Deep learning quiz

Question 1

Why does Batch Normalization sometimes hurt performance in very small mini-batch regimes, and what alternatives address this?

Accepted Answer

BatchNorm statistics become noisy with small batches; GroupNorm/LayerNorm are batch-independent alternatives

Answer

BatchNorm always improves performance regardless of batch size

Answer

Small batches make BatchNorm deterministic, removing regularization

Answer

Only Dropout can fix small-batch normalization issues

Question 2

What is the key computational advantage of FlashAttention over a naive softmax attention implementation?

Accepted Answer

It is IO-aware, tiling computation to avoid materializing the N×N matrix in HBM

Answer

It approximates attention with low-rank factorization

Answer

It replaces softmax with a linear kernel

Answer

It reduces FLOPs from O(N^2) to O(N log N)

Question 3

In a Transformer, why are positional encodings necessary, and how does Rotary Position Embedding (RoPE) differ from absolute sinusoidal encodings?

Accepted Answer

RoPE rotates Q/K by position-dependent angles, encoding relative position in the dot product

Answer

RoPE adds a learned bias to logits; sinusoidal multiplies embeddings

Answer

Both are identical except RoPE uses cosine only

Answer

Positional encodings are unnecessary in decoder-only models

Question 4

What problem does Grouped-Query Attention (GQA) solve relative to Multi-Head Attention during inference?

Accepted Answer

It reduces the KV cache by sharing K/V heads across query-head groups

Answer

It increases the number of attention heads for accuracy

Answer

It removes the need for a KV cache entirely

Answer

It applies quantization to attention weights

Question 5

Explain why residual connections enable training of very deep networks.

Accepted Answer

They provide identity shortcuts so gradients flow and layers learn residuals

Answer

They reduce the parameter count of deep networks

Answer

They eliminate the need for activation functions

Answer

They force each layer to learn the full mapping from scratch

Question 6

What is the difference between LoRA and full fine-tuning, and why is LoRA parameter-efficient?

Accepted Answer

LoRA learns a low-rank update BA while freezing W, cutting trainable params

Answer

LoRA trains all weights but at lower precision

Answer

LoRA distills a large model into a smaller one

Answer

LoRA prunes attention heads to save memory

Question 7

Why does the Adam optimizer often converge faster than SGD, and what is a known generalization concern?

Accepted Answer

Adam adapts per-parameter rates for fast convergence but may generalize worse; AdamW decouples weight decay

Answer

Adam is always strictly better than SGD in generalization

Answer

Adam removes the need for a learning rate entirely

Answer

SGD uses second moments while Adam uses only first moments

Question 8

What is the vanishing gradient problem in RNNs and how do LSTMs mitigate it?

Accepted Answer

LSTMs add a gated additive cell state enabling stable gradient flow over time

Answer

LSTMs use deeper layers to avoid vanishing gradients

Answer

LSTMs remove recurrence entirely

Answer

LSTMs clip gradients automatically at each step

Question 9

In diffusion models, what does the forward process do and what is the network trained to predict?

Accepted Answer

The forward process adds noise; the network predicts the added noise (or score) to reverse it

Answer

The network predicts the original class label

Answer

The forward process denoises and the network adds noise

Answer

The network predicts the next pixel autoregressively

Question 10

What is classifier-free guidance (CFG) in diffusion models and what tradeoff does the guidance scale control?

Accepted Answer

CFG extrapolates between conditional and unconditional predictions; scale trades fidelity vs diversity

Answer

The scale controls learning rate during sampling

Answer

CFG requires a separately trained classifier

Answer

Higher guidance always improves diversity

Question 11

Why is layer normalization preferred over batch normalization in Transformers?

Accepted Answer

LayerNorm normalizes per-token over features, independent of batch/sequence statistics

Answer

LayerNorm is faster to compute than BatchNorm on GPUs

Answer

BatchNorm cannot be used with GPUs

Answer

LayerNorm requires running statistics at inference

Question 12

What is the dead ReLU problem and how do variants like Leaky ReLU or GELU address it?

Accepted Answer

Negative-region zero gradient can freeze neurons; Leaky ReLU/GELU keep nonzero gradients

Answer

Dead ReLU is caused by too-large output values; Leaky ReLU caps them

Answer

ReLU dies due to overflow; GELU uses lower precision

Answer

The problem only occurs at inference time

Question 13

How does Mixture-of-Experts (MoE) reduce compute relative to a dense model of equal capacity?

