such as the Distress Analysis Interview Corpus - Wizard of Oz have been created
30 to aid research in this area. However, on top of the challenges inherent in
31 accurately detecting depression, biases in datasets may result in skewed
32 classification performance. In this paper we examine gender bias in the
33 DAIC-WOZ dataset using audio-based deep neural networks. We show that gender
34 biases in DAIC-WOZ can lead to an overreporting of performance, which has been
35 overlooked in the past due to the same gender biases being present in the test
36 set. By using raw audio and different concepts from Fair Machine Learning, such
37 as data re-distribution, we can mitigate against the harmful effects of bias.
38 </p>
39 </description>
40 </item>
41 <item>
42 <title>papaya2: 2D Irreducible Minkowski Tensor computation. (arXiv:2010.15138v1 [cs.GR])</title>
43 <link>http://fr.arxiv.org/abs/2010.15138</link>
44 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Schaller_F/0/1/0/all/0/1">Fabian M. Schaller</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Wagner_J/0/1/0/all/0/1">Jenny Wagner</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Kapfer_S/0/1/0/all/0/1">Sebastian C. Kapfer</a></p>
45
46 <p>A common challenge in scientific and technical domains is the quantitative
47 description of geometries and shapes, e.g. in the analysis of microscope
48 imagery or astronomical observation data. Frequently, it is desirable to go
49 beyond scalar shape metrics such as porosity and surface to volume ratios
50 because the samples are anisotropic or because direction-dependent quantities
51 such as conductances or elasticity are of interest. Minkowski Tensors are a
52 systematic family of versatile and robust higher-order shape descriptors that
53 allow for shape characterization of arbitrary order and promise a path to
54 systematic structure-function relationships for direction-dependent properties.
55 Papaya2 is a software to calculate 2D higher-order shape metrics with a library
56 interface, support for Irreducible Minkowski Tensors and interpolated marching
57 squares. Extensions to Matlab, JavaScript and Python are provided as well.
58 While the tensor of inertia is computed by many tools, we are not aware of
59 other open-source software which provides higher-rank shape characterization in
60 2D.
61 </p>
62 </description>
63 </item>
64 <item>
65 <title>DeSMOG: Detecting Stance in Media On Global Warming. (arXiv:2010.15149v1 [cs.CL])</title>
66 <link>http://fr.arxiv.org/abs/2010.15149</link>
67 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Luo_Y/0/1/0/all/0/1">Yiwei Luo</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Card_D/0/1/0/all/0/1">Dallas Card</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Jurafsky_D/0/1/0/all/0/1">Dan Jurafsky</a></p>
68
69 <p>Citing opinions is a powerful yet understudied strategy in argumentation. For
70 example, an environmental activist might say, "Leading scientists agree that
71 global warming is a serious concern," framing a clause which affirms their own
72 stance ("that global warming is serious") as an opinion endorsed ("[scientists]
73 agree") by a reputable source ("leading"). In contrast, a global warming denier
74 might frame the same clause as the opinion of an untrustworthy source with a
75 predicate connoting doubt: "Mistaken scientists claim [...]." Our work studies
76 opinion-framing in the global warming (GW) debate, an increasingly partisan
77 issue that has received little attention in NLP. We introduce DeSMOG, a dataset
78 of stance-labeled GW sentences, and train a BERT classifier to study novel
79 aspects of argumentation in how different sides of a debate represent their own
80 and each other's opinions. From 56K news articles, we find that similar
81 linguistic devices for self-affirming and opponent-doubting discourse are used
82 across GW-accepting and skeptic media, though GW-skeptical media shows more
83 opponent-doubt. We also find that authors often characterize sources as
84 hypocritical, by ascribing opinions expressing the author's own view to source
85 entities known to publicly endorse the opposing view. We release our stance
86 dataset, model, and lexicons of framing devices for future work on
87 opinion-framing and the automatic detection of GW stance.
88 </p>
89 </description>
90 </item>
91 <item>
92 <title>On the Optimality and Convergence Properties of the Learning Model Predictive Controller. (arXiv:2010.15153v1 [math.OC])</title>
93 <link>http://fr.arxiv.org/abs/2010.15153</link>
94 <description><p>Authors: <a href="http://fr.arxiv.org/find/math/1/au:+Rosolia_U/0/1/0/all/0/1">Ugo Rosolia</a>, <a href="http://fr.arxiv.org/find/math/1/au:+Lian_Y/0/1/0/all/0/1">Yingzhao Lian</a>, <a href="http://fr.arxiv.org/find/math/1/au:+Maddalena_E/0/1/0/all/0/1">Emilio T. Maddalena</a>, <a href="http://fr.arxiv.org/find/math/1/au:+Ferrari_Trecate_G/0/1/0/all/0/1">Giancarlo Ferrari-Trecate</a>, <a href="http://fr.arxiv.org/find/math/1/au:+Jones_C/0/1/0/all/0/1">Colin N. Jones</a></p>
95
96 <p>In this technical note we analyse the performance improvement and optimality
97 properties of the Learning Model Predictive Control (LMPC) strategy for linear
98 deterministic systems. The LMPC framework is a policy iteration scheme where
99 closed-loop trajectories are used to update the control policy for the next
100 execution of the control task. We show that, when a Linear Independence
101 Constraint Qualification (LICQ) condition holds, the LMPC scheme guarantees
102 strict iterative performance improvement and optimality, meaning that the
103 closed-loop cost evaluated over the entire task converges asymptotically to the
104 optimal cost of the infinite-horizon control problem. Compared to previous
105 works this sufficient LICQ condition can be easily checked, it holds for a
106 larger class of systems and it can be used to adaptively select the prediction
107 horizon of the controller, as demonstrated by a numerical example.
108 </p>
109 </description>
110 </item>
111 <item>
112 <title>Kernel Aggregated Fast Multipole Method: Efficient summation of Laplace and Stokes kernel functions. (arXiv:2010.15155v1 [math.NA])</title>
113 <link>http://fr.arxiv.org/abs/2010.15155</link>
114 <description><p>Authors: <a href="http://fr.arxiv.org/find/math/1/au:+Yan_W/0/1/0/all/0/1">Wen Yan</a>, <a href="http://fr.arxiv.org/find/math/1/au:+Blackwell_R/0/1/0/all/0/1">Robert Blackwell</a></p>
115
116 <p>Many different simulation methods for Stokes flow problems involve a common
117 computationally intense task---the summation of a kernel function over $O(N^2)$
118 pairs of points. One popular technique is the Kernel Independent Fast Multipole
119 Method (KIFMM), which constructs a spatial adaptive octree and places a small
120 number of equivalent multipole and local points around each octree box, and
121 completes the kernel sum with $O(N)$ performance. However, the KIFMM cannot be
122 used directly with nonlinear kernels, can be inefficient for complicated linear
123 kernels, and in general is difficult to implement compared to less-efficient
124 alternatives such as Ewald-type methods. Here we present the Kernel Aggregated
125 Fast Multipole Method (KAFMM), which overcomes these drawbacks by allowing
126 different kernel functions to be used for specific stages of octree traversal.
