-
Notifications
You must be signed in to change notification settings - Fork 81
/
Copy pathneural_networks_unsupervised.html
382 lines (350 loc) · 22.2 KB
/
neural_networks_unsupervised.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
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
<!DOCTYPE html>
<!-- data-theme below is forced to be "light" but should be changed if we use pydata-theme-sphinx in the future -->
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" data-content_root="../" data-theme="light"> <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" data-content_root="../" data-theme="light"> <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1" />
<meta property="og:title" content="2.9. Neural network models (unsupervised)" />
<meta property="og:type" content="website" />
<meta property="og:url" content="https://scikit-learn/stable/modules/neural_networks_unsupervised.html" />
<meta property="og:site_name" content="scikit-learn" />
<meta property="og:description" content="Restricted Boltzmann machines: Restricted Boltzmann machines (RBM) are unsupervised nonlinear feature learners based on a probabilistic model. The features extracted by an RBM or a hierarchy of RBM..." />
<meta property="og:image" content="https://scikit-learn/stable/_images/sphx_glr_plot_rbm_logistic_classification_001.png" />
<meta property="og:image:alt" content="scikit-learn" />
<meta name="description" content="Restricted Boltzmann machines: Restricted Boltzmann machines (RBM) are unsupervised nonlinear feature learners based on a probabilistic model. The features extracted by an RBM or a hierarchy of RBM..." />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>2.9. Neural network models (unsupervised) — scikit-learn 1.5.dev0 documentation</title>
<link rel="canonical" href="http://scikit-learn.org/stable/modules/neural_networks_unsupervised.html" />
<link rel="shortcut icon" href="../_static/favicon.ico"/>
<link rel="stylesheet" href="../_static/css/vendor/bootstrap.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<link rel="stylesheet" href="../_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="../_static/plot_directive.css" type="text/css" />
<link rel="stylesheet" href="../https://fonts.googleapis.com/css?family=Vibur" type="text/css" />
<link rel="stylesheet" href="../_static/jupyterlite_sphinx.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-binder.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-dataframe.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-rendered-html.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<script id="documentation_options" data-url_root="../" src="../_static/documentation_options.js"></script>
<script src="../_static/js/vendor/jquery-3.6.3.slim.min.js"></script>
<script src="../_static/js/details-permalink.js"></script>
</head>
<body>
<nav id="navbar" class="sk-docs-navbar navbar navbar-expand-md navbar-light bg-light py-0">
<div class="container-fluid sk-docs-container px-0">
<a class="navbar-brand py-0" href="../index.html">
<img
class="sk-brand-img"
src="../_static/scikit-learn-logo-small.png"
alt="logo"/>
</a>
<button
id="sk-navbar-toggler"
class="navbar-toggler"
type="button"
data-toggle="collapse"
data-target="#navbarSupportedContent"
aria-controls="navbarSupportedContent"
aria-expanded="false"
aria-label="Toggle navigation"
>
<span class="navbar-toggler-icon"></span>
</button>
<div class="sk-navbar-collapse collapse navbar-collapse" id="navbarSupportedContent">
<ul class="navbar-nav mr-auto">
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../install.html">Install</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../user_guide.html">User Guide</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" href="classes.html">API</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../auto_examples/index.html">Examples</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" target="_blank" rel="noopener noreferrer" href="https://blog.scikit-learn.org/">Community</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../getting_started.html" >Getting Started</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../tutorial/index.html" >Tutorial</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../whats_new/v1.5.html" >What's new</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../glossary.html" >Glossary</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../developers/index.html" >Development</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../faq.html" >FAQ</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../support.html" >Support</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../related_projects.html" >Related packages</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../roadmap.html" >Roadmap</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../governance.html" >Governance</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../about.html" >About us</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="https://github.com/scikit-learn/scikit-learn" >GitHub</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="https://scikit-learn.org/dev/versions.html" >Other Versions and Download</a>
</li>
<li class="nav-item dropdown nav-more-item-dropdown">
<a class="sk-nav-link nav-link dropdown-toggle" href="#" id="navbarDropdown" role="button" data-toggle="dropdown" aria-haspopup="true" aria-expanded="false">More</a>
