-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
177 lines (164 loc) · 5.67 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
<!doctype html>
<! -- TODO: create a custom Spell theme and apply it to my slides. Learn all the reveal.js goodies. -->
<html>
<head>
<link rel="stylesheet" href="css/reveal.css">
<link rel="stylesheet" href="css/theme/spell.css" id="theme">
</head>
<body>
<div class="reveal">
<div class="slides">
<section data-markdown>
<textarea data-template>
<img src="static/logo.png" style="width:100px;"/>
### Battle-Tested Techniques for Scoping Machine Learning Projects
</textarea>
</section>
<section data-background-image="static/Spell-Overview.png">
</section>
<section data-markdown>
<textarea data-template>
Observation 1: model development projects follow a well-defined performance curve.
</textarea>
</section>
<!-- <section data-markdown>
<textarea data-template>
</textarea>
</section> -->
<section data-markdown>
<textarea data-template>
</textarea>
</section>
<section data-markdown>
<textarea data-template>
</textarea>
</section>
<section data-markdown>
<textarea data-template>
Observation 2: model performance on a novel dataset is impossible to forecast.
</textarea>
</section>
<section>
<section data-markdown>
<textarea data-template>
In the influential 2014 paper (["Hidden Technical Debt in Machine Learning Systems"](https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdf)) a group of Google engineers coined the CACE principle: <b>changing anything changes everything</b>.
</textarea>
</section>
<section data-markdown>
<textarea data-template>
<span style="color:lightgray">In the influential 2014 paper ("Hidden Technical Debt in Machine Learning Systems") a group of Google engineers coined the CACE principle:</span> <b>changing anything changes everything</b>.
This is the source of most of the complexity in practical deep learning systems!
</textarea>
</section>
</section>
<section data-markdown>
<textarea data-template>
Some machine learning problems are intrinsically hard.
<img src="static/hard-problem.png" style="width:66%;"/>
</textarea>
</section>
<section data-markdown>
<textarea data-template>
Some machine learning problems are intrinsically easy.
<img src="static/easy-problem.png" style="width:66%;"/>
</textarea>
</section>
<section data-markdown>
<textarea data-template>
</textarea>
</section>
<section data-markdown>
<textarea data-template>
Project scoping is a hard problem in software.
It's even harder in ML.
</textarea>
</section>
<section>
<section data-markdown>
<textarea data-template>
### Tip 1
<p>
Separate the project into <b>explore</b> and <b>implement</b> phases.
</p>
<p>
Use explore time to estimate project cost.
</p>
<br/><br/>
[h/t "Doing Freelance Data Consulting in 2019"](https://www.ethanrosenthal.com/2020/01/08/freelance-ds-consulting/)
</textarea>
</section>
</section>
<section>
<section data-markdown>
<textarea data-template>
### Tip 2
Collect human and model performance benchmarks on the dataset.
</textarea>
</section>
<section data-markdown>
<textarea data-template>
</textarea>
<!-- <span class="article-reference">
<a href="https://paperswithcode.com/">PapersWithCode</a>
</span> -->
</section>
<section data-markdown>
<textarea data-template>
</textarea>
<!-- <span class="article-reference">
<a href="https://karpathy.github.io/2014/09/02/what-i-learned-from-competing-against-a-convnet-on-imagenet/">Andrej Karpathy, "What I Learned From Competiing Against a Convnet on Imagenet"</a>
</span> -->
</section>
</section>
<section data-markdown>
<textarea data-template>
### Tip 3
Negotiate performance breakpoints with stakeholders ahead of time.
What is "good enough"? What is "great"? This is typically a business decision.
</textarea>
</section>
<section data-markdown>
<textarea data-template>
### Tip 4
Keep updating your time cost estimate as you proceed.
Keep in contact with stakeholders and make sure they understand that performance improvements will be a non-linear.
</textarea>
</section>
<section data-markdown>
<textarea data-template>
### Tip 5
Have a library (or a zoo!) of well-optimized models you're comfortable using.
</textarea>
</section>
<section data-markdown>
<textarea data-template>
@residentmario
</textarea>
</section>
</div>
</div>
<script src="js/reveal.js"></script>
<!-- This section configures plugins. -->
<script>
Reveal.initialize({
controls: true,
progress: true,
center: true,
hash: true,
transition: 'slide',
dependencies: [
// Enables markdown in slides, using the <section data-markdown><textarea data-template></></> construction.
{ src: 'plugin/markdown/marked.js', condition: function() { return !!document.querySelector( '[data-markdown]' ); } },
{ src: 'plugin/markdown/markdown.js', condition: function() { return !!document.querySelector( '[data-markdown]' ); } },
]
});
</script>
</body>
</html>