127 In many cases a simpler linear kernel suffices during the most extensive stage
128 of octree traversal, even for nonlinear kernel summation problems. The KAFMM
129 thereby improves computational efficiency in general and also allows efficient
130 evaluation of some nonlinear kernel functions such as the regularized
131 Stokeslet. We have implemented our method as an open-source software library
132 STKFMM with support for Laplace kernels, the Stokeslet, regularized Stokeslet,
133 Rotne-Prager-Yamakawa (RPY) tensor, and the Stokes double-layer and traction
134 operators. Open and periodic boundary conditions are supported for all kernels,
135 and the no-slip wall boundary condition is supported for the Stokeslet and RPY
136 tensor. The package is designed to be ready-to-use as well as being readily
137 extensible to additional kernels. Massive parallelism is supported with mixed
138 OpenMP and MPI.
139 </p>
140 </description>
141 </item>
142 <item>
143 <title>Diagnostic data integration using deep neural networks for real-time plasma analysis. (arXiv:2010.15156v1 [physics.comp-ph])</title>
144 <link>http://fr.arxiv.org/abs/2010.15156</link>
145 <description><p>Authors: <a href="http://fr.arxiv.org/find/physics/1/au:+Garola_A/0/1/0/all/0/1">A. Rigoni Garola</a>, <a href="http://fr.arxiv.org/find/physics/1/au:+Cavazzana_R/0/1/0/all/0/1">R. Cavazzana</a>, <a href="http://fr.arxiv.org/find/physics/1/au:+Gobbin_M/0/1/0/all/0/1">M. Gobbin</a>, <a href="http://fr.arxiv.org/find/physics/1/au:+Delogu_R/0/1/0/all/0/1">R.S. Delogu</a>, <a href="http://fr.arxiv.org/find/physics/1/au:+Manduchi_G/0/1/0/all/0/1">G. Manduchi</a>, <a href="http://fr.arxiv.org/find/physics/1/au:+Taliercio_C/0/1/0/all/0/1">C. Taliercio</a>, <a href="http://fr.arxiv.org/find/physics/1/au:+Luchetta_A/0/1/0/all/0/1">A. Luchetta</a></p>
146
147 <p>Recent advances in acquisition equipment is providing experiments with
148 growing amounts of precise yet affordable sensors. At the same time an improved
149 computational power, coming from new hardware resources (GPU, FPGA, ACAP), has
150 been made available at relatively low costs. This led us to explore the
151 possibility of completely renewing the chain of acquisition for a fusion
152 experiment, where many high-rate sources of data, coming from different
153 diagnostics, can be combined in a wide framework of algorithms. If on one hand
154 adding new data sources with different diagnostics enriches our knowledge about
155 physical aspects, on the other hand the dimensions of the overall model grow,
156 making relations among variables more and more opaque. A new approach for the
157 integration of such heterogeneous diagnostics, based on composition of deep
158 \textit{variational autoencoders}, could ease this problem, acting as a
159 structural sparse regularizer. This has been applied to RFX-mod experiment
160 data, integrating the soft X-ray linear images of plasma temperature with the
161 magnetic state.
162 </p>
163 <p>However to ensure a real-time signal analysis, those algorithmic techniques
164 must be adapted to run in well suited hardware. In particular it is shown that,
165 attempting a quantization of neurons transfer functions, such models can be
166 modified to create an embedded firmware. This firmware, approximating the deep
167 inference model to a set of simple operations, fits well with the simple logic
168 units that are largely abundant in FPGAs. This is the key factor that permits
169 the use of affordable hardware with complex deep neural topology and operates
170 them in real-time.
171 </p>
172 </description>
173 </item>
174 <item>
175 <title>Panoster: End-to-end Panoptic Segmentation of LiDAR Point Clouds. (arXiv:2010.15157v1 [cs.CV])</title>
176 <link>http://fr.arxiv.org/abs/2010.15157</link>
177 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Gasperini_S/0/1/0/all/0/1">Stefano Gasperini</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Mahani_M/0/1/0/all/0/1">Mohammad-Ali Nikouei Mahani</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Marcos_Ramiro_A/0/1/0/all/0/1">Alvaro Marcos-Ramiro</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Navab_N/0/1/0/all/0/1">Nassir Navab</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Tombari_F/0/1/0/all/0/1">Federico Tombari</a></p>
178
179 <p>Panoptic segmentation has recently unified semantic and instance
180 segmentation, previously addressed separately, thus taking a step further
181 towards creating more comprehensive and efficient perception systems. In this
182 paper, we present Panoster, a novel proposal-free panoptic segmentation method
183 for point clouds. Unlike previous approaches relying on several steps to group
184 pixels or points into objects, Panoster proposes a simplified framework
185 incorporating a learning-based clustering solution to identify instances. At
186 inference time, this acts as a class-agnostic semantic segmentation, allowing
187 Panoster to be fast, while outperforming prior methods in terms of accuracy.
188 Additionally, we showcase how our approach can be flexibly and effectively
189 applied on diverse existing semantic architectures to deliver panoptic
190 predictions.
