Mercurial > pylearn

annotate doc/v2_planning/sampler.txt @ 1067:4f287324a5ad

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Merged
author Olivier Delalleau <delallea@iro>
date Fri, 10 Sep 2010 10:00:49 -0400
parents 875d53754bd0
children 0e12ea6ba661
rev   line source
1027
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1018
diff changeset
1
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1018
diff changeset
2 Inference / Sampling committee: JB, GD, AC
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1018
diff changeset
3
1018
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
4 OVERVIEW
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
5 ========
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
6
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
7 Before we start defining what a sampler is and how it should be defined in
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
8 pylearn, we should first know what we're up against.
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
9
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
10 The workflow I have in mind is the following:
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
11 1. identify the most popular sampling algorithms in the litterature
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
12 2. get up to speed with methods we're not familiar with
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
13 3. identify common usage patterns, properties of the algorithm, etc.
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
14 4. decide on an API / best way to implement them
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
15 5. prioritize the algorithms
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
16 6. code away
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
17
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
18 1.BACKGROUND
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
19 =============
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
20
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
21 This section should provide a brief overview of what exists in the litterature.
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
22 We should make sure to have a decent understanding of all of these (not everyone
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
23 has to be experts though), so that we can *intelligently* design our sampler
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
24 interface based on common usage patterns, properties, etc.
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
25
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
26 Sampling from basic distributions
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
27 * already supported: uniform, normal, binomial, multinomial
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
28 * wish list: beta, poisson, others ?
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
29
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
30 List of sampling algorithms:
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
31 * inversion sampling
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
32 * rejection sampling
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
33 * importance sampling
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
34 * Markov Chain Monte Carlo
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
35 * Gibbs sampling
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
36 * Metropolis Hastings
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
37 * Slice Sampling
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
38 * Annealing
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
39 * Parallel Tempering, Tempered Transitions, Simulated Tempering
790376d986a3 initial document for sampling
gdesjardins
parents:
diff changeset
40 * Nested Sampling (?)
1027
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1018
diff changeset
41 * Hamiltonian Monte Carlo --> or is it Hybrid Monte Carlo?
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1018
diff changeset
42
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1018
diff changeset
43 3. USAGE PATTERNS
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1018
diff changeset
44 =================
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1018
diff changeset
45
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1018
diff changeset
46 * MCMC methods have a usage pattern that is quite different from the kind of univariate sampling methods
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1018
diff changeset
47 needed for nice-and-easy parametric families.
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1018
diff changeset
48
a1b6ccd5b6dc few comments added
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents: 1018
diff changeset
49