Inference
Chris Gregg
CS109, Stanford University
Autumn 2026
Today: Stanford Eye Test
Chris Piech, CS109
Learning Goals
Chris Piech, CS109
Four Prototypical Trajectories
Review
Chris Piech, CS109
Probabilistic Models
| Roommates | 2RoomDbl | Shared Partner | Single | |
Frosh | 0.30 | 0.07 | 0.00 | 0.00 | 0.37 |
Soph | 0.12 | 0.17 | 0.00 | 0.03 | 0.32 |
Junior | 0.04 | 0.01 | 0.00 | 0.10 | 0.15 |
Senior | 0.01 | 0.01 | 0.02 | 0.01 | 0.05 |
5+ | 0.02 | 0.00 | 0.05 | 0.04 | 0.11 |
| 0.49 | 0.26 | 0.07 | 0.18 | 1.00 |
Probability
Joint
Marginals
Chris Piech, CS109
Today: Inference
Inference noun
Updating one’s belief about a random variable (or multiple) based on conditional knowledge regarding another random variable (or multiple) in a probabilistic model.
TLDR: conditional probability with random variables.
Chris Piech, CS109
7
Chris Piech, CS109
8
Baby Delivery
Its 17 days before the due date and still no baby
Chris Piech, CS109
9
Baby Delivery
Its 17 days before the due date and still no baby
Chris Piech, CS109
10
Baby Delivery
Its 17 days before the due date and still no baby
Probability baby is born on due date increases by 12%
Chris Piech, CS109
Big idea for re-calculating PMF
11
We have a dictionary of probabilities mapped to days, based on the due date being day 0. We have probability values from -50 days before the due date to 30 days after the due date. The peak is after the due date because the due date is an estimate.
Day,Probability
-50,0.00009390
-49,0.00011268
-48,0.00014085
...
-3,0.02529675
-2,0.02633363
-1,0.02708620
0,0.02769864
1,0.02845020
2,0.02910833
3,0.02967256
...
28,0.00093990
29,0.00056337
30,0.00028168
Chris Piech, CS109
Big idea for re-calculating PMF
12
Once we reach a day (say NB17, or “no baby 17 days prior to the due date”), we have the information that we have not yet seen a baby 17 days prior (and, of course, all the days before that).
Day,Probability
-50,0.00009390
-49,0.00011268
-48,0.00014085
...
-3,0.02529675
-2,0.02633363
-1,0.02708620
0,0.02769864
1,0.02845020
2,0.02910833
3,0.02967256
...
28,0.00093990
29,0.00056337
30,0.00028168
Chris Piech, CS109
Big idea for re-calculating PMF
13
We need to apply Bayes’ Theorem for all the values (-50, 30) to get an updated PMF. We’ll do this in a loop. D is the delivery day relative to the due date with d being the day the baby is actually born in our model, and NB is the event that the baby is not born nb days before the due date.)
Chris Piech, CS109
Big idea for re-calculating PMF
14
Notably: is 0 for any day, d, up to and including nb, and 1 for any day after nb. That is, the probability that we have not yet seen the baby by day nb is zero if the baby was born on or before nb, and one if the baby is born after nb.
Chris Piech, CS109
To the code!
15
Chris Piech, CS109
Update Belief PMF
Before observation
After
Observation
Chris Piech, CS109
Bayesian Carbon Dating: Inference Overview
17
Let A be how many years old the sample is (A = 100 means the sample is 100 years old)
Let M be the observed amount of C14 left in the sample
Such that
Chris Piech, CS109
Probability of Having 900 Remain
18
There were originally 1000 C14 molecules.
Each molecule remains independently with equal probability pi
What is the probability that 900 remain?