191 </p>
192 </description>
193 </item>
194 <item>
195 <title>CNN Profiler on Polar Coordinate Images for Tropical Cyclone Structure Analysis. (arXiv:2010.15158v1 [cs.CV])</title>
196 <link>http://fr.arxiv.org/abs/2010.15158</link>
197 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Chen_B/0/1/0/all/0/1">Boyo Chen</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Chen_B/0/1/0/all/0/1">Buo-Fu Chen</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Hsiao_C/0/1/0/all/0/1">Chun-Min Hsiao</a></p>
198
199 <p>Convolutional neural networks (CNN) have achieved great success in analyzing
200 tropical cyclones (TC) with satellite images in several tasks, such as TC
201 intensity estimation. In contrast, TC structure, which is conventionally
202 described by a few parameters estimated subjectively by meteorology
203 specialists, is still hard to be profiled objectively and routinely. This study
204 applies CNN on satellite images to create the entire TC structure profiles,
205 covering all the structural parameters. By utilizing the meteorological domain
206 knowledge to construct TC wind profiles based on historical structure
207 parameters, we provide valuable labels for training in our newly released
208 benchmark dataset. With such a dataset, we hope to attract more attention to
209 this crucial issue among data scientists. Meanwhile, a baseline is established
210 with a specialized convolutional model operating on polar-coordinates. We
211 discovered that it is more feasible and physically reasonable to extract
212 structural information on polar-coordinates, instead of Cartesian coordinates,
213 according to a TC's rotational and spiral natures. Experimental results on the
214 released benchmark dataset verified the robustness of the proposed model and
215 demonstrated the potential for applying deep learning techniques for this
216 barely developed yet important topic.
217 </p>
218 </description>
219 </item>
220 <item>
221 <title>Sizeless: Predicting the optimal size of serverless functions. (arXiv:2010.15162v1 [cs.DC])</title>
222 <link>http://fr.arxiv.org/abs/2010.15162</link>
223 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Eismann_S/0/1/0/all/0/1">Simon Eismann</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Bui_L/0/1/0/all/0/1">Long Bui</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Grohmann_J/0/1/0/all/0/1">Johannes Grohmann</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Abad_C/0/1/0/all/0/1">Cristina L. Abad</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Herbst_N/0/1/0/all/0/1">Nikolas Herbst</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Kounev_S/0/1/0/all/0/1">Samuel Kounev</a></p>
224
225 <p>Serverless functions are a cloud computing paradigm that reduces operational
226 overheads for developers, because the cloud provider takes care of resource
227 management tasks such as resource provisioning, deployment, and auto-scaling.
228 The only resource management task that developers are still in charge of is
229 resource sizing, that is, selecting how much resources are allocated to each
230 worker instance. However, due to the challenging nature of resource sizing,
231 developers often neglect it despite its significant cost and performance
232 benefits. Existing approaches aiming to automate serverless functions resource
233 sizing require dedicated performance tests, which are time consuming to
234 implement and maintain.
235 </p>
236 <p>In this paper, we introduce Sizeless -- an approach to predict the optimal
237 resource size of a serverless function using monitoring data from a single
238 resource size. As our approach requires only production monitoring data,
239 developers no longer need to implement and maintain representative performance
240 tests. Furthermore, it enables cloud providers, which cannot engage in testing
241 the performance of user functions, to implement resource sizing on a platform
242 level and automate the last resource management task associated with serverless
243 functions. In our evaluation, Sizeless was able to predict the execution time
244 of the serverless functions of a realistic server-less application with a
245 median prediction accuracy of 93.1%. Using Sizeless to optimize the memory size
246 of this application results in a speedup of 16.7% while simultaneously
247 decreasing costs by 2.5%.
248 </p>
249 </description>
250 </item>
251 <item>
252 <title>Polymer Informatics with Multi-Task Learning. (arXiv:2010.15166v1 [cond-mat.mtrl-sci])</title>
253 <link>http://fr.arxiv.org/abs/2010.15166</link>
254 <description><p>Authors: <a href="http://fr.arxiv.org/find/cond-mat/1/au:+Kunneth_C/0/1/0/all/0/1">Christopher K&#xfc;nneth</a>, <a href="http://fr.arxiv.org/find/cond-mat/1/au:+Rajan_A/0/1/0/all/0/1">Arunkumar Chitteth Rajan</a>, <a href="http://fr.arxiv.org/find/cond-mat/1/au:+Tran_H/0/1/0/all/0/1">Huan Tran</a>, <a href="http://fr.arxiv.org/find/cond-mat/1/au:+Chen_L/0/1/0/all/0/1">Lihua Chen</a>, <a href="http://fr.arxiv.org/find/cond-mat/1/au:+Kim_C/0/1/0/all/0/1">Chiho Kim</a>, <a href="http://fr.arxiv.org/find/cond-mat/1/au:+Ramprasad_R/0/1/0/all/0/1">Rampi Ramprasad</a></p>
255
256 <p>Modern data-driven tools are transforming application-specific polymer
257 development cycles. Surrogate models that can be trained to predict the
258 properties of new polymers are becoming commonplace. Nevertheless, these models
259 do not utilize the full breadth of the knowledge available in datasets, which
260 are oftentimes sparse; inherent correlations between different property
261 datasets are disregarded. Here, we demonstrate the potency of multi-task
262 learning approaches that exploit such inherent correlations effectively,
263 particularly when some property dataset sizes are small. Data pertaining to 36
264 different properties of over $13, 000$ polymers (corresponding to over $23,000$
265 data points) are coalesced and supplied to deep-learning multi-task
266 architectures. Compared to conventional single-task learning models (that are
267 trained on individual property datasets independently), the multi-task approach
268 is accurate, efficient, scalable, and amenable to transfer learning as more
269 data on the same or different properties become available. Moreover, these
270 models are interpretable. Chemical rules, that explain how certain features
271 control trends in specific property values, emerge from the present work,
272 paving the way for the rational design of application specific polymers meeting
273 desired property or performance objectives.
274 </p>
275 </description>
276 </item>
277 <item>
278 <title>Semi-Grant-Free NOMA: Ergodic Rates Analysis with Random Deployed Users. (arXiv:2010.15169v1 [cs.IT])</title>
279 <link>http://fr.arxiv.org/abs/2010.15169</link>
280 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Zhang_C/0/1/0/all/0/1">Chao Zhang</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Liu_Y/0/1/0/all/0/1">Yuanwei Liu</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Yi_W/0/1/0/all/0/1">Wenqiang Yi</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Qin_Z/0/1/0/all/0/1">Zhijin Qin</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Ding_Z/0/1/0/all/0/1">Zhiguo Ding</a></p>
281
282 <p>Semi-grant-free (Semi-GF) non-orthogonal multiple access (NOMA) enables
283 grant-free (GF) and grant-based (GB) users to share the same resource blocks,
284 thereby balancing the connectivity and stability of communications. This letter
285 analyzes ergodic rates of Semi-GF NOMA systems. First, this paper exploits a
286 Semi-GF protocol, denoted as dynamic protocol, for selecting GF users into the
287 occupied GB channels via the GB user's instantaneous received power. Under this
288 protocol, the closed-form analytical and approximated expressions for ergodic
289 rates are derived. The numerical results illustrate that the GF user (weak NOMA
290 user) has a performance upper limit, while the ergodic rate of the GB user
291 (strong NOMA user) increases linearly versus the transmit signal-to-noise
292 ratio.
