Exercise Type 14: GMM, VFE, FEP & Concept Questions

What the exam asks: A mixed bag of questions testing your ability to spot the correct GMM form, understand VFE properties, reason about FEP conceptually, and distinguish true/false statements about Bayesian vs ML approaches, Gaussian properties, and overfitting.


Part 0: What Do All These Symbols Mean?

The Key Notation

Symbol How to Read It What It Means
$\pi_k$ "pi sub k" Mixing coefficient — how common is component k?
$z_{nk}$ "z sub n k" Is data point n in cluster k? (1 if yes, 0 if no)
$K$ Capital K Total number of clusters/components
$F[q]$ "F of q" The VFE functional — a function of the variational distribution q
$q(z)$ "q of z" The variational distribution — our approximation to the true posterior
$p(x,z)$ "p of x comma z" The joint distribution over observed and latent variables
$KL(q \,|\, p)$ "KL divergence of q from p" How different q is from p — always ≥ 0
$p(y_\bullet \mid x_\bullet, D)$ "prob of y bullet given x bullet and data" Predictive distribution for a new data point
$L(\theta)$ "L of theta" The likelihood function — a function of the parameters

The Core Ideas — Plain English

GMM (Gaussian Mixture Model): Data comes from multiple hidden clusters, each generating points from its own Gaussian. The joint distribution uses a product over components, raised to the power $z_{nk}$, which acts as a switch — only the active cluster's term survives.

VFE (Variational Free Energy): When the true posterior is too hard to compute, we approximate it with a simpler distribution $q(z)$. VFE measures how good our approximation is. Minimizing VFE simultaneously improves the approximation and gives an estimate of the model evidence.

FEP (Free Energy Principle): A theory of intelligent behavior. Agents have an internal model of the world. They perceive by updating beliefs to match observations (minimize free energy). They act by choosing actions that minimize expected future surprise. Goals are encoded as target priors — preferred future states built into the model.

Bayesian vs MLE: MLE maximizes fit only. Bayesian averages over uncertainty. As data grows, MLE ≈ MAP because the likelihood dominates the prior. Bayesian evidence automatically penalizes complexity — no separate test set needed.


Part 1: The Key Formulas (MEMORIZE)

GMM Joint Distribution

Spotting trick: Both $\pi_k$ AND $\mathcal{N}$ must be inside the parentheses raised to $z_{nk}$ (not $z_n$). Product ∏, not sum Σ.

VFE Functional

Property Expression
Upper bound $F[q] \geq -\log p(x)$ for any $q(z)$
Equality $F[q] = -\log p(x)$ when $q(z) = p(z \mid x)$
Minimization Minimizes $KL(q(z) \,|\, p(z \mid x))$ AND gives evidence estimate

FEP Core Principles

Concept What It Means
Perception Minimize free energy (update beliefs to match observations)
Action Minimize expected free energy of future states
Goals Encoded as target priors in the generative model
Decision making Minimization of a functional of beliefs (not costs)

Concept Facts

Statement True/False Why
Likelihood = function of parameters True $L(\theta) = p(D \mid \theta)$
Linear combination of Gaussians = Gaussian True $Z = aX + bY$ preserves Gaussianity
Product of Gaussians = Gaussian False $Z = XY$ is NOT Gaussian
MLE = MAP always False Only when prior is uniform
Bayesian evidence = fit − complexity True Built-in overfitting protection
Discriminative classification ≈ regression True Predicts label, not full density

Part 2: Tricks & Shortcuts

TRICK 1: GMM Joint — Both Must Be in the Power

The correct answer ALWAYS has both $\pi_k$ and $\mathcal{N}$ inside the parentheses raised to $z_{nk}$.

If only $\mathcal{N}$ has the exponent → wrong. If only $\pi_k$ has the exponent → wrong. If it uses a sum (Σ) instead of product (∏) → that's the marginal, not the joint.

TRICK 2: VFE — Always ≥ Negative Log Evidence

Think: "Variational Free Energy is an Upper bound" → VFEU.