Let T be the time to decay for any one molecule
Each molecules’ time to live is exponential with λ = 1/8267
Chris Piech, CS109
Inference as Code
19
def update_belief(m = 900):
"""
Returns a dictionary A, where A[i] contains the
corresponding probability, P(A = i| M = 900).
m is the number of C14 molecules remaining and i
is age in years. i is in the range 100 to 10000
"""
pr_A = {}
n_years = 9901
for i in range(100,10000+1):
prior = 1 / n_years # P(A = i)
likelihood = calc_likelihood(m, i) # P(M=m | A=i)
pr_A[i] = prior * likelihood
# implicitly computes the normalization constant
normalize(pr_A)
return pr_A
Chris Piech, CS109
Probability of Having 900 Remain
20
def calc_likelihood(m = 900, age):
"""
Computes P(M = m | A = age), the probability of
having m molecules left given the sample is age
years old. Uses the exponential decay of C14
"""
n_original = 1000
p_remain = math.exp(-age/C14_MEAN_LIFE)
return stats.binom.pmf(m, n_original, p_remain)
Chris Piech, CS109
Probability of Having 900 Remain
21
def calc_likelihood(m = 900, age):
"""
Computes P(M = m | A = age), the probability of
having m molecules left given the sample is age
years old. Uses the exponential decay of C14
"""
n_original = 1000 + delta_start(age)
p_remain = math.exp(-age/C14_MEAN_LIFE)
return stats.binom.pmf(m, n_original, p_remain)
Chris Piech, CS109
Four Prototypical Trajectories
Why is inference hard?
Bayes with variables
Can mix discrete and continuous
Normalization term is trippy at first
Likelihood term can be complex
Chris Piech, CS109
Bayes + Continuous Random Variables
Let X be a continuous random variable
Let N be a discrete random variable
Chris Piech, CS109
Four Prototypical Trajectories
End Review
Chris Piech, CS109
Today: Compare and Contrast Many Examples
25
Age from C14
Updated Delivery Prob
Age from Name
Hidden Chambers
Stanford Eye Test
Updating Lidar Belief
Chris Piech, CS109
Name 2 Age
26
Birth year
Name
Size of Dataset
Number of people with this name born this year
Chris Piech, CS109
Name 2 Age
27
Birth year
Name
Size of Dataset
Number of people with this name born this year
How can you use the joint to compute this?
Chris Piech, CS109
Name 2 Age
28
Number of people with this name born this year
How can you use the joint to compute this?
I.e., loop over all the possible birth years, and what is the probability they were born in that year and their name was Michael?
Chris Piech, CS109
Name 2 Age
29
How would you compute the entire PMF using code?
Chris Piech, CS109
30
Chris Piech, CS109
Compare and Contrast Code: (1) Bayesian Carbon Dating
31
After
Observation
Chris Piech, CS109
Compare and Contrast Code: (2) Baby Delivery
32
For each value d:
Chris Piech, CS109
Compare and Contrast Code: (2) Baby Delivery
33
Its 19 days until the due date and no baby
For each value d:
Chris Piech, CS109
Compare and Contrast Code: (3) Name To Age
34
Chris Piech, CS109
35
def update_belief_carbon_dating(m = 900):
# pr_A[i] is P(Age = i| m = 900).
pr_A = {}
for i in range(100,10000+1):
prior = 1 / n_years # P(A = i)
likelihood = calc_likelihood(m, i) #P(M=m|A=i)
pr_A[i] = likelihood * prior
# implicitly computes the normalization constant
normalize(pr_A)
return pr_A
def update_belief_baby(prior, today = 10):
# pr_D[i] is P(D = i| No Baby Yet).
pr_D = {}
for i in range(-50,25):
# P(NoBaby | D = i)
likelihood = 0 if i < today else 1
pr_D[i] = likelihood * prior[i]
# implicitly computes the LOTP
normalize(pr_D)
return pr_D
def update_belief_name_to_age(name = 'Laura'):
# pr_age[i] is P(Age = i| name).
# prob_name_and_age is just a counting from the US
# Social Security database.
pr_age = {}
for i in range(10,110):
pr_age[i] = calc_prob_name_and_age(name, i)
# implicitly computes the normalization constant
normalize(pr_age)
return pr_age
What do you notice is the same. What is different?