293 </p>
294 </description>
295 </item>
296 <item>
297 <title>Slicing a single wireless collision channel among throughput- and timeliness-sensitive services. (arXiv:2010.15171v1 [cs.IT])</title>
298 <link>http://fr.arxiv.org/abs/2010.15171</link>
299 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Leyva_Mayorga_I/0/1/0/all/0/1">Israel Leyva-Mayorga</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Chiariotti_F/0/1/0/all/0/1">Federico Chiariotti</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Stefanovic_C/0/1/0/all/0/1">&#x10c;edomir Stefanovi&#x107;</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Kalor_A/0/1/0/all/0/1">Anders E. Kal&#xf8;r</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Popovski_P/0/1/0/all/0/1">Petar Popovski</a></p>
300
301 <p>The fifth generation (5G) wireless system has a platform-driven approach,
302 aiming to support heterogeneous connections with very diverse requirements. The
303 shared wireless resources should be sliced in a way that each user perceives
304 that its requirement has been met. Heterogeneity challenges the traditional
305 notion of resource efficiency, as the resource usage has cater for, e.g. rate
306 maximization for one user and timeliness requirement for another user. This
307 paper treats a model for radio access network (RAN) uplink, where a
308 throughput-demanding broadband user shares wireless resources with an
309 intermittently active user that wants to optimize the timeliness, expressed in
310 terms of latency-reliability or Age of Information (AoI). We evaluate the
311 trade-offs between throughput and timeliness for Orthogonal Multiple Access
312 (OMA) as well as Non-Orthogonal Multiple Access (NOMA) with successive
313 interference cancellation (SIC). We observe that NOMA with SIC, in a
314 conservative scenario with destructive collisions, is just slightly inferior to
315 that of OMA, which indicates that it may offer significant benefits in
316 practical deployments where the capture effect is frequently encountered. On
317 the other hand, finding the optimal configuration of NOMA with SIC depends on
318 the activity pattern of the intermittent user, to which OMA is insensitive.
319 </p>
320 </description>
321 </item>
322 <item>
323 <title>Improving Perceptual Quality by Phone-Fortified Perceptual Loss for Speech Enhancement. (arXiv:2010.15174v1 [cs.SD])</title>
324 <link>http://fr.arxiv.org/abs/2010.15174</link>
325 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Hsieh_T/0/1/0/all/0/1">Tsun-An Hsieh</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Yu_C/0/1/0/all/0/1">Cheng Yu</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Fu_S/0/1/0/all/0/1">Szu-Wei Fu</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Lu_X/0/1/0/all/0/1">Xugang Lu</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Tsao_Y/0/1/0/all/0/1">Yu Tsao</a></p>
326
327 <p>Speech enhancement (SE) aims to improve speech quality and intelligibility,
328 which are both related to a smooth transition in speech segments that may carry
329 linguistic information, e.g. phones and syllables. In this study, we took
330 phonetic characteristics into account in the SE training process. Hence, we
331 designed a phone-fortified perceptual (PFP) loss, and the training of our SE
332 model was guided by PFP loss. In PFP loss, phonetic characteristics are
333 extracted by wav2vec, an unsupervised learning model based on the contrastive
334 predictive coding (CPC) criterion. Different from previous deep-feature-based
335 approaches, the proposed approach explicitly uses the phonetic information in
336 the deep feature extraction process to guide the SE model training. To test the
337 proposed approach, we first confirmed that the wav2vec representations carried
338 clear phonetic information using a t-distributed stochastic neighbor embedding
339 (t-SNE) analysis. Next, we observed that the proposed PFP loss was more
340 strongly correlated with the perceptual evaluation metrics than point-wise and
341 signal-level losses, thus achieving higher scores for standardized quality and
342 intelligibility evaluation metrics in the Voice Bank--DEMAND dataset.
343 </p>
344 </description>
345 </item>
346 <item>
347 <title>A Study on Efficiency in Continual Learning Inspired by Human Learning. (arXiv:2010.15187v1 [cs.LG])</title>
348 <link>http://fr.arxiv.org/abs/2010.15187</link>
349 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Ball_P/0/1/0/all/0/1">Philip J. Ball</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Li_Y/0/1/0/all/0/1">Yingzhen Li</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Lamb_A/0/1/0/all/0/1">Angus Lamb</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Zhang_C/0/1/0/all/0/1">Cheng Zhang</a></p>
350
351 <p>Humans are efficient continual learning systems; we continually learn new
352 skills from birth with finite cells and resources. Our learning is highly
353 optimized both in terms of capacity and time while not suffering from
354 catastrophic forgetting. In this work we study the efficiency of continual
355 learning systems, taking inspiration from human learning. In particular,
356 inspired by the mechanisms of sleep, we evaluate popular pruning-based
357 continual learning algorithms, using PackNet as a case study. First, we
358 identify that weight freezing, which is used in continual learning without
359 biological justification, can result in over $2\times$ as many weights being
360 used for a given level of performance. Secondly, we note the similarity in
361 human day and night time behaviors to the training and pruning phases
362 respectively of PackNet. We study a setting where the pruning phase is given a
363 time budget, and identify connections between iterative pruning and multiple
364 sleep cycles in humans. We show there exists an optimal choice of iteration
365 v.s. epochs given different tasks.
366 </p>
367 </description>
368 </item>
369 <item>
370 <title>Explicit stabilized multirate method for stiff stochastic differential equations. (arXiv:2010.15193v1 [math.NA])</title>
371 <link>http://fr.arxiv.org/abs/2010.15193</link>
372 <description><p>Authors: <a href="http://fr.arxiv.org/find/math/1/au:+Abdulle_A/0/1/0/all/0/1">Assyr Abdulle</a>, <a href="http://fr.arxiv.org/find/math/1/au:+Souza_G/0/1/0/all/0/1">Giacomo Rosilho de Souza</a></p>
373
374 <p>Stabilized explicit methods are particularly efficient for large systems of
375 stiff stochastic differential equations (SDEs) due to their extended stability
376 domain. However, they loose their efficiency when a severe stiffness is induced
377 by very few "fast" degrees of freedom, as the stiff and nonstiff terms are
378 evaluated concurrently. Therefore, inspired by [A. Abdulle, M. J. Grote, and G.
