Core ML and Neural Network Foundations — NVIDIA NCA Gen AI LLMs practice questions

Gradient clipping is applied during LLM training MAINLY to:

1. Bound update magnitude when rare large gradients occur
2. Force sparsity in the gradient vector so optimizer memory is saved
3. Compensate for a learning rate that is too low
4. Speed up convergence by amplifying small gradients

Answer: A — Bound update magnitude when rare large gradients occur

A) Correct — clipping caps the norm of the update so an outlier batch cannot destabilize the run. B) Clipping rescales rather than sparsifies, and it saves no memory. C) It does nothing for an under-sized learning rate. D) It never amplifies gradients; it only limits large ones.

During backpropagation, what quantity is actually propagated backward through the network layers?

1. The raw input features, re-scaled per layer
2. The updated weight values from the output layer
3. The activation values that were recorded during the forward pass step
4. The partial derivatives of the loss with respect to each layer's inputs

Answer: D — The partial derivatives of the loss with respect to each layer's inputs

D) Correct — backprop applies the chain rule, passing gradients of the loss backward to compute each parameter's derivative. A) Inputs move forward, not backward. B) Weights are updated locally by the optimizer after gradients exist; they are not propagated. C) Activations are cached from the forward pass and used, but they are not what flows backward.

A residual (skip) connection helps very deep networks train PRIMARILY by:

1. Reducing the total number of parameters in the block
2. Replacing the need for nonlinear activations
3. Forcing each block to learn a strictly orthogonal representation
4. Providing a direct path for gradients to reach earlier layers

Answer: D — Providing a direct path for gradients to reach earlier layers

D) Correct — the identity path carries gradient backward largely undiminished, mitigating depth-related vanishing. A) Skip connections add no parameters but do not reduce them either. C) Nothing enforces orthogonality between blocks. B) Nonlinearities remain essential inside each residual block.

A common cause of training divergence (NaN losses) when using FP16 is:

1. Using BF16 tensors everywhere instead of FP16 tensors
2. Running the training job across far too many GPUs
3. Overflow of attention logits in FP16's narrow range
4. Using the Adam optimizer rather than plain SGD

Answer: C — Overflow of attention logits in FP16's narrow range

C) FP16 has a narrow exponent range, so large attention logits and gradients overflow to inf and then NaN; loss scaling or BF16 addresses it. A) BF16 has the wider range and helps rather than hurts. B) Scale alone does not produce NaNs. D) Adam is the standard optimizer for these runs.

Compared with sigmoid, the tanh activation offers which specific advantage for hidden layers?

1. It cannot saturate even for very large-magnitude inputs
2. It has a strictly monotonic derivative
3. It is zero-centered, so activations do not drift positive
4. It is computationally cheaper to evaluate

Answer: C — It is zero-centered, so activations do not drift positive

C) Correct — a zero-centered output range avoids the systematically positive activations that bias sigmoid-based gradient updates. A) Tanh saturates at both tails just as sigmoid does. B) Its derivative rises then falls; it is not monotonic. D) Both involve exponentials with comparable cost.

Which pairing correctly matches a task to its standard loss?

1. Binary classification paired with a categorical hinge loss
2. Multi-label tagging with softmax cross-entropy
3. Continuous price prediction with mean squared error
4. Next-token prediction with mean absolute error

Answer: C — Continuous price prediction with mean squared error

C) Correct — squared error is the conventional objective for continuous-valued regression targets. A) Binary classification standardly uses binary cross-entropy. B) Multi-label needs per-class sigmoids since labels are not mutually exclusive. D) Token prediction is a classification problem over the vocabulary, trained with cross-entropy.

Initializing every weight in a dense network to the same constant fails because:

1. The loss function becomes non-differentiable at that particular point
2. Activations immediately overflow to infinity
3. The optimizer cannot compute a second moment estimate
4. Units in a layer receive identical gradients and never differentiate

Answer: D — Units in a layer receive identical gradients and never differentiate

D) Correct — symmetric weights produce symmetric gradients, so all units in a layer remain identical and the layer's capacity collapses to one unit. A) The loss remains differentiable; symmetry is the issue. C) Moment estimates compute fine on identical gradients. B) Overflow depends on scale, not on symmetry, and is not the general failure.

The softmax function converts a vector of logits into:

1. A set of independent probabilities that each range 0 to 1
2. A probability distribution whose components sum to one
3. A zero-mean, unit-variance normalized score vector
4. A one-hot vector that selects the maximum logit

Answer: B — A probability distribution whose components sum to one

B) Correct — exponentiating and normalizing yields nonnegative values summing to one across the classes. A) Independent per-class probabilities that need not sum to one come from elementwise sigmoids. C) That describes standardization, not softmax. D) Hard selection of the maximum is argmax; softmax is a smooth relaxation.

L2 regularization (weight decay) influences a model by:

1. Penalizing large weight magnitudes, favoring smoother functions
2. Driving the majority of weights exactly to zero for feature selection
3. Normalizing activations to zero mean and unit variance
4. Halting training once validation loss stops improving

Answer: A — Penalizing large weight magnitudes, favoring smoother functions

A) Correct — adding a squared-norm penalty biases the optimizer toward smaller weights and lower-complexity functions. B) Exact sparsity is the hallmark of L1, not L2. C) Activation standardization is what batch or layer normalization does. D) That is early stopping, a separate regularizer.

A deep sigmoid network trains extremely slowly, and gradient magnitudes in early layers are near zero. The MOST likely cause is:

1. Exploding gradients from an oversized learning rate
2. A batch size too small to estimate the gradient
3. Label noise in the training set steadily overwhelming the loss signal
4. Vanishing gradients from repeated multiplication of small derivatives

Answer: D — Vanishing gradients from repeated multiplication of small derivatives

D) Correct — sigmoid derivatives peak at 0.25, so chaining them across many layers shrinks gradients toward zero. A) Exploding gradients produce huge values or NaNs, the opposite symptom. C) Label noise raises loss and hurts accuracy but does not systematically zero early-layer gradients. B) Small batches add variance to gradients, not a depth-dependent decay.

Parameters vs Compute vs Data — scaling laws?

Chinchilla: optimal compute split is balanced parameters + tokens. ~20 tokens per parameter for training.

Why did ReLU-family activations largely replace sigmoid and tanh in deep networks?

Sigmoid and tanh saturate, driving derivatives toward zero and causing vanishing gradients in deep stacks. ReLU has a constant gradient of 1 on the positive side, so signal propagates through many layers, and it is cheap to compute. Variants like GELU and SwiGLU smooth the kink and are common in transformer feed-forward blocks.

Which loss function trains an autoregressive language model, and what quantity is it minimizing?

Cross-entropy loss over the vocabulary at each position. It minimizes the negative log-likelihood the model assigns to the actual next token, which is equivalent to minimizing the average number of nats or bits needed to encode the training text.

Distinguish supervised, unsupervised, and self-supervised learning by what supplies the training signal.

Supervised learning uses human-provided labels paired with inputs. Unsupervised learning finds structure in unlabeled data, e.g. clustering or dimensionality reduction. Self-supervised learning generates labels from the data itself — next-token prediction on raw text is the canonical LLM example.