Probabilistic Models: Joints, marginals, independence, and Bayes | Stanford CS109 Lecture 10

🌏 中文版

Video status: Recordings require sign-in or course authorization. Source details

This is article 11 in Reading Stanford CS109, covering Summer 2026 Lecture 10: Probabilistic Models on July 7 with Chris Gregg. Its Summer agenda follows the worksheet, answer key, LLM Learning Guide, and shared Spring-dated reader chapters on joint distributions and inference. The Canvas recording is inaccessible, so spoken material is not reconstructed.

The lecture expands from one random variable to how several variables move together. A joint distribution is the complete starting point. Summing out an unneeded variable produces a marginal; normalizing a row or column produces a conditional; independence asks whether the joint factors into a product of marginals.

P1: Open with a Normal review

For an IQ-like score X~N(100,225), the second parameter is variance, so σ=15. The z-score of 130 is two:

P(X>130) = 1-Φ((130-100)/15)
         = 1-Φ(2)
         = 1-0.9772 = 0.0228

This preserves the previous lecture's distinction among σ², σ, and a right-tail complement before introducing two-variable models.

P2: Read a joint PMF

Let X be relationship status—single (S) or in a relationship (R)—and Y be class year: freshman (Fr), sophomore (So), or junior (Jr). The official table is:

All six cells are nonnegative and sum to 1.00, making this a valid joint PMF. Read joint probabilities directly: P(R,Jr)=0.30 and P(S,Fr)=0.18. The junior probability sums its column: P(Y=Jr)=0.10+0.30=0.40.

P3: Sum a joint into marginals

Fix each value of X and sum across every possible Y:

P(X=S) = 0.18+0.12+0.10 = 0.40
P(X=R) = 0.12+0.18+0.30 = 0.60

Similarly, summing over relationship status gives

P(Y=Fr)=0.30,  P(Y=So)=0.30,  P(Y=Jr)=0.40

In general, P(X=x)=Σy P(X=x,Y=y). The possible values of Y are mutually exclusive and exhaustive, so this is the law of total probability. The word “marginal” comes from writing these sums in the table's margins.

P4: Conditioning and independence

Given that someone is a junior, the conditional probability of status R is

P(X=R | Y=Jr) = P(R,Jr)/P(Y=Jr)
               = 0.30/0.40 = 0.75

The denominator renormalizes the junior column to sum to one. The table cell 0.30 is a joint probability; 0.75 is the proportion within the restricted junior sample space.

Independence would require P(x,y)=P(x)P(y) for every cell. One counterexample disproves it:

P(S,Fr)=0.18
P(S)P(Fr)=0.40×0.30=0.12

Therefore X and Y are dependent. One matching cell cannot prove full independence; proving it requires every value pair or an equivalent structural argument.

P5: Bayes with a discrete hidden variable and continuous likelihood

Let Y=1 mean a baby can hear, with prior P(Y=1)=0.75. Gaze change after a sound follows

X | Y=1 ~ N(15,25)
X | Y=0 ~ N(8,25)

After observing X=14, Bayes uses conditional densities as likelihoods even though a continuous variable assigns zero probability to an exact point:

P(Y=1 | X=14)
= f(14|Y=1)P(Y=1)
  / [f(14|Y=1)P(Y=1)+f(14|Y=0)P(Y=0)]

Both Normals have σ=5, so the shared 1/(σ√(2π)) cancels. Using only the exponential parts,

f(14|Y=1) ∝ e^-0.02 ≈ 0.9802
f(14|Y=0) ∝ e^-0.72 ≈ 0.4868

posterior ≈ 0.75(0.9802)
          / [0.75(0.9802)+0.25(0.4868)]
          ≈ 0.858

Fourteen is closer to 15 than to 8, so the observation raises belief in hearing from 0.75 to about 0.858. A density serves as relative likelihood; the normalized posterior is the probability.

P6: Why a full joint table does not scale

Twenty random variables with five values each require one cell per complete assignment:

5^20 = 95,367,431,640,625 cells

This exponential growth motivates independence assumptions and Bayes networks. Rather than store every joint cell, a structured model factorizes the distribution into smaller local pieces. P6 is a pset4 item deliberately omitted from the public answer key; the count follows directly from the prompt.

Challenge: A Tired Baby with Exponential likelihoods

The prior is P(Tired)=3/4. Eye-rub time follows Exp(3) when tired and Exp(1) otherwise. Observing a rub at two minutes and using f(t)=λe^-λt gives

f(2|Tired) = 3e^-6
f(2|not Tired) = e^-2

P(Tired|t=2)
= (3/4)(3e^-6)
  / [(3/4)(3e^-6)+(1/4)e^-2]
≈ 0.142

The posterior falls from 0.75 to about 0.142. Exp(3) has mean wait 1/3 minute, versus one minute for Exp(1), so a two-minute wait is relatively more plausible when not tired. The challenge is also hidden from the public pset key. PDF extraction renders the prior fraction 3/4 as 34; the original page clearly displays a fraction.

How to use the LLM Learning Guide

The guide orders joint PMFs, marginals, conditioning within a joint, independence, random-variable Bayes, and a discrete hypothesis with continuous density. Reuse one table to ask four questions: which cell is the joint, which sum is the marginal, which marginal normalizes the conditional, and whether the joint equals the product of marginals. With a continuous observation, only the likelihood changes from a PMF to a density; prior multiplication and normalization remain unchanged.

Material boundaries

• This guide covers official P1–P6, the optional tired-baby challenge, and all six LLM-guide concepts.
• P6 and the challenge are pset4 items deliberately omitted from the public answer key; this article derives them only from the public prompts.
• The challenge prior is 3/4; 34 is a PDF fraction-extraction artifact.
• The Canvas recording is inaccessible. The worksheet and guide are two pages each, so the short-material exception applies; the article remains draft: true pending independent review.

Course video sources

Verified the official Summer 2026 archive: the course is recorded and its syllabus says recordings are available through Canvas. The archived homepage's Videos link redirects to Stanford Canvas login and requires course access. None of the lectures in the 2022 Stanford Online public playlist matches this lecture's title, so no public player is provided and recordings from other offerings are not passed off as this one.

Official sources:

• CS109 Summer 2026 — Canvas lecture recordings (login required)
• CS109 Summer 2026 — archived official schedule
• CS109 Summer 2026 — archived syllabus and recording policy
• Stanford Online CS109 2022 public playlist (different offering, supplementary)

Checked: 2026-10-10.

Update Log

• 2026-10-10: Added explicit video status and checked recording sources and access notes.
• 2026-10-10: Rechecked video status. Confirmed Summer 2026 recordings need Canvas sign-in and the 2022 public playlist has no matching lecture; status unchanged and check date added.

References

• Stanford CS109 Summer 2026 schedule
• Lecture 10: Probabilistic Models
• Lecture 10 worksheet
• Lecture 10 answer key
• Lecture 10 LLM Learning Guide
• Probability for Computer Science: Joint distributions
• Probability for Computer Science: Inference
• Probability for Computer Science: Bayesian carbon dating