379 Rosilho de Souza, Preprint (2020), <a href="/abs/2006.00744">arXiv:2006.00744</a>] we introduce a stochastic
380 modified equation whose stiffness depends solely on the "slow" terms. By
381 integrating this modified equation with a stabilized explicit scheme we devise
382 a multirate method which overcomes the bottleneck caused by a few severely
383 stiff terms and recovers the efficiency of stabilized schemes for large systems
384 of nonlinear SDEs. The scheme is not based on any scale separation assumption
385 of the SDE and therefore it is employable for problems stemming from the
386 spatial discretization of stochastic parabolic partial differential equations
387 on locally refined grids. The multirate scheme has strong order 1/2, weak order
388 1 and its stability is proved on a model problem. Numerical experiments confirm
389 the efficiency and accuracy of the scheme.
390 </p>
391 </description>
392 </item>
393 <item>
394 <title>Reinforcement Learning for Sparse-Reward Object-Interaction Tasks in First-person Simulated 3D Environments. (arXiv:2010.15195v1 [cs.LG])</title>
395 <link>http://fr.arxiv.org/abs/2010.15195</link>
396 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Carvalho_W/0/1/0/all/0/1">Wilka Carvalho</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Liang_A/0/1/0/all/0/1">Anthony Liang</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Lee_K/0/1/0/all/0/1">Kimin Lee</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Sohn_S/0/1/0/all/0/1">Sungryull Sohn</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Lee_H/0/1/0/all/0/1">Honglak Lee</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Lewis_R/0/1/0/all/0/1">Richard L. Lewis</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Singh_S/0/1/0/all/0/1">Satinder Singh</a></p>
397
398 <p>First-person object-interaction tasks in high-fidelity, 3D, simulated
399 environments such as the AI2Thor virtual home-environment pose significant
400 sample-efficiency challenges for reinforcement learning (RL) agents learning
401 from sparse task rewards. To alleviate these challenges, prior work has
402 provided extensive supervision via a combination of reward-shaping,
403 ground-truth object-information, and expert demonstrations. In this work, we
404 show that one can learn object-interaction tasks from scratch without
405 supervision by learning an attentive object-model as an auxiliary task during
406 task learning with an object-centric relational RL agent. Our key insight is
407 that learning an object-model that incorporates object-attention into forward
408 prediction provides a dense learning signal for unsupervised representation
409 learning of both objects and their relationships. This, in turn, enables faster
410 policy learning for an object-centric relational RL agent. We demonstrate our
411 agent by introducing a set of challenging object-interaction tasks in the
412 AI2Thor environment where learning with our attentive object-model is key to
413 strong performance. Specifically, we compare our agent and relational RL agents
414 with alternative auxiliary tasks to a relational RL agent equipped with
415 ground-truth object-information, and show that learning with our object-model
416 best closes the performance gap in terms of both learning speed and maximum
417 success rate. Additionally, we find that incorporating object-attention into an
418 object-model's forward predictions is key to learning representations which
419 capture object-category and object-state.
420 </p>
421 </description>
422 </item>
423 <item>
424 <title>A fast and scalable computational framework for large-scale and high-dimensional Bayesian optimal experimental design. (arXiv:2010.15196v1 [math.NA])</title>
425 <link>http://fr.arxiv.org/abs/2010.15196</link>
426 <description><p>Authors: <a href="http://fr.arxiv.org/find/math/1/au:+Wu_K/0/1/0/all/0/1">Keyi Wu</a>, <a href="http://fr.arxiv.org/find/math/1/au:+Chen_P/0/1/0/all/0/1">Peng Chen</a>, <a href="http://fr.arxiv.org/find/math/1/au:+Ghattas_O/0/1/0/all/0/1">Omar Ghattas</a></p>
427
428 <p>We develop a fast and scalable computational framework to solve large-scale
429 and high-dimensional Bayesian optimal experimental design problems. In
430 particular, we consider the problem of optimal observation sensor placement for
431 Bayesian inference of high-dimensional parameters governed by partial
432 differential equations (PDEs), which is formulated as an optimization problem
433 that seeks to maximize an expected information gain (EIG). Such optimization
434 problems are particularly challenging due to the curse of dimensionality for
435 high-dimensional parameters and the expensive solution of large-scale PDEs. To
436 address these challenges, we exploit two essential properties of such problems:
437 the low-rank structure of the Jacobian of the parameter-to-observable map to
438 extract the intrinsically low-dimensional data-informed subspace, and the high
439 correlation of the approximate EIGs by a series of approximations to reduce the
440 number of PDE solves. We propose an efficient offline-online decomposition for
441 the optimization problem: an offline stage of computing all the quantities that
442 require a limited number of PDE solves independent of parameter and data
443 dimensions, and an online stage of optimizing sensor placement that does not
444 require any PDE solve. For the online optimization, we propose a swapping
445 greedy algorithm that first construct an initial set of sensors using leverage
446 scores and then swap the chosen sensors with other candidates until certain
447 convergence criteria are met. We demonstrate the efficiency and scalability of
448 the proposed computational framework by a linear inverse problem of inferring
449 the initial condition for an advection-diffusion equation, and a nonlinear
450 inverse problem of inferring the diffusion coefficient of a log-normal
451 diffusion equation, with both the parameter and data dimensions ranging from a
452 few tens to a few thousands.
453 </p>
454 </description>
455 </item>
456 <item>
457 <title>Forecasting Hamiltonian dynamics without canonical coordinates. (arXiv:2010.15201v1 [cs.LG])</title>
458 <link>http://fr.arxiv.org/abs/2010.15201</link>
459 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Choudhary_A/0/1/0/all/0/1">Anshul Choudhary</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Lindner_J/0/1/0/all/0/1">John F. Lindner</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Holliday_E/0/1/0/all/0/1">Elliott G. Holliday</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Miller_S/0/1/0/all/0/1">Scott T. Miller</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Sinha_S/0/1/0/all/0/1">Sudeshna Sinha</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Ditto_W/0/1/0/all/0/1">William L. Ditto</a></p>
460
461 <p>Conventional neural networks are universal function approximators, but
462 because they are unaware of underlying symmetries or physical laws, they may
463 need impractically many training data to approximate nonlinear dynamics.