  • If an option says ≤ → wrong direction.
  • If it says equality at $q(z) = p(z)$ (prior) → wrong, it's the posterior $p(z \mid x)$.
  • If it says equality at $q(z) = 0$ → nonsense, not a valid distribution.

TRICK 3: FEP — Always MINIMIZE, Never Maximize

FEP is about MINIMIZING free energy (or expected free energy).

  • "Maximize free energy" → wrong direction.
  • "Cost function" → wrong framework (traditional engineering, not FEP).
  • "Desired" vs "actual" future → it's desired (target priors), not actual.
  • "Generative model" → if a statement says agents hold one, it's true.

TRICK 4: Gaussian Properties

  • Linear combination ($aX + bY$) → Gaussian ✅
  • Product ($XY$) → NOT Gaussian ❌
  • Sum of distributions (mixture) → NOT Gaussian in general ❌
  • Ratio ($X/Y$) → NOT Gaussian ❌

TRICK 5: Bayesian vs MLE

  • More data → likelihood gets narrower (more certain), prior stays fixed.
  • MLE ≈ MAP eventually because likelihood dominates.
  • Bayesian methods are NOT faster — often more expensive (integration).
  • No train/test split needed — complexity penalty handles overfitting.

Part 3: FULL Walkthrough of Real Exam Questions

EXAM QUESTION 1 (2021-Part-B, Question 1e — GMM Form)

Which is a correct GMM specification for $p(x_n, z_n)$ with one-hot $z_n$?

Options: - (a) $p(x_n,z_n) = \prod_{k=1}^K (\pi_k \cdot \mathcal{N}(x_n \mid \mu_k,\Sigma_k))^{z_n}$ - (b) $p(x_n,z_n) = \prod_{k=1}^K \pi_k \cdot \mathcal{N}(x_n \mid \mu_k,\Sigma_k)^{z_{nk}}$ - (c) $p(x_n,z_n) = \sum_{k=1}^K \pi_k \cdot \mathcal{N}(x_n \mid \mu_k,\Sigma_k)$ - (d) $p(x_n,z_n) = \prod_{k=1}^K (\pi_k \cdot \mathcal{N}(x_n \mid \mu_k,\Sigma_k))^{z_{nk}}$

STEP-BY-STEP SOLUTION

Step 1: Check the exponent

We need $z_{nk}$ (the k-th element), not $z_n$ (the whole vector).

(a) uses $z_n$ → ELIMINATE.

Step 2: Check what's raised to the power

Both $\pi_k$ AND $\mathcal{N}$ must be inside the exponentiation — the power selects the right cluster entirely.

(b) only $\mathcal{N}$ has the exponent, $\pi_k$ is outside → ELIMINATE.

Step 3: Check product vs sum

The joint uses a product (∏) — we're selecting one component per data point. The marginal uses a sum (Σ) — we've summed out the hidden variable.

(c) uses a sum → this is the marginal, not the joint → ELIMINATE.

Step 4: Verify (d)

(d) has: product over k ✓, both $\pi_k$ and $\mathcal{N}$ inside ✓, raised to $z_{nk}$ ✓

Answer: (d)


EXAM QUESTION 2 (2021-Resit, Question 4e — VFE Property)

$F[q] = \int q(z) \log \frac{q(z)}{p(x,z)} \, dz$. Which statement is true?

Options: - (a) $F[q] = -\log p(x)$ if $q(z) = 0$ - (b) $F[q] \leq -\log p(x)$ for any $q(z)$ - (c) $F[q] \geq -\log p(x)$ for any $q(z)$ - (d) $F[q] = -\log p(x)$ if $q(z) = p(z)$

STEP-BY-STEP SOLUTION

Step 1: Recall the fundamental VFE property

$F[q] \geq -\log p(x)$ for any $q(z)$. It's an upper bound on the negative log evidence.

This immediately gives us (c).

