-
Notifications
You must be signed in to change notification settings - Fork 18
/
Copy path03-07-01-ACA-Input-Post-Processing.tex
234 lines (215 loc) · 10.9 KB
/
03-07-01-ACA-Input-Post-Processing.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
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
% move all configuration stuff into includes file so we can focus on the content
\input{include}
\subtitle{module 3.7.1: feature post-processing}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{document}
% generate title page
\input{include/titlepage}
\section[overview]{lecture overview}
\begin{frame}{introduction}{overview}
\begin{block}{corresponding textbook section}
%\href{http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6331120}{Chapter 3~---~Instantaneous Features}: pp.~63--66
sections~3.7.1--3.7.3
\end{block}
\begin{itemize}
\item \textbf{lecture content}
\begin{itemize}
\item derived features
\item feature aggregation
\item feature normalization
\end{itemize}
\bigskip
\item<2-> \textbf{learning objectives}
\begin{itemize}
\item discuss the advantages of specific derived features
\item summarize the principles of feature aggregation
\item list two forms of feature normalization and explain their usefulness
\end{itemize}
\end{itemize}
\inserticon{directions}
\end{frame}
\section[intro]{introduction}
\begin{frame}{feature post-processing}{introduction 1/2}
\begin{itemize}
\item extracting multiple instantaneous features leads to
\begin{itemize}
\item[$\rightarrow$] one feature vector per block, or
\item[$\rightarrow$] one feature matrix per audio file
\end{itemize}
\end{itemize}
\bigskip
\begin{eqnarray*}
\mat{V} &=& \left[\vec{v}(0)\; \vec{v}(1)\; \ldots\; \vec{v}(\mathcal{N}-1)\right] \nonumber\\
&=&
\left[
\begin{array}{cccc}
v_0(0) & v_0(1) & \ldots & v_0(\mathcal{N}-1)\\
v_1(0) & v_1(1) & \ldots & v_1(\mathcal{N}-1)\\
\vdots & \vdots & \ddots & \vdots \\
v_{\mathcal{F}-1}(0) & v_{\mathcal{F}-1}(1) & \ldots & v_{\mathcal{F}-1}(\mathcal{N}-1)\\
\end{array}
\right]
\end{eqnarray*}
\bigskip
\begin{footnotesize}
dimensions: $\mathcal{F}\times \mathcal{N}$ (number of features and number of blocks, resp.)
\end{footnotesize}
\end{frame}
\begin{frame}{feature post-processing}{introduction 2/2}
multiple options for feature matrix processing:
\begin{enumerate}
\item derive additional features
\item aggregate existing features (e.g., one feature vector per file)
\item ensure similar scale and distribution
\end{enumerate}
\end{frame}
\section[derived]{derived features}
\begin{frame}{feature post-processing}{examples of derived features}
%\begin{columns}
%\column{.5\linewidth}
\begin{itemize}
\item \textbf{diff}: use the change in value
\begin{equation*}
v_{j,\Delta}(n) = v_j(n) - v_j(n-1)
\end{equation*}
\smallskip
\item<2-> \textbf{smoothed}: remove high frequency content by low-pass filtering
\begin{itemize}
\item (anticausal) single-pole
\begin{equation*}
v_{j,\mathrm{LP}}(n) = (1-\alpha)\cdot v_j(n) - \alpha\cdot v_{j,\mathrm{LP}}(n-1)
\end{equation*}
\item moving average
\end{itemize}
\end{itemize}
%\column{.5\linewidth}
%\vspace{-10mm}
%\begin{figure}%
%\includegraphics{DerivedFeatures}%
%\end{figure}
%\end{columns}
%\addreference{matlab source: \href{https://github.com/alexanderlerch/ACA-Slides/blob/master/matlab/displayDerivedFeatures.m}{matlab/displayDerivedFeatures.m}}
\end{frame}
\section[normalization]{feature normalization}
\begin{frame}{feature post-processing}{feature normalization}
\vspace{-3mm}
\begin{columns}
\column{.5\linewidth}
\begin{itemize}
\item \textbf{reasons}
\begin{itemize}
\item features have different ranges and distributions
\item ensure that one feature does not have outsized impact
\end{itemize}
\smallskip
\item \textbf{z-score normalization}
\begin{equation*}\label{eq:featnorm_zscore}
v_{j,\mathrm{N}}(n) = \frac{v_j(n) - \mu_{v_j}}{\sigma_{v_j}}.