464 Recently introduced Hamiltonian neural networks can efficiently learn and
465 forecast dynamical systems that conserve energy, but they require special
466 inputs called canonical coordinates, which may be hard to infer from data. Here
467 we significantly expand the scope of such networks by demonstrating a simple
468 way to train them with any set of generalised coordinates, including easily
469 observable ones.
470 </p>
471 </description>
472 </item>
473 <item>
474 <title>Micromobility in Smart Cities: A Closer Look at Shared Dockless E-Scooters via Big Social Data. (arXiv:2010.15203v1 [cs.SI])</title>
475 <link>http://fr.arxiv.org/abs/2010.15203</link>
476 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Feng_Y/0/1/0/all/0/1">Yunhe Feng</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Zhong_D/0/1/0/all/0/1">Dong Zhong</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Sun_P/0/1/0/all/0/1">Peng Sun</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Zheng_W/0/1/0/all/0/1">Weijian Zheng</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Cao_Q/0/1/0/all/0/1">Qinglei Cao</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Luo_X/0/1/0/all/0/1">Xi Luo</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Lu_Z/0/1/0/all/0/1">Zheng Lu</a></p>
477
478 <p>The micromobility is shaping first- and last-mile travels in urban areas.
479 Recently, shared dockless electric scooters (e-scooters) have emerged as a
480 daily alternative to driving for short-distance commuters in large cities due
481 to the affordability, easy accessibility via an app, and zero emissions.
482 Meanwhile, e-scooters come with challenges in city management, such as traffic
483 rules, public safety, parking regulations, and liability issues. In this paper,
484 we collected and investigated 5.8 million scooter-tagged tweets and 144,197
485 images, generated by 2.7 million users from October 2018 to March 2020, to take
486 a closer look at shared e-scooters via crowdsourcing data analytics. We
487 profiled e-scooter usages from spatial-temporal perspectives, explored
488 different business roles (i.e., riders, gig workers, and ridesharing
489 companies), examined operation patterns (e.g., injury types, and parking
490 behaviors), and conducted sentiment analysis. To our best knowledge, this paper
491 is the first large-scale systematic study on shared e-scooters using big social
492 data.
493 </p>
494 </description>
495 </item>
496 <item>
497 <title>Rosella: A Self-Driving Distributed Scheduler for Heterogeneous Clusters. (arXiv:2010.15206v1 [cs.DC])</title>
498 <link>http://fr.arxiv.org/abs/2010.15206</link>
499 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Wu_Q/0/1/0/all/0/1">Qiong Wu</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Manandhar_S/0/1/0/all/0/1">Sunil Manandhar</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Liu_Z/0/1/0/all/0/1">Zhenming Liu</a></p>
500
501 <p>Large-scale interactive web services and advanced AI applications make
502 sophisticated decisions in real-time, based on executing a massive amount of
503 computation tasks on thousands of servers. Task schedulers, which often operate
504 in heterogeneous and volatile environments, require high throughput, i.e.,
505 scheduling millions of tasks per second, and low latency, i.e., incurring
506 minimal scheduling delays for millisecond-level tasks. Scheduling is further
507 complicated by other users' workloads in a shared system, other background
508 activities, and the diverse hardware configurations inside datacenters.
509 </p>
510 <p>We present Rosella, a new self-driving, distributed approach for task
511 scheduling in heterogeneous clusters. Our system automatically learns the
512 compute environment and adjust its scheduling policy in real-time. The solution
513 provides high throughput and low latency simultaneously, because it runs in
514 parallel on multiple machines with minimum coordination and only performs
515 simple operations for each scheduling decision. Our learning module monitors
516 total system load, and uses the information to dynamically determine optimal
517 estimation strategy for the backends' compute-power. Our scheduling policy
518 generalizes power-of-two-choice algorithms to handle heterogeneous workers,
519 reducing the max queue length of $O(\log n)$ obtained by prior algorithms to
520 $O(\log \log n)$. We implement a Rosella prototype and evaluate it with a
521 variety of workloads. Experimental results show that Rosella significantly
522 reduces task response times, and adapts to environment changes quickly.
523 </p>
524 </description>
525 </item>
526 <item>
527 <title>Ground Roll Suppression using Convolutional Neural Networks. (arXiv:2010.15209v1 [eess.IV])</title>
528 <link>http://fr.arxiv.org/abs/2010.15209</link>
529 <description><p>Authors: <a href="http://fr.arxiv.org/find/eess/1/au:+Oliveira_D/0/1/0/all/0/1">Dario Augusto Borges Oliveira</a>, <a href="http://fr.arxiv.org/find/eess/1/au:+Semin_D/0/1/0/all/0/1">Daniil Semin</a>, <a href="http://fr.arxiv.org/find/eess/1/au:+Zaytsev_S/0/1/0/all/0/1">Semen Zaytsev</a></p>
530
531 <p>Seismic data processing plays a major role in seismic exploration as it
532 conditions much of the seismic interpretation performance. In this context,
533 generating reliable post-stack seismic data depends also on disposing of an
534 efficient pre-stack noise attenuation tool. Here we tackle ground roll noise,
535 one of the most challenging and common noises observed in pre-stack seismic
536 data. Since ground roll is characterized by relative low frequencies and high
537 amplitudes, most commonly used approaches for its suppression are based on
538 frequency-amplitude filters for ground roll characteristic bands. However, when
539 signal and noise share the same frequency ranges, these methods usually deliver
540 also signal suppression or residual noise. In this paper we take advantage of
541 the highly non-linear features of convolutional neural networks, and propose to
542 use different architectures to detect ground roll in shot gathers and
543 ultimately to suppress them using conditional generative adversarial networks.
544 Additionally, we propose metrics to evaluate ground roll suppression, and
545 report strong results compared to expert filtering. Finally, we discuss
546 generalization of trained models for similar and different geologies to better
547 understand the feasibility of our proposal in real applications.
