-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
277 lines (257 loc) · 17.3 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
<!DOCTYPE HTML>
<!--
Strata by HTML5 UP
-->
<html>
<head>
<title>Anomaly detection at 40 MHz: Data challenge</title>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1, user-scalable=no" />
<link rel="stylesheet" href="assets/css/main.css" />
<script>
!function(t,e){var o,n,p,r;e.__SV||(window.posthog=e,e._i=[],e.init=function(i,s,a){function g(t,e){var o=e.split(".");2==o.length&&(t=t[o[0]],e=o[1]),t[e]=function(){t.push([e].concat(Array.prototype.slice.call(arguments,0)))}}(p=t.createElement("script")).type="text/javascript",p.async=!0,p.src=s.api_host+"/static/array.js",(r=t.getElementsByTagName("script")[0]).parentNode.insertBefore(p,r);var u=e;for(void 0!==a?u=e[a]=[]:a="posthog",u.people=u.people||[],u.toString=function(t){var e="posthog";return"posthog"!==a&&(e+="."+a),t||(e+=" (stub)"),e},u.people.toString=function(){return u.toString(1)+".people (stub)"},o="capture identify alias people.set people.set_once set_config register register_once unregister opt_out_capturing has_opted_out_capturing opt_in_capturing reset isFeatureEnabled onFeatureFlags".split(" "),n=0;n<o.length;n++)g(u,o[n]);e._i.push([i,s,a])},e.__SV=1)}(document,window.posthog||[]);
posthog.init('phc_sOP2e7XELOKn4kXPgzjV2L6VSi5JuGPP0RhsIt2rRVF',{api_host:'https://app.posthog.com'})
</script>
</head>
<body class="is-preload">
<!-- Header -->
<header id="header">
<div class="inner">
<!-- <a href="#" class="image avatar"><img src="images/logo.png" alt="" /></a> -->
<h1><strong>Welcome to the <br />
Anomaly Detection <br />
Data Challenge 2021!</strong></h1>
</div>
<!-- <div class="right">
<ul class="selector">
<li><a href="https://zenodo.org/record/5046389#.YNyCfi0Rpqs" class="button">Dataset</a></li>
<li><a href="https://zenodo.org/record/" class="button">Data Reader</a></li>
<li><a href="https://zenodo.org/record/" class="button">Models</a></li>
<li><a href="https://github.com/mpp-hep/ADC2021-results" class="button">Upload contribution!</a></li>
</ul>
</div> -->
</header>
<!-- Main -->
<div id="main">
<!-- One -->
<section id="one">
<header class="major">
<h1>Unsupervised New Physics detection at 40 MHz</h1>
</header>
<p> In this challenge, you will develop algorithms for detecting New Physics by reformulating the problem as an out-of-distribution detection task.
Armed with four-vectors of the highest-momentum jets, electrons, and muons produced in a LHC collision event, together with the missing transverse energy (missing E<sub>T</sub>),
the goal is to find a-priori unknown and rare New Physics hidden in a data sample dominated by ordinary Standard Model processes, using anomaly detection approaches.
<br />
<header class="minor">
<h2>Real-time event filtering</h2>
</header>
The algorithms are intended to be deployed in the first stage of the real-time event filter processing system of LHC experiments (Level 1 or L1 trigger), where the available bandwidth, latency and resources are strictly limited.
Such limitations constrain the design of the algorithm. To emulate the constraints in terms of bandwith only the leading 10 jets, 4 muons, 4 electrons and the missing E<sub>T</sub> will be provided to be used as input to the algorithm.
Furthermore, only a maximum number of bits is available for the representation of the η, φ, and the transverse momentum p<sub>T</sub> of each physics object.
The effect of such <em>quantization</em> of the inputs can be studied for instance with QKeras (see below).
<br />
<br />
Further complicating things, is that trigger algorithms must have a very low latency, of around one microsecond, and in addition must be highly resource efficient.
The task is therefore to design an architecture that maximises the sensitivity for New Physics, but at the lowest possible resource and latency cost.
An estimate of the algorithm efficiency can be obtained by calculating the number of floating-point operations of the model (FLOPs).
An example of how to calculate FLOPs is provided <a href="https://github.com/mpp-hep/ADC2021-examplecode/blob/main/computeFLOPs.ipynb">here</a>.<br />
<br />
<header class="minor">
<h2>FPGA inference with hls4ml</h2>
</header>
To meet the latency constraint, field-programmable gate arrays (FPGAs) are typically deployed in the L1 trigger farm of LHC experiments.
Algorithm complexity, and possibly accuracy, could be enhanced by deploying deep learning algorithms on these devices.
To do so, the participants can use the <span style="font-family:'Courier New'">hls4ml</span> library (see below), which was introduced as a tool to automatically translate a given DL model
into high-level synthesis (HLS) code describing the electronic circuit.
The HLS code is automatically optimized to achieve the lowest latency and resources for your favourite DL model.