Chris Piech, CS109
Normalize in Python
36
# list normalization
def normalize_list(data_list):
total_sum = np.sum(data_list)
return np.array(data_list) / total_sum
>>> norm = normalize_list([10, 20, 30, 40])
>>> np.sum(norm) # 1.0, always (within floating point error)
# dictionary normalization
def normalize_dict(data_dict):
total_sum = sum(data_dict.values())
normalized = {}
for key, value in data_dict.items():
normalized[key] = value / total_sum
return normalized
>>> norm = normalize_dict({'a': 100, 'b': 200, 'c': 300})
>>> np.sum(norm.values()) # 1.0, always (within floating point error)
return
[0.1 0.2 0.3 0.4]
return
{'a': 0.166, 'b': 0.333, 'c': 0.5}
Chris Piech, CS109
37
Chris Piech, CS109
Four Prototypical Trajectories
Are you ready
For Hidden Chambers???
Chris Piech, CS109
Hidden Pyramid Chambers with Poisson + Bayes
Chris Piech, CS109
Basics of Muography
40
Beer Lambert Law
Rate of muons depends on x, amount of limestone
Chris Piech, CS109
41
Beer Lambert Law
Rate of muons depends on x, amount of limestone
Chris Piech, CS109
42
Chris Piech, CS109
43
Beer Lambert Law
Rate of muons depends on x, amount of limestone
Chris Piech, CS109
44
Chris Piech, CS109
45
Chris Piech, CS109
Four Prototypical Trajectories
Beyond Inference:
What hidden chambers make all the muon readings most likely?
Chris Piech, CS109
Four Prototypical Trajectories
Are you ready
For Stanford Acuity Test???
Chris Piech, CS109
A Better Eye Test
48
https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(21)02149-8/fulltext
https://www.science.org/content/article/eye-robot-artificial-intelligence-dramatically-improves-accuracy-classic-eye-exam
https://ojs.aaai.org/index.php/AAAI/article/view/5384/5240
Chris Piech, CS109
Prior Belief in Ability to See (Random Var A)
49
Can’t see
Standard vision
Probability Mass
Chris Piech, CS109
PMF is Actually Stored as a Dictionary
50
a | P(A=a) | | a | P(A=a) | | | | a | P(A=a) |
0.00 | 0.00198 | | 0.20 | 0.00372 | | | | 0.80 | 0.01684 |
0.01 | 0.00205 | | 0.21 | 0.00384 | | | | 0.81 | 0.01708 |
0.02 | 0.00211 | | 0.22 | 0.00396 | | | | 0.82 | 0.01731 |
0.03 | 0.00218 | | 0.23 | 0.00408 | | | | 0.83 | 0.01753 |
0.04 | 0.00225 | | 0.24 | 0.00421 | | | | 0.84 | 0.01774 |
0.05 | 0.00233 | | 0.25 | 0.00434 | | | | 0.85 | 0.01795 |
0.06 | 0.0024 | | 0.26 | 0.00447 | | | | 0.86 | 0.01814 |
0.07 | 0.00248 | | 0.27 | 0.00461 | | | | 0.87 | 0.01832 |
0.08 | 0.00256 | | 0.28 | 0.00475 | | … | | 0.88 | 0.01848 |
0.09 | 0.00264 | | 0.29 | 0.00489 | | | | 0.89 | 0.01864 |
0.10 | 0.00273 | | 0.30 | 0.00504 | | | | 0.90 | 0.01877 |
0.11 | 0.00281 | | 0.31 | 0.00519 | | | | 0.91 | 0.0189 |
0.12 | 0.0029 | | 0.32 | 0.00535 | | | | 0.92 | 0.019 |
0.13 | 0.00299 | | 0.33 | 0.00551 | | | | 0.93 | 0.01909 |
0.14 | 0.00309 | | 0.34 | 0.00567 | | | | 0.94 | 0.01916 |
0.15 | 0.00319 | | 0.35 | 0.00584 | | | | 0.95 | 0.01921 |
0.16 | 0.00329 | | 0.36 | 0.00601 | | | | 0.96 | 0.01924 |
0.17 | 0.00339 | | 0.37 | 0.00619 | | | | 0.97 | 0.01925 |
0.18 | 0.0035 | | 0.38 | 0.00637 | | | | 0.98 | 0.01924 |
0.19 | 0.00361 | | 0.39 | 0.00655 | | | | 0.99 | 0.01921 |
def main():
belief = get_prior_belief()
…
Chris Piech, CS109
Prior Belief in Ability to See (Random Var A)
51
belief = get_prior_belief()
As a graph
As a dictionary
Chris Piech, CS109
The Patient is Shown One Letter and They Get it Wrong
52
User: wrong
Observation Y = 0
Font size 0.7
Chris Piech, CS109
Number or Dictionary?
