Question bank — Neuro-symbolic AI

False — it trades some raw fit for consistency, sample-efficiency, and interpretability; on huge clean datasets a pure net can match or beat it, the payoff shows on small or rule-governed data.

Partly false — the rules are often human-given, but systems like differentiable theorem provers learn the confidence of facts and even soft rules by gradient descent, so the symbolic side is not fully manual.

False — too large a $\lambda$ makes the model obey rules at the expense of the data, so it fits training labels poorly; $\lambda$ is a balance to be tuned, not maximised.

True at the corners — plug in the four crisp cases: $0\cdot 0=0$, $0\cdot 1=0$, $1\cdot 0=0$, $1\cdot 1=1$, which is exactly the classical AND truth table; the relaxation only differs in the fuzzy middle, which is precisely where we need gradients.

False — the whole point is reuse: a small library of modules (find, left_of, query_color) is composed differently per question, so one left_of serves every "left of" query.

False — it is one of several t-norms (families of fuzzy-AND): Gödel $\min (a,b)$, Łukasiewicz $\max (0,a+b-1)$, and product $a\cdot b$ all reduce to classical AND at the corners; product is preferred here for reasons beyond "it works" (see the dedicated why-question).

False — it keeps the logic explicit and inspectable (you can still read the rule parent∧parent→grandparent); it only softens the true/false values to $[0,1]$ so gradients can flow.

Mostly-but-careful — the symbolic inference chain is readable, yet the neural front-end that produced the predicates (e.g. "this pixel is a cube") is still opaque; you inherit XAI problems at the perception boundary.

False as stated — plain nets struggle with it, but architectures with explicit compositional structure (modules, attention over parts, attention mechanisms) can achieve it; it is a weakness, not a law.

Wrong sign of the constraint — mutual exclusivity means the two probabilities should not both be high, so the violation is $\max (0,p(c_1)+p(c_2)-1)$; the stated form instead punishes them for summing to less than 1.

Order flipped — the rule "pneumonia $\Rightarrow$ high opacity" means opacity must not fall below pneumonia, so violation $=\max (0,p(\text{pneumonia})-p(\text{opacity}))=\max (0,0.8-0.3)=0.5$, i.e. it was violated.

$\min$ is not smooth — it has a kink along $a=b$ (the crease in the left panel of the figure), so its gradient is undefined there; the product t-norm $a\cdot b$ is chosen precisely because it is differentiable everywhere.

Wrong quantifier — deriving $S(x,z)$ needs "there exists some linking $y$ with $P(x,y)$ AND $Q(y,z)$", so we take ${\max }_y[{\mu }_{P(x,y)}\cdot {\mu }_{Q(y,z)}]$; existential $\exists y$ relaxes to ${\max }_y$, whereas ${\min }_y$ would encode $\forall y$ (every $y$ must work), the wrong logic.

Leaves the interval — $-a$ can go negative and is not a truth degree; the correct relaxation is $\neg a=1-a$, which maps $[0,1]\to [0,1]$ and still flips $0\leftrightarrow 1$.

Overreach — the differentiable form is for learning (gradients); many systems still round to crisp symbols at inference for exact, verifiable answers, so discrete logic returns at deployment.

Backwards — a good symbolic prior shrinks the hypothesis space, so fewer examples are needed; it typically adds constraints, not free parameters, aligning with few-shot learning.

No conversion needed — on this page both are truth values in $[0,1]$ sharing one scale; a classifier's $p(\text{blue})=0.92$ is handed to the logic engine as ${\mu }_{\text{blue}}=0.92$ directly.

Its outputs are independent learned numbers with no built-in logical link; nothing forces $p(\text{awake})+p(\text{asleep})\le 1$ unless a symbolic constraint injects that structure.

Crisp values are flat almost everywhere and jump at the boundary, giving zero or undefined gradient; a continuous relaxation gives a slope everywhere so gradient descent can actually learn confidences.

Plain $a+b$ can exceed $1$ (e.g. $0.7+0.6=1.3$, the dashed line leaving the box in the figure), so it is no longer a truth degree; subtracting $ab$ keeps the result in $[0,1]$ and matches classical OR at the corners — this is exactly the probabilistic sum (t-conorm of the product t-norm).

Two extra reasons: (1) product/probabilistic-sum have non-zero gradients almost everywhere, whereas Gödel's $\min$/$\max$ pass gradient to only one input (the winner) and Łukasiewicz clips to a flat zero region that kills learning; (2) if the truth values are read as independent probabilities, $a\cdot b$ and $a+b-ab$ are the exact probabilities of "both true" and "at least one true", giving the relaxation a clean probabilistic meaning.

A rule like a+b or parent∧parent→grandparent encodes structure that a net would otherwise have to rediscover from millions of samples; the prior supplies that knowledge for free.

It is subsymbolic — it names an internal number, not a human-meaningful reason; the symbolic layer instead yields a chain like "income < threshold AND debt-ratio high", which a person can contest.

Because a new question is just a new arrangement of already-learned operations; if find, above, query_color each work alone, their fresh composition works without any new training.

Both express knowledge as explicit relations over named entities (edges, causes(X,Y)), which is exactly the symbolic format the reasoning engine consumes, so the neural side can feed extracted predicates straight in.

The violation is $0$, so its gradient contribution vanishes — the constraint becomes silent and the model is driven purely by data until it drifts back into violation.

The AND $a\cdot b$ collapses to $0$ and, being multiplication, its gradient with respect to the other premise is also $0$ — a dead branch that cannot recover through that path, a known vanishing-signal trap.

The product underflows toward zero ($0.3^{20}\approx 3.5\times 10^{-11}$), so even a genuinely-supported conclusion looks false and its gradient shrinks with it; practitioners fix this by computing in the log-domain ($\sum \log \mu$ instead of $\prod \mu$) or by tempering the t-norm (see next item).

Introduce a temperature $\tau >0$ and replace ${\prod }_i{\mu }_i$ with $\exp \left(\frac{1}{\tau }{\sum }_i\log {\mu }_i\right)$; as $\tau$ grows the exponent is shrunk toward $0$, so the chain's truth is pulled away from $0$ and toward $1$, keeping values numerically alive and gradients non-vanishing. At $\tau =1$ this is exactly the ordinary product; large $\tau$ softens long conjunctions.

The ${\max }_y$ keeps only the strongest supporting path; adding a second route does not increase the derived truth beyond the best single one, matching existential "at least one $y$ works".

The system treats it as maximally uncertain; through the product t-norm it dampens every conclusion drawn from it, and gradients can push it toward $0$ or $1$ as evidence accumulates.

The constraint then fights the data — either it forces incorrect predictions (if $\lambda$ is high) or the data overrides it (if $\lambda$ is low); garbage rules give a garbage prior, so rule correctness is a real assumption, not a guarantee.

Every relaxation ($a\cdot b$, $a+b-ab$, $1-a$, $\max$) converges exactly to classical logic, so a fully-confident differentiable prover behaves identically to a discrete one.

You are left with a pure symbolic reasoner — perfect on the given predicates but blind to raw images/audio/text, which is exactly the perception gap neuro-symbolic AI was built to close (compare with RL agents that still need a perceptual encoder).