This repository has been archived by the owner on Sep 13, 2022. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 0
/
presentation.tex
198 lines (169 loc) · 6.1 KB
/
presentation.tex
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
% Created 2020-04-23 Thu 16:45
% Intended LaTeX compiler: pdflatex
\documentclass[bigger]{beamer}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{graphicx}
\usepackage{grffile}
\usepackage{longtable}
\usepackage{wrapfig}
\usepackage{rotating}
\usepackage[normalem]{ulem}
\usepackage{amsmath}
\usepackage{textcomp}
\usepackage{amssymb}
\usepackage{capt-of}
\usepackage{hyperref}
\usepackage{natbib}
\usepackage{bm}
\usepackage{pgfplots}
\usepgfplotslibrary{groupplots,dateplot}
\usetikzlibrary{patterns,shapes.arrows}
\pgfplotsset{compat=newest}
\usepackage{dsfont}
\usepackage{xcolor}
\usepackage{listings}
\DeclareMathOperator{\TopHat}{TH}
\DeclareMathOperator{\CDF}{CDF}
\usetheme{Madrid}
\author{Aleksandr Petrosyan}
\date{\today}
\title{Cosmological Parameter estimation using Bayesian accelerated Machine learning.}
\hypersetup{
pdfauthor={Aleksandr Petrosyan},
pdftitle={Cosmological Parameter estimation using Bayesian accelerated Machine learning.},
pdfkeywords={},
pdfsubject={},
pdfcreator={Emacs 26.3 (Org mode 9.1.9)},
pdflang={English}}
\begin{document}
\maketitle
\begin{frame}[label={sec:orge6b6c72}]{What is \(\Lambda\)CDM.}
It's the accepted standard model of Cosmology.
Has the following main parameters:
\begin{itemize}
\item \(\Omega_{b}h^{2}\) --- Physical baryon density.
\item \(\Omega_{c}h^{2}\) --- Physical dark matter density.
\item \(\tau_{0}\) --- Age of the universe.
\item \(n_{s}\) --- Scalar spectral index.
\item \(\tau_\text{reio}\) --- Re-ionization optical depth.
\item \(\Delta_{R}^{2}\) --- Curvature fluctuation amplitude.
\end{itemize}
\end{frame}
\begin{frame}[label={sec:orga7c365a}]{How do we measure it?}
Have Planck, WMap and many other sources of data.
Do Bayesian inference. Long discussion what this actually means.
The gist:
\begin{itemize}
\item Data using \(\chi^{2}\)defines \({\cal L}\) --- likelihood.
\item Prior information (other experiments) give us \(\pi\) --- the prior.
\item From Bayes' theorem, compute \emph{evidence} \({\cal Z}\) and \emph{posterior} \({\cal P}\).
\item From posterior, \emph{marginalise} the \emph{calibration parameters}.
\end{itemize}
\end{frame}
\begin{frame}[label={sec:orgf5b3dfe}]{Example of posterior.}
Started with a \alert{Gaussian}. Then Gave an \alert{offset}.
\(\ln {\cal Z} = -62 \pm 1\) exactly what we expect. Except for
\alert{one} case.
\includegraphics[width=.75\linewidth]{./illustrations/convergence.pdf}
\end{frame}
\begin{frame}[label={sec:org083fab8}]{How do we do it?}
We are spoiled for choice. But can only use Monte-Carlo in many
dimensions (lots of parameters).
The fastest methods given below in order of stability/speed.
\begin{block}{Monte-Carlo methods}
\begin{itemize}
\item Nested sampling.
\item Hamiltonian Monte-Carlo.
\item Metropolis-Hastings.
\end{itemize}
\end{block}
\end{frame}
\begin{frame}[label={sec:orgcf084f1}]{Can we make it go fast?}
Yes and no.
If we give the prior as the posterior, we will converge the
fastest. But the evidence is all wrong. The posterior is also wrong.