Step 2: Eliminate the rest

(a) $q(z) = 0$ is not a valid probability distribution (doesn't integrate to 1) → ELIMINATE.

(b) Wrong direction — it's ≥ not ≤ → ELIMINATE.

(d) Equality is at $q(z) = p(z \mid x)$ (the true posterior), not $q(z) = p(z)$ (the prior) → ELIMINATE.

Answer: (c)


EXAM QUESTION 3 (2023, Question 4f — VFE & Bayesian Inference)

Why can VFE minimization approximate Bayesian inference?

Options: - (a) VFE is Bayesian inference plus Gaussian noise - (b) VFE minimizes KL-divergence to evidence, and VFE is upper bound on posterior - (c) VFE minimizes evidence by optimizing variational posterior - (d) VFE minimizes KL-divergence between variational and true posterior, and VFE is upper bound on negative log evidence

STEP-BY-STEP SOLUTION

Step 1: Two key facts about VFE

  1. $F[q] = KL(q(z) \,|\, p(z \mid x)) - \log p(x)$. Since $-\log p(x)$ doesn't depend on $q$, minimizing $F[q]$ minimizes the KL divergence to the true posterior.
  2. $F[q] \geq -\log p(x)$ (upper bound on negative log evidence).

Step 2: Match the options

(a) "Plus Gaussian noise" — nonsense → ELIMINATE.

(b) "KL to evidence" — wrong, it's KL to the posterior. "Upper bound on posterior" — wrong, it's on negative log evidence → ELIMINATE.

(c) "Minimizes evidence" — wrong, it minimizes KL to the posterior → ELIMINATE.

(d) Both parts are correct → ✓

Answer: (d)


EXAM QUESTION 4 (2021-Part-B, Question 1b — VB Properties)

Which is NOT a property of the Variational Bayesian approach?

Options: - (a) Transfers Bayesian ML to optimization problem - (b) VB finds posteriors by maximizing Bayesian evidence - (c) VFE minimization gives posterior AND evidence - (d) Global VFE minimization realizes Bayes rule

STEP-BY-STEP SOLUTION

(a) True — VB turns integration into optimization → NOT the answer.

(b) VB finds posteriors by minimizing VFE (upper bound on negative log evidence), not "maximizing evidence" → This is wrong → ELIMINATE as the answer.

(c) True — VFE gives both $q(z)$ (approximate posterior) and a lower bound on log evidence → NOT the answer.

(d) True — at the global minimum, $q(z) = p(z \mid x)$ (exact Bayes rule) → NOT the answer.

Answer: (b)


EXAM QUESTION 5 (2022, Question 4a — FEP Comprehension)

Which statement is most consistent with FEP?

Options: - (a) Actions aim to minimize free energy of future states - (b) Actions aim to minimize complexity of future states - (c) Intelligent decision making requires minimization of a functional of beliefs about future states - (d) Intelligent decision making requires minimization of a cost function of future states

STEP-BY-STEP SOLUTION

(a) Close, but imprecise — it's about minimizing a functional of beliefs, not directly "free energy of future states."

(b) "Minimize complexity" — not the core idea of FEP → ELIMINATE.

(c) Yes — expected free energy is a functional of probability distributions (beliefs) about future states → ✓

(d) "Cost function" — that's traditional control theory, not FEP → ELIMINATE.

Answer: (c)


EXAM QUESTION 6 (2021-Part-B, Question 1d — FEP Goal-Driven Behavior)

How to equip an agent with goal-driven behavior in FEP?

Options: - (a) Specify cost function, minimize costs - (b) Extend generative model with target priors for future observations. Minimize Free Energy. - (c) Specify cost function of actions, minimize - (d) Extend with posterior for actions, maximize posterior

STEP-BY-STEP SOLUTION

In FEP, goals are encoded as target priors — preferred future states built into the generative model. The agent then minimizes expected free energy, which naturally achieves these goals.

(a), (c) "Cost function" — wrong framework → ELIMINATE. (d) Doesn't explain how goals are encoded → ELIMINATE.