\end{equation*}
\item \textbf{min-max normalization}
\begin{equation*}\label{eq:featnorm_minmax}
v_{j,\mathrm{N}}(n) = \frac{v_j(n) - \min(v_j)}{\max(v_j) - \min(v_j)}.
\end{equation*}
\end{itemize}
\column{.5\linewidth}
\begin{block}{normalization}
The normalization constants $\mu_{v_j},\sigma_{v_j},\max(v_j),\min(v_j)$ have to be estimated from the \emph{{Training Set}}. The same (training) constants are then applied during inference. Extracting constants from the \emph{Test Set} is meaningless as the system has to infer with exactly the same parameters as during training.
\end{block}
\end{columns}
\end{frame}
\section[aggregation]{feature aggregation}
\begin{frame}{feature post-processing}{feature aggregation}
feature aggregation:\footnote{also compare \textit{pooling} operation in machine learning} compute \textit{summary features} from feature series $\Rightarrow$ \textbf{subfeatures}
\bigskip
\begin{itemize}
\item \textbf{reasons}
\begin{itemize}
\item only one feature vector required per file
\item data reduction
\item characteristics of distribution or change over time contain additional info
\end{itemize}
\smallskip
\item \textbf{examples}
\begin{itemize}
\item \textit{statistical descriptors}
\begin{itemize}
\item mean, median, max, standard deviation
\end{itemize}
\item \textit{hand crafted }
\begin{itemize}
\item anything that might be meaningful --- periodicity, slope, \ldots
\end{itemize}
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}{feature post-processing}{feature aggregation}
\begin{columns}
\column{.4\linewidth}
\begin{itemize}
\item could be for whole file or \textbf{texture window}:\\ split feature series in overlapping blocks of a few seconds length
\bigskip
\item<2-> could be \textbf{hierarchical} process:
\begin{enumerate}
\item compute subfeatures per window
\item compute subfeatures of subfeature series
\item (go to 1.)
\end{enumerate}
\end{itemize}
%\begin{footnotesize}
%\begin{align}
%%\mat{V} &=&
%\left[
%\begin{array}{ccccc}
%v_0(0) & v_0(1) & v_0(2) & v_0(3) & v_0(4) \\
%v_1(0) & v_1(1) & v_1(2) & v_1(3) & v_1(4) \\
%\vdots & \vdots & \vdots & \vdots & \vdots \\
%\undermat{\left[\begin{array}{c}
%\mu_0(0)\\
%\sigma_0(0)\\
%\mu_1(0)\\
%\sigma_1(0)\\
%\vdots\\
%\mu_{\mathcal{F}-1}(0)\\
%\sigma_{\mathcal{F}-1}(0)
%\end{array}\right] }{v_{\mathcal{F}-1}(0) & v_{\mathcal{F}-1}(1) & v_{\mathcal{F}-1}(2)& v_{\mathcal{F}-1}(3)} & v_{\mathcal{F}-1}(4)\\
%\end{array}
%\right]
%\end{align}
%\end{footnotesize}
%\vspace{8\baselineskip}
%\caption[Feature Aggregation]{Example aggregation with a texture window length of $4$ with the arithmetic mean $\mu$ and the standard deviation $\sigma$}
%\label{fig:feature_aggregation}
\column{.6\linewidth}
\figwithmatlab{TextureWindow}
\end{columns}
\end{frame}
\section{summary}
\begin{frame}{summary}{lecture content}
\begin{itemize}
\item \textbf{feature matrix should be processed to adapt to task and classifier}
\begin{itemize}
\item derive additional features
\item aggregate features
\item normalize features
\end{itemize}
\bigskip
\item \textbf{derived features}
\begin{itemize}
\item take existing features and ``create'' new ones
\end{itemize}
\bigskip
\item \textbf{feature normalization}
\begin{itemize}
\item avoid different value ranges might impacting classifier
\item handle different feature distributions
\end{itemize}
\bigskip
\item \textbf{aggregate features: subfeatures}
\begin{itemize}
\item combine blocks of features by computing, e.g., statistical features from them (mean, standard deviation, \ldots)
\item subfeature vector is used as classifier input or as intermediate feature series
\end{itemize}
\end{itemize}
\inserticon{summary}
\end{frame}
\end{document}