<div class="dropdown-menu" aria-labelledby="navbarDropdown">
<a class="sk-nav-dropdown-item dropdown-item" href="../getting_started.html" >Getting Started</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../tutorial/index.html" >Tutorial</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../whats_new/v1.5.html" >What's new</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../glossary.html" >Glossary</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../developers/index.html" >Development</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../faq.html" >FAQ</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../support.html" >Support</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../related_projects.html" >Related packages</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../roadmap.html" >Roadmap</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../governance.html" >Governance</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../about.html" >About us</a>
<a class="sk-nav-dropdown-item dropdown-item" href="https://github.com/scikit-learn/scikit-learn" >GitHub</a>
<a class="sk-nav-dropdown-item dropdown-item" href="https://scikit-learn.org/dev/versions.html" >Other Versions and Download</a>
</div>
</li>
</ul>
<div id="searchbox" role="search">
<div class="searchformwrapper">
<form class="search" action="../search.html" method="get">
<input class="sk-search-text-input" type="text" name="q" aria-labelledby="searchlabel" />
<input class="sk-search-text-btn" type="submit" value="Go" />
</form>
</div>
</div>
</div>
</div>
</nav>
<div class="d-flex" id="sk-doc-wrapper">
<input type="checkbox" name="sk-toggle-checkbox" id="sk-toggle-checkbox">
<label id="sk-sidemenu-toggle" class="sk-btn-toggle-toc btn sk-btn-primary" for="sk-toggle-checkbox">Toggle Menu</label>
<div id="sk-sidebar-wrapper" class="border-right">
<div class="sk-sidebar-toc-wrapper">
<div class="btn-group w-100 mb-2" role="group" aria-label="rellinks">
<a href="density.html" role="button" class="btn sk-btn-rellink py-1" sk-rellink-tooltip="2.8. Density Estimation">Prev</a><a href="../unsupervised_learning.html" role="button" class="btn sk-btn-rellink py-1" sk-rellink-tooltip="2. Unsupervised learning">Up</a>
<a href="../model_selection.html" role="button" class="btn sk-btn-rellink py-1" sk-rellink-tooltip="3. Model selection and evaluation">Next</a>
</div>
<div class="alert alert-danger p-1 mb-2" role="alert">
<p class="text-center mb-0">
<strong>scikit-learn 1.5.dev0</strong><br/>
<a href="http://scikit-learn.org/dev/versions.html">Other versions</a>
</p>
</div>
<div class="alert alert-warning p-1 mb-2" role="alert">
<p class="text-center mb-0">
Please <a class="font-weight-bold" href="../about.html#citing-scikit-learn"><string>cite us</string></a> if you use the software.
</p>
</div>
<div class="sk-sidebar-toc">
<ul>
<li><a class="reference internal" href="#">2.9. Neural network models (unsupervised)</a><ul>
<li><a class="reference internal" href="#restricted-boltzmann-machines">2.9.1. Restricted Boltzmann machines</a><ul>
<li><a class="reference internal" href="#graphical-model-and-parametrization">2.9.1.1. Graphical model and parametrization</a></li>
<li><a class="reference internal" href="#bernoulli-restricted-boltzmann-machines">2.9.1.2. Bernoulli Restricted Boltzmann machines</a></li>
<li><a class="reference internal" href="#stochastic-maximum-likelihood-learning">2.9.1.3. Stochastic Maximum Likelihood learning</a></li>
</ul>
</li>
</ul>
</li>
</ul>
</div>
</div>
</div>
<div id="sk-page-content-wrapper">
<div class="sk-page-content container-fluid body px-md-3" role="main">
<section id="neural-network-models-unsupervised">
<span id="neural-networks-unsupervised"></span><h1><span class="section-number">2.9. </span>Neural network models (unsupervised)<a class="headerlink" href="#neural-network-models-unsupervised" title="Link to this heading">¶</a></h1>
<section id="restricted-boltzmann-machines">
<span id="rbm"></span><h2><span class="section-number">2.9.1. </span>Restricted Boltzmann machines<a class="headerlink" href="#restricted-boltzmann-machines" title="Link to this heading">¶</a></h2>
<p>Restricted Boltzmann machines (RBM) are unsupervised nonlinear feature learners
based on a probabilistic model. The features extracted by an RBM or a hierarchy
of RBMs often give good results when fed into a linear classifier such as a
linear SVM or a perceptron.</p>
<p>The model makes assumptions regarding the distribution of inputs. At the moment,
scikit-learn only provides <a class="reference internal" href="generated/sklearn.neural_network.BernoulliRBM.html#sklearn.neural_network.BernoulliRBM" title="sklearn.neural_network.BernoulliRBM"><code class="xref py py-class docutils literal notranslate"><span class="pre">BernoulliRBM</span></code></a>, which assumes the inputs are
either binary values or values between 0 and 1, each encoding the probability
that the specific feature would be turned on.</p>
<p>The RBM tries to maximize the likelihood of the data using a particular
graphical model. The parameter learning algorithm used (<a class="reference internal" href="#sml"><span class="std std-ref">Stochastic
Maximum Likelihood</span></a>) prevents the representations from straying far
from the input data, which makes them capture interesting regularities, but