Answer: (b)


EXAM QUESTION 7 (2021-Part-B, Question 5b — FEP Consistent Statements)

Which statements are consistent with FEP? - (a) Active inference agent holds generative model for sensory inputs - (b) Actions inferred from differences between predicted and desired future observations - (c) Actions inferred from differences between predicted and actual future observations - (d) Agent focuses on explorative behavior only

Options: - (a) (a) and (b) - (b) (a) - (c) (b) and (d) - (d) (c) and (d)

STEP-BY-STEP SOLUTION

(a) True — generative model is fundamental to FEP → ✓ (b) True — actions bridge predicted vs. desired (target prior) states → ✓ (c) False — it's about desired, not "actual" future → ✗ (d) False — agents balance exploration AND exploitation → ✗

(a) and (b) are true.

Answer: (a) — (a) and (b)


EXAM QUESTION 8 (2021-Part-A, Question 1a — Likelihood Terminology)

"The likelihood of the parameters" is more appropriate than "the likelihood of the data".

Options: (a) true, (b) false

STEP-BY-STEP SOLUTION

The likelihood IS a function of the parameters given the data: $L(\theta) = p(D \mid \theta)$. So "likelihood of the parameters" is correct usage.

Answer: (a) true


EXAM QUESTION 9 (2021-Part-A, Question 1b — Gaussian Properties)

If X and Y are independent Gaussian distributed variables, then $Z = 3X - XY$ is also Gaussian.

Options: (a) true, (b) false

STEP-BY-STEP SOLUTION

The product $XY$ of two Gaussians is NOT Gaussian in general. Linear combinations like $3X - Y$ are Gaussian, but $3X - XY$ involves a product term.

Answer: (b) false


EXAM QUESTION 10 (2021-Part-A, Question 1d — MLE vs MAP)

MLE always selects the parameter where the Bayesian posterior distribution is maximal.

Options: (a) true, (b) false

STEP-BY-STEP SOLUTION

MLE maximizes $p(D \mid \theta)$. MAP maximizes $p(D \mid \theta)p(\theta)$. They're only equal when the prior $p(\theta)$ is uniform. MLE ≠ MAP in general.

Answer: (b) false


EXAM QUESTION 11 (2021-Resit, Question 4b — Bayesian vs MLE)

Which statement is most accurate?

Options: - (a) ML becomes better approximation as data grows, since prior becomes wider - (b) ML becomes worse as data grows, since both likelihood and prior become wider - (c) ML becomes better as data grows, since likelihood becomes wider while prior doesn't depend on data - (d) ML becomes better as data grows, since likelihood becomes narrower while prior doesn't depend on data

STEP-BY-STEP SOLUTION

With more data: - Likelihood becomes narrower (more certain about the true parameter) - Prior stays the same (doesn't depend on data) - The likelihood eventually dominates → MLE ≈ MAP

(a) Prior doesn't become wider → ELIMINATE. (b) Both don't become wider → ELIMINATE. (c) Likelihood becomes narrower, not wider → ELIMINATE.

Answer: (d)


EXAM QUESTION 12 (2021-Part-B, Question 1a — Bayesian Approach)

Which statement is FALSE about the Bayesian approach?

Options: - (a) No need to split data into train/test. All data used for training. - (b) Fundamentally sound, based on probability theory. - (c) Requires explicit model assumptions upfront. - (d) Fast alternative to maximum likelihood.

STEP-BY-STEP SOLUTION

Bayesian methods are NOT generally faster — they're often more computationally expensive due to integration over parameters.

