03 Prob

Probability and Information
Theory
Lecture slides for Chapter 3 of Deep Learning
www.deeplearningbook.org
Ian Goodfellow
2016-09-26
adapted by m.n. for CMPS 392
Probability
• Sample space Ω: set of all outcomes of a random experiment
• Set of events ℱ : collection of possible outcomes of an experiment.
• Probability measure: 𝑃: ℱ → ℝ
q Axioms of probability
o 𝑃 𝐴 ≥ 0 for all 𝐴 ∈ ℱ
o 𝑃 Ω =1
o If 𝐴! , 𝐴" , … are disjoint events then
𝑃 / 𝐴# = 0 𝐴#
# #
(Goodfellow 2016)
Random variable
• Consider an experiment in which we flip 10 coins, and we want to
know the number of coins that come up heads.
• Here, the elements of the sample space Ω are 10-length
sequences of heads and tails.
• For example, we might have
𝑤$ = < 𝐻, 𝐻, 𝑇, 𝐻, 𝑇, 𝐻, 𝐻, 𝑇, 𝑇, 𝑇 >
• However, in practice, we usually do not care about the probability
of obtaining any particular sequence of heads and tails.
• Instead we usually care about real-valued functions of outcomes,
such as
q the number of heads that appear among our 10 tosses,
q or the length of the longest run of tails.
• These functions, under some technical conditions, are known as
random variables: 𝑋: Ω → ℝ
(Goodfellow 2016)
Discrete vs. continuous
• Discrete random variable:
q 𝑃 𝑋 = 𝑘 = 𝑃 𝑤: 𝑋 𝑤 = 𝑘
• Continuous random variable:
q 𝑃 𝑎 ≤ 𝑋 ≤ 𝑏 = 𝑃 𝑤: 𝑎 ≤ 𝑋 𝑤 ≤ 𝑏
A cumulative distribution function (CDF):

𝑃(𝑋 ≤ 𝑥)
(Goodfellow 2016)
Probability Mass Function
(discrete variable)
Example: uniform distribution:
(Goodfellow 2016)
Probability Density Function
(continuous variable)
Example: uniform distribution:
The pdf at some point 𝑥 is not the

probability of 𝑥: 𝑝 𝑥 ≠ 𝑃 x = 𝑥
(Goodfellow 2016)
Computing Marginal
Probability with the Sum Rule
(Goodfellow 2016)
Conditional Probability
(Goodfellow 2016)
Chain Rule of Probability
(Goodfellow 2016)
Independence
(Goodfellow 2016)
Conditional Independence
(Goodfellow 2016)
Expectation
linearity of expectations:
(Goodfellow 2016)
Variance and Covariance
Covariance matrix:
(Goodfellow 2016)
Bernoulli Distribution
(Goodfellow 2016)
Gaussian Distribution
Parametrized by variance:
Parametrized by precision:
(Goodfellow 2016)
Gaussian Distribution
Figure 3.1
(Goodfellow 2016)
Multivariate Gaussian
Parametrized by covariance matrix:
Parametrized by precision matrix:
(Goodfellow 2016)
More Distributions
Exponential:
Laplace:
Dirac:
(Goodfellow 2016)
Empirical Distribution
(Goodfellow 2016)
Mixture Distributions
Gaussian mixture
with three
components
Figure 3.2 (Goodfellow 2016)

Logistic Sigmoid
Commonly used to parametrize Bernoulli distributions

(Goodfellow 2016)
Softplus Function
A
smoothed
version of
(Goodfellow 2016)
Useful Properties
%&'())
• 𝜎 𝑥 = %&' ) +%&' $
,
• 𝜎 𝑥 =𝜎 𝑥 1−𝜎 𝑥
,)
• 1 − 𝜎 𝑥 = 𝜎 −𝑥
• log 𝜎 𝑥 = −𝜁 −𝑥
,
• 𝜁 𝑥 =𝜎 𝑥
,)
-! )
• ∀𝑥 ∈ 0,1 , 𝜎 𝑥 = log !-)
• ∀𝑥 > 0, 𝜁 -! 𝑥 = log exp 𝑥 − 1
)
• 𝜁 𝑥 = ∫-. 𝜎 𝑦 𝑑𝑦
• 𝜁 𝑥 − 𝜁 −𝑥 = 𝑥
(Goodfellow 2016)
Bayes’ Rule
(Goodfellow 2016)
Bayes Rule
Prior
Posterior Likelihood
𝑃(𝐵|𝐴) 𝑃(𝐴)
𝑃(𝐴|𝐵) =
𝑃(𝐵)
Marginal
likelihood
computed by the total probability rule:
𝑃 𝐵 = $ 𝑃 𝐵 𝐴 = 𝑎 𝑃(𝐴 = 𝑎)
!
Bag-of-words Naïve Bayes:
¨ Features: Wi is the word at position i

¨ Called “bag-of-words” because model is insensitive to word order or
reordering:
¤ In a bag-of-words model, each position is identically distributed
𝑃 𝑆𝑝𝑎𝑚 𝑊! , 𝑊" , … , 𝑊# ∝ 𝑃 𝑊! 𝑆𝑝𝑎𝑚 𝑃 𝑊" 𝑆𝑝𝑎𝑚 … 𝑃 𝑊# 𝑆𝑝𝑎𝑚
¨ Start with a bunch of probabilities:
¤ Prior distribution 𝑃 𝑆𝑝𝑎𝑚 , 𝑃 𝐻𝑎𝑚

¤ and the likelihood probabilities (The CPT tables) 𝑃(𝑊" |𝑌)
¨ Use standard inference to compute the posterior probabilities

𝑃(𝑌|𝑊" … 𝑊# )
¨ We can use the normalization trick:
𝑃(𝐻𝑎𝑚|𝑊" … 𝑊# )+𝑃 𝑆𝑝𝑎𝑚 𝑊" … 𝑊# = 1
¨ Computing the log posterior (instead of the posterior) prevents numerical
errors
Example
Word P(w|spam) P(w|ham) Total spam Total ham

(log) (log)
(prior) 0.333 0.666 -1.1 -0.4
The 0.005 0.013 -5.27 -4.27
Year … … … …
2002
…
Bag of words exercise
¨ Spam messages: ¨ Size of vocabulary= ?