With this tool the user can obtain an estimate of these algorithm performance metrics for a deployment on FPGAs for the L1 trigger.
<br />
<br />
The participants should aim at achieving the highest true positive rate at a fixed false positive rate of 10<sup>-5</sup>,
roughly corresponding to the available output data rate budget for a trigger algorithm, at the lowest possible latency and resource utilization.
<br />
<br />
Do not miss the introduction to the challenge at <a href="https://indico.cern.ch/event/980214/contributions/4413658/">ML4Jets2021</a>!<br />
<br />
<header class="minor">
<h2>Datasets and example code</h2>
</header>
Links to the data samples are given below. The training dataset contains a cocktail of Standard Model events intended for training and testing.
In addition, we provide four different Beyond Standard Model signals to study the algorithm performance for the New Physics detection task.
Finally, we provide a "blackbox" test sample, containing a mixture of SM and an unknown New Physics signal.
Information on how to upload your contribution based on this sample are provided below.
A full description of the SM and BSM processes considered, the simulation setup, and observables in the dataset can be found in <a href="https://arxiv.org/abs/2107.02157">this paper</a>.
<br />
<br />
As deep autoencoders are becoming more and more popular for the anomaly detection task we have prepared <a href="https://github.com/mpp-hep/ADC2021-examplecode">examples</a> of these architectures.
In this example code you will also see how to read in the dataset and how to evaluate the performance.
<br><br />
<ul class="actions">
<li><a href="https://zenodo.org/record/5072068#.YOSIaG6xVbU" class="button" style="width:230px;">Blackbox Dataset</a></li>
<li><a href="https://zenodo.org/record/5046428#.YNyFtRMzZqt" class="button" style="width:230px;">SM Dataset: Training</a></li>
</ul>
<ul class="actions">
<li><a href="https://zenodo.org/record/7152590#.YNyFtxMzZqt" class="button" style="width:230px;">BSM Dataset: A to 4 leptons</a></li>
<li><a href="https://zenodo.org/record/7152599#.YN7TyW6xVbV" class="button" style="width:230px;">BSM Dataset: LQ to bτ</a></li>
</ul>
<ul class="actions">
<li><a href="https://zenodo.org/record/7152614#.YOLJ8m6xXfY" class="button" style="width:230px;">BSM Dataset: h<sup>0</sup> to ττ</a></li>
<li><a href="https://zenodo.org/record/7152617#.YOLLU26xVbU" class="button" style="width:230px;">BSM Dataset: h<sup>+</sup> to τν</a></li>
</ul>
<!---ul class="actions">
<li><a href="https://github.com/anomalyHackathon/L1_anomalyDetection/" class="button">Code: Data Reader</a></li>
<li><a href="https://github.com/anomalyHackathon/L1_anomalyDetection/" class="button">Code: Example architecture</a></li>
</ul--->
<!---header class="minor">
<h2>The <span style="font-family:'Courier New'">hls4ml</span> library</h2>
</header--->
<header class="minor">
<h2>How to submit a contribution</h2>
</header>
In order to participate in the challenge, you need to evaluate your model on the blackbox dataset,
containing a mixture of Standard Model and New Physics events. Each team will be asked to submit
the list of the 10<sup>-5</sup> fraction of events passing a selection on the algorithm metrics as well as the total number of FLOPS for the architecture.
We also encourage the teams to achieve a full design for deployment on FPGAs to obtain a more precise estimate
of the feasibility for your algorithm to run in the L1 trigger. The design can be achieved through <span style="font-family:'Courier New'">hls4ml</span> (see tutorials and other resources below) or other tools.
In this case, the team will also submit the estimated performance on the FPGA in terms of latency and resources.
Note that while the tool support the translation for the building blocks of most of the popular DL architectures, more custom ingredients might be missing.
In this case, the <span style="font-family:'Courier New'">hls4ml</span> developers team can help on a best-effort basis.
The submission should be in a form of a HDF5 file containing a numpy array with the identification numbers of each selected event, plus a dictionary with the algorithm deployment performance, as shown in this <a href="https://github.com/mpp-hep/ADC2021-results/blob/master/upload-your-data.py">example</a>.