Answer: (d) false


EXAM QUESTION 13 (2021-Part-B, Question 5d — Discriminative Classification)

Which statements are true? - (a) Product of independent Gaussians is Gaussian - (b) Linear combination $Z = 3X - Y$ of independent Gaussians is Gaussian - (c) Sum of two Gaussians is Gaussian - (d) Discriminative classification is more similar to regression than density estimation

Options: - (a) (b) and (c) - (b) (a) and (d) - (c) (b) and (d) - (d) (b) and (c)

STEP-BY-STEP SOLUTION

(a) False — the product random variable $Z = XY$ is NOT Gaussian → ✗ (b) True — linear combinations of Gaussians are Gaussian → ✓ (c) True — sum of Gaussians is Gaussian → ✓ (d) True — discriminative classification predicts a label value (like regression), not a full density → ✓

Options with (a) → ELIMINATE. (b), (c), and (d) are all true.

Best match: (c) — (b) and (d).

Answer: (c)


EXAM QUESTION 14 (2023, Question 4b — Overfitting)

Why is a "Bayesian engineer" not concerned about overfitting?

Options: - (a) Model evidence decomposes as "fit minus complexity". Complexity prevents overfitting. - (b) Uses separate test set - (c) Minimizes probability of overfitting - (d) Evidence decomposes as "fit minus entropy"

STEP-BY-STEP SOLUTION

Bayesian model evidence naturally penalizes complexity (the "Occam factor"):

This is built-in overfitting protection — no separate test set needed.

(b) That's the frequentist approach → ELIMINATE. (c) Vague, not the reason → ELIMINATE. (d) "Entropy" — wrong term, it's complexity → ELIMINATE.

Answer: (a)


EXAM QUESTION 15 (2021-Part-B, Question 1c — Bayesian Model Comparison)

Posterior model probability $p(m_k \mid D)$:

Options: - (a) $p(m_k \mid D) = p(m_k) \int p(\theta \mid D,m_k) \, p(D \mid m_k) \, d\theta$ - (b) $p(m_k \mid D) = p(m_k) \int p(D \mid \theta,m_k) \, p(\theta \mid m_k) \, d\theta$ - (c) $p(m_k \mid D) = \sum_n p(m_k) \, p(D \mid m_k)$ - (d) $p(m_k \mid D) = \int p(D \mid \theta,m_k) \, p(\theta \mid m_k) \, d\theta$

STEP-BY-STEP SOLUTION

Where the evidence is:

So without the normalizing constant $p(D)$:

(a) Integrates $p(\theta \mid D,m_k) \cdot p(D \mid m_k)$ — wrong terms → ELIMINATE. (c) Sums over data points — nonsense → ELIMINATE. (d) Missing the prior $p(m_k)$ — that's just the evidence, not the posterior → ELIMINATE.

Answer: (b)


Part 4: Practice Exercises

Exercise 1 (GMM Form)

Which is a correct GMM for $p(x_n, z_n)$ with one-hot $z_n$?

Options: - (a) $\prod_{k=1}^K (\pi_k \cdot \mathcal{N}(x_n \mid \mu_k,\Sigma_k))^{z_n}$ - (b) $\prod_{k=1}^K \pi_k \cdot \mathcal{N}(x_n \mid \mu_k,\Sigma_k)^{z_{nk}}$ - (c) $\sum_{k=1}^K \pi_k \cdot \mathcal{N}(x_n \mid \mu_k,\Sigma_k)$ - (d) $\prod_{k=1}^K (\pi_k \cdot \mathcal{N}(x_n \mid \mu_k,\Sigma_k))^{z_{nk}}$


Exercise 2 (VFE Property)

$F[q] = \int q(z) \log \frac{q(z)}{p(x,z)} \, dz$. Which is true?

Options: - (a) $F[q] = -\log p(x)$ if $q(z) = 0$ - (b) $F[q] \leq -\log p(x)$ for any $q(z)$ - (c) $F[q] \geq -\log p(x)$ for any $q(z)$ - (d) $F[q] = -\log p(x)$ if $q(z) = p(z)$


Exercise 3 (FEP Action)

In FEP, how does an agent choose actions?

Options: - (a) Minimize cost function of future states - (b) Minimize expected free energy in future states - (c) Maximize free energy in future states - (d) Maximize expected accuracy of future states


Exercise 4 (Gaussian Property)

Which is true?