548 </p>
549 </description>
550 </item>
551 <item>
552 <title>On Linearizability and the Termination of Randomized Algorithms. (arXiv:2010.15210v1 [cs.DC])</title>
553 <link>http://fr.arxiv.org/abs/2010.15210</link>
554 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Hadzilacos_V/0/1/0/all/0/1">Vassos Hadzilacos</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Hu_X/0/1/0/all/0/1">Xing Hu</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Toueg_S/0/1/0/all/0/1">Sam Toueg</a></p>
555
556 <p>We study the question of whether the "termination with probability 1"
557 property of a randomized algorithm is preserved when one replaces the atomic
558 registers that the algorithm uses with linearizable (implementations of)
559 registers. We show that in general this is not so: roughly speaking, every
560 randomized algorithm A has a corresponding algorithm A' that solves the same
561 problem if the registers that it uses are atomic or strongly-linearizable, but
562 does not terminate if these registers are replaced with "merely" linearizable
563 ones. Together with a previous result shown in [15], this implies that one
564 cannot use the well-known ABD implementation of registers in message-passing
565 systems to automatically transform any randomized algorithm that works in
566 shared-memory systems into a randomized algorithm that works in message-passing
567 systems: with a strong adversary the resulting algorithm may not terminate.
568 </p>
569 </description>
570 </item>
571 <item>
572 <title>Safety-Aware Cascade Controller Tuning Using Constrained Bayesian Optimization. (arXiv:2010.15211v1 [eess.SY])</title>
573 <link>http://fr.arxiv.org/abs/2010.15211</link>
574 <description><p>Authors: <a href="http://fr.arxiv.org/find/eess/1/au:+Konig_C/0/1/0/all/0/1">Christopher K&#xf6;nig</a>, <a href="http://fr.arxiv.org/find/eess/1/au:+Khosravi_M/0/1/0/all/0/1">Mohammad Khosravi</a>, <a href="http://fr.arxiv.org/find/eess/1/au:+Maier_M/0/1/0/all/0/1">Markus Maier</a>, <a href="http://fr.arxiv.org/find/eess/1/au:+Smith_R/0/1/0/all/0/1">Roy S. Smith</a>, <a href="http://fr.arxiv.org/find/eess/1/au:+Rupenyan_A/0/1/0/all/0/1">Alisa Rupenyan</a>, <a href="http://fr.arxiv.org/find/eess/1/au:+Lygeros_J/0/1/0/all/0/1">John Lygeros</a></p>
575
576 <p>This paper presents an automated, model-free, data-driven method for the safe
577 tuning of PID cascade controller gains based on Bayesian optimization. The
578 optimization objective is composed of data-driven performance metrics and
579 modeled using Gaussian processes. We further introduce a data-driven constraint
580 that captures the stability requirements from system data. Numerical evaluation
581 shows that the proposed approach outperforms relay feedback autotuning and
582 quickly converges to the global optimum, thanks to a tailored stopping
583 criterion. We demonstrate the performance of the method in simulations and
584 experiments on a linear axis drive of a grinding machine. For experimental
585 implementation, in addition to the introduced safety constraint, we integrate a
586 method for automatic detection of the critical gains and extend the
587 optimization objective with a penalty depending on the proximity of the current
588 candidate points to the critical gains. The resulting automated tuning method
589 optimizes system performance while ensuring stability and standardization.
590 </p>
591 </description>
592 </item>
593 <item>
594 <title>Away from Trolley Problems and Toward Risk Management. (arXiv:2010.15217v1 [cs.CY])</title>
595 <link>http://fr.arxiv.org/abs/2010.15217</link>
596 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Goodall_N/0/1/0/all/0/1">Noah J. Goodall</a></p>
597
598 <p>As automated vehicles receive more attention from the media, there has been
599 an equivalent increase in the coverage of the ethical choices a vehicle may be
600 forced to make in certain crash situations with no clear safe outcome. Much of
601 this coverage has focused on a philosophical thought experiment known as the
602 "trolley problem," and substituting an automated vehicle for the trolley and
603 the car's software for the bystander. While this is a stark and straightforward
604 example of ethical decision making for an automated vehicle, it risks
605 marginalizing the entire field if it is to become the only ethical problem in
606 the public's mind. In this chapter, I discuss the shortcomings of the trolley
607 problem, and introduce more nuanced examples that involve crash risk and
608 uncertainty. Risk management is introduced as an alternative approach, and its
609 ethical dimensions are discussed.
610 </p>
611 </description>
612 </item>
613 <item>
614 <title>StencilFlow: Mapping Large Stencil Programs to Distributed Spatial Computing Systems. (arXiv:2010.15218v1 [cs.DC])</title>
615 <link>http://fr.arxiv.org/abs/2010.15218</link>
616 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Licht_J/0/1/0/all/0/1">Johannes de Fine Licht</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Kuster_A/0/1/0/all/0/1">Andreas Kuster</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Matteis_T/0/1/0/all/0/1">Tiziano De Matteis</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Ben_Nun_T/0/1/0/all/0/1">Tal Ben-Nun</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Hofer_D/0/1/0/all/0/1">Dominic Hofer</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Hoefler_T/0/1/0/all/0/1">Torsten Hoefler</a></p>
617
618 <p>Spatial computing devices have been shown to significantly accelerate stencil
619 computations, but have so far relied on unrolling the iterative dimension of a
620 single stencil operation to increase temporal locality. This work considers the
621 general case of mapping directed acyclic graphs of heterogeneous stencil
622 computations to spatial computing systems, assuming large input programs
623 without an iterative component. StencilFlow maximizes temporal locality and
624 ensures deadlock freedom in this setting, providing end-to-end analysis and
625 mapping from a high-level program description to distributed hardware. We
626 evaluate the generated architectures on an FPGA testbed, demonstrating the
627 highest single-device and multi-device performance recorded for stencil
628 programs on FPGAs to date, then leverage the framework to study a complex
629 stencil program from a production weather simulation application. Our work
630 enables productively targeting distributed spatial computing systems with large
631 stencil programs, and offers insight into architecture characteristics required
632 for their efficient execution in practice.
633 </p>
634 </description>
635 </item>
636 <item>
637 <title>Geometric Sampling of Networks. (arXiv:2010.15221v1 [math.DG])</title>
638 <link>http://fr.arxiv.org/abs/2010.15221</link>
639 <description><p>Authors: <a href="http://fr.arxiv.org/find/math/1/au:+Barkanass_V/0/1/0/all/0/1">Vladislav Barkanass</a>, <a href="http://fr.arxiv.org/find/math/1/au:+Jost_J/0/1/0/all/0/1">J&#xfc;rgen Jost</a>, <a href="http://fr.arxiv.org/find/math/1/au:+Saucan_E/0/1/0/all/0/1">Emil Saucan</a></p>
640
641 <p>Motivated by the methods and results of manifold sampling based on Ricci
642 curvature, we propose a similar approach for networks. To this end we make
643 appeal to three types of discrete curvature, namely the graph Forman-, full
644 Forman- and Haantjes-Ricci curvatures for edge-based and node-based sampling.
645 We present the results of experiments on real life networks, as well as for
646 square grids arising in Image Processing. Moreover, we consider fitting Ricci
647 flows and we employ them for the detection of networks' backbone. We also
648 develop embedding kernels related to the Forman-Ricci curvatures and employ
649 them for the detection of the coarse structure of networks, as well as for
650 network visualization with applications to SVM. The relation between the Ricci
651 curvature of the original manifold and that of a Ricci curvature driven
652 discretization is also studied.