53
Chris Piech, CS109
Number or Dictionary?
54
belief[a] = 0.001
belief
Chris Piech, CS109
Today: I am going to simplify the units of vision
55
Normally doctors measure ability to see in logarithmic units. To make today’s demo easier to understand
I have translated both onto a
[0, 1] scale.
Where 0 means can’t see and 1.0 is standard vision
Chris Piech, CS109
Posterior Belief in Ability to See (Random Var A)
56
Can’t see
Standard vision
Probability Mass
Chris Piech, CS109
Bayes with Random Variables
57
User: wrong
1
2
3
Observation Y = 0
(At font size s1)
Chris Piech, CS109
Inference on a non-bernoulli random variable
58
# RV bayes as code
def update(belief, obs):
for a in support:
prior_a = belief[a]
likelihood = calc_likelihood(a, obs)
belief[a] = prior_a * likelihood
normalize(belief)
likelihood
In plain English: run bayes for each value of a
Chris Piech, CS109
Normalize???
59
# RV bayes as code
def update(belief, obs):
for a in support:
prior_a = belief[a]
likelihood = calc_likelihood(a, obs)
belief[a] = prior_a * likelihood
normalize(belief)
In plain English: this is the numerator, summed over all values of A
Chris Piech, CS109
Inference
60
In general Bayes theorem with a random variable is like the cellphone problem: multiple possible assignments to keep track of
Still true when some variables are continuous
Chris Piech, CS109
Random Variables
61
Not all beliefs can be represented as a function. Dictionary / table is a great way to represent a random variable belief.
This is formally called non-parametric
Chris Piech, CS109
Representing Continuous Variables
62
Dictionary can also be used to represent a discretization of a continuous random var
We do it all the time! Yay compute!
Chris Piech, CS109
Four Prototypical Trajectories
Multiple observations??
Chris Piech, CS109
Multiple Observations
User: wrong
Obs 1
Observation Y = 0
(At font size 0.7)
Observation Y = 0
(At font size 0.8)
User: correct
Obs 2
Chris Piech, CS109
Multiple Observations
65
Single Observation:
Multiple Observations:
2
2
Chris Piech, CS109
Multiple Observations
66
Single Observation:
Multiple Observations:
2
2
Chris Piech, CS109
Posterior becomes new prior
67
User: correct
1
2
3
Observation Y = 1
(At font size 0.8)
Chris Piech, CS109
In the Code
Chris Piech, CS109
Four Prototypical Trajectories
Beyond Inference:
How do you select the next size to show?
Chris Piech, CS109
Today: Five New Real + Exciting Problems
70
Age from C14
Updated Delivery Prob
Age from Name
Hidden Chambers
Stanford Eye Test
Cellphone Tracking Cont.
Chris Piech, CS109
Chris Piech, CS109
Likelihood function f(X=x | T=t)? (soln)
Chris Piech, CS109
Observe LiDAR measurement of 4m. f(X=4 | T=t)=?
Chris Piech, CS109
We want a new belief in the true dist., f(T=t | X=4)
Chris Piech, CS109
Today: Compare and Contrast Many Examples
75
Age from C14
Updated Delivery Prob
Age from Name
Hidden Chambers
Stanford Eye Test
Updating Lidar Belief
Chris Piech, CS109