Options: - (a) Product of independent Gaussians is Gaussian - (b) Linear combination of independent Gaussians is Gaussian - (c) Ratio of Gaussians is Gaussian - (d) Product $Z = XY$ of Gaussians is Gaussian


Exercise 5 (Bayesian vs MLE)

As data size grows:

Options: - (a) ML becomes worse, prior becomes wider - (b) ML becomes better, likelihood becomes narrower, prior unchanged - (c) ML becomes worse, both become wider - (d) ML becomes better, prior becomes narrower


Exercise 6 (Discriminative Classification)

Discriminative approach to classification for new input $x_\bullet$:

Options: - (a) $p(y_\bullet \mid x_\bullet,D) = \int p(y_\bullet \mid x_\bullet,\theta,D) \, d\theta$ - (b) $p(y_\bullet \mid x_\bullet) = \int p(y_\bullet \mid x_\bullet,\theta) \, p(\theta) \, d\theta$ - (c) $p(y_\bullet \mid x_\bullet,D) = \int p(y_\bullet \mid x_\bullet,\theta) \, p(\theta \mid D) \, d\theta$ - (d) $p(y_\bullet \mid x_\bullet) = \int p(y_\bullet \mid x_\bullet,\theta) \, d\theta$


Exercise 7 (Overfitting)

Why are Bayesian engineers not concerned about overfitting?

Options: - (a) Evidence = fit − complexity - (b) Use separate test set - (c) Minimize overfitting probability - (d) Evidence = fit − entropy


Exercise 8 (Generative Model for Signal Recovery)

Bayesian records signal $x$, wants to recover speech $s$. How?

Options: - (a) Model $p(x,s,z) = p(s \mid x,z)p(x,z)$, compute $p(s \mid x) \propto \int p(s \mid x,z) \, dz$ - (b) Model $p(x,s,z) = p(s \mid x,z)p(x,z)$, compute $p(s \mid x) \propto \int p(s \mid x,z)p(x,z) \, dz$ - (c) Model $p(x,s,z) = p(x \mid s,z)p(s,z)$, compute $p(s \mid x,z) = \frac{p(x \mid s,z)p(s,z)}{p(x,z)}$ - (d) Model $p(x,s,z) = p(x \mid s,z)p(s,z)$, compute $p(s \mid x) \propto \int p(x \mid s,z)p(s,z) \, dz$


Part 5: Answers

Exercise 1 **Answer: (d)** (a) uses $z_n$ instead of $z_{nk}$. (b) only $\mathcal{N}$ has the exponent. (c) is a sum (that's the marginal, not the joint). (d) is correct: both inside, power $z_{nk}$, product.
Exercise 2 **Answer: (c)** VFE is always ≥ negative log evidence (upper bound). (a) $q(z)=0$ is not a valid distribution. (b) wrong direction. (d) equality at posterior, not prior.
Exercise 3 **Answer: (b)** Actions minimize expected free energy, not maximize. Not cost functions.
Exercise 4 **Answer: (b)** Linear combinations of Gaussians are Gaussian. Products and ratios are not.
Exercise 5 **Answer: (b)** With more data, likelihood narrows (more certain), prior stays fixed. ML ≈ MAP eventually.
Exercise 6 **Answer: (c)** Bayesian predictive: average the likelihood $p(y \mid x,\theta)$ over the posterior $p(\theta \mid D)$. (a) has wrong conditioning. (b), (d) don't use the data $D$.
Exercise 7 **Answer: (a)** Evidence = training fit − complexity. The complexity term prevents overfitting. (b) is the frequentist approach. (d) says "entropy" — wrong.
Exercise 8 **Answer: (d)** Joint model $p(x,s,z) = p(x \mid s,z)p(s,z)$. Marginalize $z$: $p(s \mid x) \propto \int p(x \mid s,z)p(s,z) \, dz$. (a), (b) use wrong factorization. (c) conditions on $z$ which is unobserved.