¤ Offer is secret ¨ P(SPAM) = ?
¤ Click secret link ¨ PML(“Secret”|SPAM) = ?
¤ Secret sports link
¨ PML(“Secret”|HAM)=?
¨ Ham messages: ¨ how many parameters to
¤ Play sports today represent the Naïve Bayes
¤ Went play sports
Network?
¤ Secret sports event
¨ P(SPAM|”Sports”)=?
¤ Sports is today
¨ P(SPAM| “Secret is secret”) = ?
¤ Sports costs money
¨ P(SPAM|”Today is secret”)=?
Change of Variables
Example of common mistake:
2
𝑦 = and 𝑥~𝑈(0,1)
3
𝑥 𝑦 = 1 if 𝑥 ∈ [0, ½]
𝑝4 𝑦 = 𝑝2 ⇒7
2 𝑦 = 0 𝑒𝑙𝑠𝑒𝑤ℎ𝑒𝑟𝑒
1
C 𝑝 𝑦 𝑑𝑦 = ‼
2
The right thing is: 𝑝4 𝑔(𝑥) 𝑑𝑦 = 𝑝2 𝑥 𝑑𝑥
𝜕𝑦
𝑝2 𝑥 = 𝑝4 𝑔 𝑥
𝜕𝑥
In higher dimensions:
(Goodfellow 2016)
Information theory
• Learning that an unlikely event has occurred is more
informative that learning that a likely event has occurred!
• Which statement has more information?
q “The sun rose this morning”
q “There was a solar eclipse this morning”
• Independent events should have additive information:

q Finding out that a tossed coin has come up heads
twice has two time more information that finding out
that a tossed coin has come up heads one time!
(Goodfellow 2016)
Self-Information
Log base e => unit is nats

Log base 2 => unit is bits or shannons
(Goodfellow 2016)
Entropy
Entropy:
• Entropy is a lower bound on the number of bits

needded on average to encode symbols drawn from
a distribution P.
• Distributions that are nearly deterministic have low
entropy
• Distributions that are nearly uniform have high
entropy
(Goodfellow 2016)
Entropy of a Bernoulli
Variable
𝐻 𝜙 = (𝜙 − 1) log 1 − 𝜙 − 𝜙log(𝜙)
Bernoulli parameter

Kullback-Leibler Divergence
KL divergence:
• KL-divergence is the extra amount of information

needed to send a message containing symbols
drawn from P, when we use a code designed to
minimize the length of messages containing
symbols drawn from Q
• KL-divergence is non-negative
• KL-divergence = 0 if P and Q are the same
distribution
(Goodfellow 2016)
KL-divergence
• It can be used as a distance measure between
distributions
• But it is not a true distance measure since it is not
symmetric:
q 𝐷56 (𝑃| 𝑄 ≠ 𝐷56 (𝑄| 𝑃
(Goodfellow 2016)
The KL Divergence is
Asymmetric
Mixture of two Gaussians for P, One Gaussian for Q

Cross-entropy
• 𝐻 𝑃, 𝑄 = 𝐻 𝑃 + 𝐷56 (𝑃| 𝑄
= −𝔼2~8 (log 𝑃 𝑥 ) + 𝔼2~8 (log 𝑃 𝑥 ) − 𝔼2~8 (log 𝑄 𝑥 )
= −𝔼2~8 (log 𝑄 𝑥 )
• Minimzing the cross entropy with respect to Q is
equivalent to minimize the KL divergence!
• Remark: usually we consider 0 log 0 = 0
(Goodfellow 2016)
Directed Model
Figure 3.7
𝑝 x = T 𝑝(x9 |𝑃𝑎𝒢 x9 )
9
(Goodfellow 2016)

03 Prob

Uploaded by

Copyright:

Available Formats

03 Prob

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

03 Prob

Uploaded by

Copyright:

Available Formats

Probability and Information

A cumulative distribution function (CDF):

Example: uniform distribution:

Example: uniform distribution:

The pdf at some point 𝑥 is not the

Parametrized by covariance matrix:

Parametrized by precision matrix:

Figure 3.2 (Goodfellow 2016)

Commonly used to parametrize Bernoulli distributions

computed by the total probability rule:

¨ Features: Wi is the word at position i

¤ Prior distribution 𝑃 𝑆𝑝𝑎𝑚 , 𝑃 𝐻𝑎𝑚

¨ Use standard inference to compute the posterior probabilities

Word P(w|spam) P(w|ham) Total spam Total ham

¨ Spam messages: ¨ Size of vocabulary= ?

• Independent events should have additive information:

Log base e => unit is nats

• Entropy is a lower bound on the number of bits

Figure 3.5 (Goodfellow 2016)

• KL-divergence is the extra amount of information

Figure 3.6 (Goodfellow 2016)

• Remark: usually we consider 0 log 0 = 0

You might also like