We can mitigate this, by using an \emph{intuitive Gaussian}. But it gets
the wrong answer if we were wrong about either the place or the size
of the Posterior.
\begin{block}{My Intuitive Gaussian.}
Take \(\tilde{\pi}(\theta) = {\cal P}(\theta)\), change \({\cal L}\), such that.
\begin{equation*}
\tilde{\cal L}(\theta)\tilde{\pi}(\theta) = {\cal L}(\theta)\pi(\theta)
\end{equation*}
\end{block}
\end{frame}
\begin{frame}[label={sec:org958d1e9}]{Can we safely make it go fast? Posterior re-partitioning.}
\begin{block}{Chen-Ferroz-Hobson PPR.}
\begin{equation*}
\tilde{\pi}(\bm{\bm{\theta}};\beta) = \cfrac{\pi(\bm{\theta})^{\beta}}{Z(\beta)\{\pi\}},
\end{equation*}
\begin{equation*}
Z(\beta)\{\pi\} = \int_{\bm{\theta} \in \Psi} \pi(\bm{\bm{\theta}})^{\beta}d\bm{\bm{\theta}}.
\end{equation*}
\begin{equation*}
\tilde{\cal L}(\bm{\theta}) = {\cal L}(\bm{\theta}) Z(\beta)\{\pi\} \cdot \pi^{1-\beta}(\bm{\theta}).
\end{equation*}
\end{block}
This can mitigate the issues of size, but not place.
\end{frame}
\begin{frame}[label={sec:org73e18f6}]{Works well, but not if we have an offset.}
\begin{center}
\includegraphics[width=.9\linewidth]{./illustrations/convergence.pdf}
\end{center}
\end{frame}
\begin{frame}[label={sec:orgb21b8c0}]{My discovery.}
\begin{block}{My Stochastic superpositional isometric model mixture.}
\begin{equation*}
\tilde{\pi}(\bm{\theta}; \beta) \triangleq \begin{cases}
\tilde{\pi}_{1}(\bm{\theta}) & \text{with probability } \beta_{1},\\
& \vdots,\\
\tilde{\pi}_{n}(\bm{\theta}) & \text{with probability } (1- \sum_{i}^{m}\beta_{i}),
\end{cases}
\end{equation*}
\begin{equation*}
\tilde{\cal L}(\bm{\theta}; \bm{\beta}) \triangleq
\begin{cases}
\tilde{\cal L}_{1}(\bm{\theta}) & \text{with probability } \beta_{1},\\
&\vdots,\\
\tilde{\cal L}_{m}(\bm{\theta}) & \text{with probability} (1- \sum_{i}^{m}\beta_{i}).
\end{cases}
\end{equation*}
\begin{equation*}
\tilde{\pi}(\bm{\theta}; \bm{\beta}) = \tilde{\pi}(\bm{\theta})_{i} \Leftrightarrow \tilde{\cal L}(\bm{\theta}; \bm{\beta}) = \tilde{\cal L}_{m}(\bm{\theta}; \bm{\beta}),
\end{equation*}
\end{block}
\end{frame}
\begin{frame}
\frametitle{This is what it looks like}
\input{./illustrations/mixture_2.tex}
\end{frame}
\begin{frame}
\frametitle{This is what it looks like}
\input{./illustrations/mixture_3.tex}
\end{frame}
\begin{frame}
\frametitle{This is what it looks like}
\input{./illustrations/mixture_4.tex}
\end{frame}
\begin{frame}[label={sec:orgcde55a9}]{This is really FAST.}
\begin{figure}
\input{./illustrations/benchmark.tex}
\end{figure}
\end{frame}
\begin{frame}[label={sec:org34912c4}]{And Accurate.}
\begin{figure}
\input{./illustrations/evidence-drift.tex}
\end{figure}
\end{frame}
\begin{frame}[fragile,label={sec:org7a68446}]{What this can do?}
Examples:
\begin{itemize}
\item Can accelerate if you know \alert{roughly} what to expect.
\item Can make more robust if you are wrong about the where.
\item Can make more robust if you are unsure about biases.
\item This can allow you to chain models.
\item You can run a \texttt{CosmoChord} on a laptop, \texttt{Cobaya} on a desktop.
\end{itemize}
\end{frame}
\end{document}