makes the model less useful for small datasets, and usually not useful for
density estimation.</p>
<p>The method gained popularity for initializing deep neural networks with the
weights of independent RBMs. This method is known as unsupervised pre-training.</p>
<figure class="align-center">
<a class="reference external image-reference" href="../auto_examples/neural_networks/plot_rbm_logistic_classification.html"><img alt="../_images/sphx_glr_plot_rbm_logistic_classification_001.png" src="../_images/sphx_glr_plot_rbm_logistic_classification_001.png" style="width: 420.0px; height: 400.0px;" /></a>
</figure>
<aside class="topic">
<p class="topic-title">Examples:</p>
<ul class="simple">
<li><p><a class="reference internal" href="../auto_examples/neural_networks/plot_rbm_logistic_classification.html#sphx-glr-auto-examples-neural-networks-plot-rbm-logistic-classification-py"><span class="std std-ref">Restricted Boltzmann Machine features for digit classification</span></a></p></li>
</ul>
</aside>
<section id="graphical-model-and-parametrization">
<h3><span class="section-number">2.9.1.1. </span>Graphical model and parametrization<a class="headerlink" href="#graphical-model-and-parametrization" title="Link to this heading">¶</a></h3>
<p>The graphical model of an RBM is a fully-connected bipartite graph.</p>
<img alt="../_images/rbm_graph.png" class="align-center" src="../_images/rbm_graph.png" />
<p>The nodes are random variables whose states depend on the state of the other
nodes they are connected to. The model is therefore parameterized by the
weights of the connections, as well as one intercept (bias) term for each
visible and hidden unit, omitted from the image for simplicity.</p>
<p>The energy function measures the quality of a joint assignment:</p>
<div class="math notranslate nohighlight">
\[E(\mathbf{v}, \mathbf{h}) = -\sum_i \sum_j w_{ij}v_ih_j - \sum_i b_iv_i
- \sum_j c_jh_j\]</div>
<p>In the formula above, <span class="math notranslate nohighlight">\(\mathbf{b}\)</span> and <span class="math notranslate nohighlight">\(\mathbf{c}\)</span> are the
intercept vectors for the visible and hidden layers, respectively. The
joint probability of the model is defined in terms of the energy:</p>
<div class="math notranslate nohighlight">
\[P(\mathbf{v}, \mathbf{h}) = \frac{e^{-E(\mathbf{v}, \mathbf{h})}}{Z}\]</div>
<p>The word <em>restricted</em> refers to the bipartite structure of the model, which
prohibits direct interaction between hidden units, or between visible units.
This means that the following conditional independencies are assumed:</p>
<div class="math notranslate nohighlight">
\[\begin{split}h_i \bot h_j | \mathbf{v} \\
v_i \bot v_j | \mathbf{h}\end{split}\]</div>
<p>The bipartite structure allows for the use of efficient block Gibbs sampling for
inference.</p>
</section>
<section id="bernoulli-restricted-boltzmann-machines">
<h3><span class="section-number">2.9.1.2. </span>Bernoulli Restricted Boltzmann machines<a class="headerlink" href="#bernoulli-restricted-boltzmann-machines" title="Link to this heading">¶</a></h3>
<p>In the <a class="reference internal" href="generated/sklearn.neural_network.BernoulliRBM.html#sklearn.neural_network.BernoulliRBM" title="sklearn.neural_network.BernoulliRBM"><code class="xref py py-class docutils literal notranslate"><span class="pre">BernoulliRBM</span></code></a>, all units are binary stochastic units. This
means that the input data should either be binary, or real-valued between 0 and
1 signifying the probability that the visible unit would turn on or off. This
is a good model for character recognition, where the interest is on which
pixels are active and which aren’t. For images of natural scenes it no longer
fits because of background, depth and the tendency of neighbouring pixels to
take the same values.</p>
<p>The conditional probability distribution of each unit is given by the
logistic sigmoid activation function of the input it receives:</p>
<div class="math notranslate nohighlight">
\[\begin{split}P(v_i=1|\mathbf{h}) = \sigma(\sum_j w_{ij}h_j + b_i) \\
P(h_i=1|\mathbf{v}) = \sigma(\sum_i w_{ij}v_i + c_j)\end{split}\]</div>
<p>where <span class="math notranslate nohighlight">\(\sigma\)</span> is the logistic sigmoid function:</p>
<div class="math notranslate nohighlight">
\[\sigma(x) = \frac{1}{1 + e^{-x}}\]</div>
</section>
<section id="stochastic-maximum-likelihood-learning">
<span id="sml"></span><h3><span class="section-number">2.9.1.3. </span>Stochastic Maximum Likelihood learning<a class="headerlink" href="#stochastic-maximum-likelihood-learning" title="Link to this heading">¶</a></h3>
<p>The training algorithm implemented in <a class="reference internal" href="generated/sklearn.neural_network.BernoulliRBM.html#sklearn.neural_network.BernoulliRBM" title="sklearn.neural_network.BernoulliRBM"><code class="xref py py-class docutils literal notranslate"><span class="pre">BernoulliRBM</span></code></a> is known as
Stochastic Maximum Likelihood (SML) or Persistent Contrastive Divergence
(PCD). Optimizing maximum likelihood directly is infeasible because of
the form of the data likelihood:</p>
<div class="math notranslate nohighlight">
\[\log P(v) = \log \sum_h e^{-E(v, h)} - \log \sum_{x, y} e^{-E(x, y)}\]</div>
<p>For simplicity the equation above is written for a single training example.