The name of the file should be unique and correspond to a participant/group name. Email us to confirm you have submit your contribution! <strong> Deadline: ML4Jets 2022! </strong>
<br><br />
<ul class="actions">
<li><a href="https://github.com/mpp-hep/ADC2021-results" class="button">Upload contribution!</a></li>
</ul>
<br />
<strong> Good luck! </strong>
<br />
<strong> Deadline: ML4Jets 2022! </strong>
</section>
<section id="features">
<div class="container">
<header>
<h2><strong>Further Reading</strong></h2>
</header>
<div class="row aln-center">
<!-- Feature -->
<div class="col-4 col-6-medium col-12-small">
<section>
<a href="https://arxiv.org/abs/2006.10159" class="image fit thumb image"><img src="images/pynq_floorplan_final.jpeg" alt="" height='100'/></a>
<header>
<h3><span style="font-family:'Courier New'">hls4ml</span><br /> and Google QKeras</h3>
</header>
</section>
</div>
<div class="col-4 col-6-medium col-12-small">
<section>
<a href="https://arxiv.org/search/?query=hls4ml&searchtype=all&source=header" class="image fit thumb image"><img src="images/hls4ml-arxiv.png" alt="" height='100'/></a>
<header>
<h3><span style="font-family:'Courier New'">hls4ml</span>:<br /> published work</h3>
</header>
</section>
</div>
<div class="col-4 col-6-medium col-12-small">
<section>
<a href="https://fastmachinelearning.org/hls4ml/index.html" class="image fit thumb image"><img src="images/hls4ml-doc.png" alt="" height='100'/></a>
<header>
<h3><span style="font-family:'Courier New'">hls4ml</span>:<br /> library documentation</h3>
</header>
<!-- <p>An overview over the hls4ml project</p> -->
</section>
</div>
<!-- Feature end -->
<!-- Feature -->
<div class="col-4 col-6-medium col-12-small">
<section>
<a href="https://www.youtube.com/watch?v=0fe_e5OEF54" class="image fit thumb image"><img src="images/hls4ml-tutorial.png" alt="" height='100'/></a>
<header>
<h3><span style="font-family:'Courier New'">hls4ml</span>:<br /> video tutorial</h3>
</header>
<!-- <p>Jupyter notebooks on how to design ML firmware with hls4ml</p> -->
</section>
</div>
<!-- Feature end -->
<!-- Feature -->
<div class="col-4 col-6-medium col-12-small">
<section>
<a href="https://github.com/fastmachinelearning/hls4ml-tutorial" class="image fit thumb image"><img src="images/hls4ml-code.png" alt="" height='100'/></a>
<header>
<h3><span style="font-family:'Courier New'">hls4ml</span>:<br /> notebooks</h3>
</header>
<!-- <p>Jupyter notebooks on how to design ML firmware with hls4ml</p> -->
</section>
</div>
<!-- Feature end -->
<div class="col-4 col-6-medium col-12-small">
<section>
<a href="https://indico.cern.ch/event/1025534/" class="image fit thumb image"><img src="images/iml.png" alt="" height='100' /></a>
<header>
<h3> Anomaly Detection:<br /> AD @ 40 MHz</h3>
</header>
<!-- <p>A talk on 40 MHz anomaly detection at IML</p> -->
</section>
</div>
<!-- Feature end -->
</div>
</section>
<!-- Three -->
<section id="three">
<h2>Get In Touch</h2>
<p>Feel free to drop us a message below!</p>
<div class="row">
<div class="col-8 col-12-small">
<form method="post" action="#">
<div class="row gtr-uniform gtr-50">
<div class="col-6 col-12-xsmall"><input type="text" name="name" id="name" placeholder="Name" /></div>
<div class="col-6 col-12-xsmall"><input type="email" name="email" id="email" placeholder="Email" /></div>
<div class="col-12"><textarea name="message" id="message" placeholder="Message" rows="4"></textarea></div>
</div>
</form>
<ul class="actions">
<li><input type="submit" value="Send Message" /></li>
</ul>
</div>
<div class="col-4 col-12-small">
<ul class="labeled-icons">
<li>
<h3 class="icon fa-home"><span class="label">Address</span></h3>
machine learning for Particle Physics (mPP)
</li>
<li>
<h3 class="icon fa-envelope-o"><span class="label">Email</span></h3>
<a href="#">[email protected]</a>
<h3 class="icon fa-envelope-o"><span class="label">Email</span></h3>
<a href="#">[email protected]</a>
<h3 class="icon fa-envelope-o"><span class="label">Email</span></h3>
<a href="#">[email protected]</a>
<h3 class="icon fa-envelope-o"><span class="label">Email</span></h3>
<a href="#">[email protected]</a>
<h3 class="icon fa-envelope-o"><span class="label">Email</span></h3>
<a href="#">[email protected]</a>
</li>
</ul>
</div>
</div>
</section>
</div>
<!-- Footer -->
<footer id="footer">
<div class="inner">
<ul class="icons">
<li><a href="https://github.com/mpp-hep/" class="icon fa-github"><span class="label">Github</span></a></li>
<li><a href="[email protected]" class="icon fa-envelope-o"><span class="label">Email</span></a></li>
</ul>
<!-- <ul class="copyright">
<li>© Untitled</li><li>Design: <a href="http://html5up.net">HTML5 UP</a></li>
</ul> -->
</div>
</footer>
<!-- Scripts -->
<script src="assets/js/jquery.min.js"></script>
<script src="assets/js/jquery.poptrox.min.js"></script>
<script src="assets/js/browser.min.js"></script>
<script src="assets/js/breakpoints.min.js"></script>
<script src="assets/js/util.js"></script>
<script src="assets/js/main.js"></script>
</body>
</html>