653 </p>
654 </description>
655 </item>
656 <item>
657 <title>Exploring complex networks with the ICON R package. (arXiv:2010.15222v1 [cs.SI])</title>
658 <link>http://fr.arxiv.org/abs/2010.15222</link>
659 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Wadhwa_R/0/1/0/all/0/1">Raoul R. Wadhwa</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Scott_J/0/1/0/all/0/1">Jacob G. Scott</a></p>
660
661 <p>We introduce ICON, an R package that contains 1075 complex network datasets
662 in a standard edgelist format. All provided datasets have associated citations
663 and have been indexed by the Colorado Index of Complex Networks - also referred
664 to as ICON. In addition to supplying a large and diverse corpus of useful
665 real-world networks, ICON also implements an S3 generic to work with the
666 network and ggnetwork R packages for network analysis and visualization,
667 respectively. Sample code in this report also demonstrates how ICON can be used
668 in conjunction with the igraph package. Currently, the Comprehensive R Archive
669 Network hosts ICON v0.4.0. We hope that ICON will serve as a standard corpus
670 for complex network research and prevent redundant work that would be otherwise
671 necessary by individual research groups. The open source code for ICON and for
672 this reproducible report can be found at https://github.com/rrrlw/ICON.
673 </p>
674 </description>
675 </item>
676 <item>
677 <title>A Visuospatial Dataset for Naturalistic Verb Learning. (arXiv:2010.15225v1 [cs.CL])</title>
678 <link>http://fr.arxiv.org/abs/2010.15225</link>
679 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Ebert_D/0/1/0/all/0/1">Dylan Ebert</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Pavlick_E/0/1/0/all/0/1">Ellie Pavlick</a></p>
680
681 <p>We introduce a new dataset for training and evaluating grounded language
682 models. Our data is collected within a virtual reality environment and is
683 designed to emulate the quality of language data to which a pre-verbal child is
684 likely to have access: That is, naturalistic, spontaneous speech paired with
685 richly grounded visuospatial context. We use the collected data to compare
686 several distributional semantics models for verb learning. We evaluate neural
687 models based on 2D (pixel) features as well as feature-engineered models based
688 on 3D (symbolic, spatial) features, and show that neither modeling approach
689 achieves satisfactory performance. Our results are consistent with evidence
690 from child language acquisition that emphasizes the difficulty of learning
691 verbs from naive distributional data. We discuss avenues for future work on
692 cognitively-inspired grounded language learning, and release our corpus with
693 the intent of facilitating research on the topic.
694 </p>
695 </description>
696 </item>
697 <item>
698 <title>Speech-Based Emotion Recognition using Neural Networks and Information Visualization. (arXiv:2010.15229v1 [cs.HC])</title>
699 <link>http://fr.arxiv.org/abs/2010.15229</link>
700 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Almahmoud_J/0/1/0/all/0/1">Jumana Almahmoud</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Kikkeri_K/0/1/0/all/0/1">Kruthika Kikkeri</a></p>
701
702 <p>Emotions recognition is commonly employed for health assessment. However, the
703 typical metric for evaluation in therapy is based on patient-doctor appraisal.
704 This process can fall into the issue of subjectivity, while also requiring
705 healthcare professionals to deal with copious amounts of information. Thus,
706 machine learning algorithms can be a useful tool for the classification of
707 emotions. While several models have been developed in this domain, there is a
708 lack of userfriendly representations of the emotion classification systems for
709 therapy. We propose a tool which enables users to take speech samples and
710 identify a range of emotions (happy, sad, angry, surprised, neutral, clam,
711 disgust, and fear) from audio elements through a machine learning model. The
712 dashboard is designed based on local therapists' needs for intuitive
713 representations of speech data in order to gain insights and informative
714 analyses of their sessions with their patients.
715 </p>
716 </description>
717 </item>
718 <item>
719 <title>Construction Payment Automation Using Blockchain-Enabled Smart Contracts and Reality Capture Technologies. (arXiv:2010.15232v1 [cs.CR])</title>
720 <link>http://fr.arxiv.org/abs/2010.15232</link>
721 <description><p>Authors: <a href="http://fr.arxiv.org/find/cs/1/au:+Hamledari_H/0/1/0/all/0/1">Hesam Hamledari</a>, <a href="http://fr.arxiv.org/find/cs/1/au:+Fischer_M/0/1/0/all/0/1">Martin Fischer</a></p>
722
723 <p>This paper presents a smart contract-based solution for autonomous
724 administration of construction progress payments. It bridges the gap between
725 payments (cash flow) and the progress assessments at job sites (product flow)
726 enabled by reality capture technologies and building information modeling
727 (BIM). The approach eliminates the reliance on the centralized and heavily
728 intermediated mechanisms of existing payment applications. The construction
729 progress is stored in a distributed manner using content addressable file
730 sharing; it is broadcasted to a smart contract which automates the on-chain
731 payment settlements and the transfer of lien rights. The method was
732 successfully used for processing payments to 7 subcontractors in two commercial
733 construction projects where progress monitoring was performed using a
734 camera-equipped unmanned aerial vehicle (UAV) and an unmanned ground vehicle
735 (UGV) equipped with a laser scanner. The results show promise for the method's
736 potential for increasing the frequency, granularity, and transparency of
737 payments. The paper is concluded with a discussion of implications for project
738 management, introducing a new model of project as a singleton state machine.
739 </p>
740 </description>
741 </item>
742 <item>
743 <title>Accurate Prostate Cancer Detection and Segmentation on Biparametric MRI using Non-local Mask R-CNN with Histopathological Ground Truth. (arXiv:2010.15233v1 [eess.IV])</title>
744 <link>http://fr.arxiv.org/abs/2010.15233</link>
codemadness.org:70 /git/sfeed_tests/file/input/sfeed/realworld/arxiv.org.rss.091.xml.gph:755: line too long