The gradient with respect to the weights is formed of two terms corresponding to
the ones above. They are usually known as the positive gradient and the negative
gradient, because of their respective signs. In this implementation, the
gradients are estimated over mini-batches of samples.</p>
<p>In maximizing the log-likelihood, the positive gradient makes the model prefer
hidden states that are compatible with the observed training data. Because of
the bipartite structure of RBMs, it can be computed efficiently. The
negative gradient, however, is intractable. Its goal is to lower the energy of
joint states that the model prefers, therefore making it stay true to the data.
It can be approximated by Markov chain Monte Carlo using block Gibbs sampling by
iteratively sampling each of <span class="math notranslate nohighlight">\(v\)</span> and <span class="math notranslate nohighlight">\(h\)</span> given the other, until the
chain mixes. Samples generated in this way are sometimes referred as fantasy
particles. This is inefficient and it is difficult to determine whether the
Markov chain mixes.</p>
<p>The Contrastive Divergence method suggests to stop the chain after a small
number of iterations, <span class="math notranslate nohighlight">\(k\)</span>, usually even 1. This method is fast and has
low variance, but the samples are far from the model distribution.</p>
<p>Persistent Contrastive Divergence addresses this. Instead of starting a new
chain each time the gradient is needed, and performing only one Gibbs sampling
step, in PCD we keep a number of chains (fantasy particles) that are updated
<span class="math notranslate nohighlight">\(k\)</span> Gibbs steps after each weight update. This allows the particles to
explore the space more thoroughly.</p>
<aside class="topic">
<p class="topic-title">References:</p>
<ul class="simple">
<li><p><a class="reference external" href="https://www.cs.toronto.edu/~hinton/absps/fastnc.pdf">“A fast learning algorithm for deep belief nets”</a>
G. Hinton, S. Osindero, Y.-W. Teh, 2006</p></li>
<li><p><a class="reference external" href="https://www.cs.toronto.edu/~tijmen/pcd/pcd.pdf">“Training Restricted Boltzmann Machines using Approximations to
the Likelihood Gradient”</a>
T. Tieleman, 2008</p></li>
</ul>
</aside>
</section>
</section>
</section>
</div>
<div class="container">
<footer class="sk-content-footer">
© 2007 - 2023, scikit-learn developers (BSD License).
<a href="../_sources/modules/neural_networks_unsupervised.rst.txt" rel="nofollow">Show this page source</a>
</footer>
</div>
</div>
</div>
<script src="../_static/js/vendor/bootstrap.min.js"></script>
<script>
window.ga=window.ga||function(){(ga.q=ga.q||[]).push(arguments)};ga.l=+new Date;
ga('create', 'UA-22606712-2', 'auto');
ga('set', 'anonymizeIp', true);
ga('send', 'pageview');
</script>
<script async src='https://www.google-analytics.com/analytics.js'></script>
<script defer data-domain="scikit-learn.org" src="https://views.scientific-python.org/js/script.js">
</script>
<script src="../_static/clipboard.min.js"></script>
<script src="../_static/copybutton.js"></script>
<script>
$(document).ready(function() {
/* Add a [>>>] button on the top-right corner of code samples to hide
* the >>> and ... prompts and the output and thus make the code
* copyable. */
var div = $('.highlight-python .highlight,' +
'.highlight-python3 .highlight,' +
'.highlight-pycon .highlight,' +
'.highlight-default .highlight')
var pre = div.find('pre');
// get the styles from the current theme
pre.parent().parent().css('position', 'relative');
// create and add the button to all the code blocks that contain >>>
div.each(function(index) {
var jthis = $(this);
// tracebacks (.gt) contain bare text elements that need to be
// wrapped in a span to work with .nextUntil() (see later)
jthis.find('pre:has(.gt)').contents().filter(function() {
return ((this.nodeType == 3) && (this.data.trim().length > 0));
}).wrap('<span>');
});
/*** Add permalink buttons next to glossary terms ***/
$('dl.glossary > dt[id]').append(function() {
return ('<a class="headerlink" href="#' +
this.getAttribute('id') +
'" title="Permalink to this term">¶</a>');
});
});
</script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-chtml.js"></script>
<script src="https://scikit-learn.org/versionwarning.js"></script>
</body>
</html>