Accepted Answer

Only top-k experts activate per token, decoupling capacity from per-token FLOPs

Answer

MoE quantizes experts to 4-bit to save compute

Answer

MoE shares all parameters across experts

Answer

MoE removes the feed-forward layers

Question 14

What is the role of temperature in softmax sampling for text generation?

Accepted Answer

Temperature scales logits; lower sharpens, higher flattens the distribution

Answer

Temperature multiplies the final probabilities directly

Answer

Temperature controls the number of layers used

Answer

Temperature only affects training, not inference

Question 15

Why can nucleus (top-p) sampling be preferable to top-k sampling?

Accepted Answer

Top-p adapts the candidate set size to the distribution's confidence, unlike fixed-k

Answer

Top-p is deterministic while top-k is random

Answer

Top-p always selects more tokens than top-k

Answer

Top-p ignores token probabilities entirely

Question 16

What is gradient checkpointing and what tradeoff does it make?

Accepted Answer

It saves a subset of activations and recomputes others in backward, trading compute for memory

Answer

It trades memory for accuracy by dropping activations permanently

Answer

It checkpoints model weights to disk every step

Answer

It reduces compute by skipping the backward pass

Question 17

Explain the difference between data parallelism and tensor (model) parallelism in distributed training.

Accepted Answer

Data parallelism splits the batch with replicated models; tensor parallelism shards layer weights

Answer

Data parallelism shards weights; tensor parallelism splits the batch

Answer

Both split only the batch dimension

Answer

Tensor parallelism replicates the full model on each GPU

Question 18

What does ZeRO (in DeepSpeed) optimize, and what are its three stages?

Accepted Answer

ZeRO partitions optimizer states, gradients, then parameters across data-parallel ranks

Answer

ZeRO compresses gradients to int8 in three stages

Answer

ZeRO is a learning-rate schedule with three phases

Answer

ZeRO replaces backprop with forward-mode autodiff

Question 19

Why does label smoothing improve generalization, and what is its mechanism?

Accepted Answer

It softens one-hot targets, reducing overconfidence and improving calibration

Answer

It adds noise to inputs to augment data

Answer

It increases the learning rate near convergence

Answer

It removes the softmax layer

Question 20

What is knowledge distillation and why can a student trained on soft labels outperform one trained on hard labels alone?

Accepted Answer

Soft targets convey inter-class similarity ('dark knowledge'), a richer signal than one-hot labels

Answer

Soft labels reduce dataset size requirements only

Answer

The student copies the teacher's weights directly

Answer

Distillation only works for identical architectures

Question 21

In contrastive learning (e.g., SimCLR), what role does the temperature parameter in the InfoNCE loss play?

Accepted Answer

It scales similarities, controlling sensitivity to hard negatives

Answer

It sets the number of negative samples

Answer

It is the learning rate for the projection head

Answer

It determines the embedding dimensionality

Question 22

What is the difference between epistemic and aleatoric uncertainty in deep learning?

Accepted Answer

Aleatoric is irreducible data noise; epistemic is reducible model uncertainty

Answer

Epistemic is data noise; aleatoric is model uncertainty

Answer

Both are reduced by adding more data

Answer

They are identical for neural networks

Question 23

Why is mixed-precision (FP16/BF16) training used, and why is BF16 often preferred over FP16?

Accepted Answer

BF16 shares FP32's exponent range, avoiding loss-scaling issues of FP16

Answer

BF16 has more mantissa bits than FP16

Answer

FP16 never overflows, so it is always preferred

Answer

Mixed precision increases memory usage for accuracy

Question 24

What problem does the Straight-Through Estimator (STE) solve in training with discrete/quantized operations?

Accepted Answer

STE passes the discrete op forward but uses an identity-like gradient backward

Answer

STE removes the discrete op during training

Answer

STE replaces backprop with reinforcement learning

Answer

STE only works for continuous activations

Question 25

How does the attention mechanism's computational complexity scale with sequence length, and name one approach to reduce it.

Accepted Answer

O(N^2) in sequence length; reduced via sparse/linear attention or SSMs

Answer

O(N log N); reduced via dropout

Answer

O(N) already; no reduction needed

Answer

O(N^3); reduced by larger batch sizes

Question 26

What is the exposure bias problem in autoregressive sequence models trained with teacher forcing?

Accepted Answer

Train/test mismatch: training uses gold tokens but inference uses model predictions

Answer

The model sees too much noise during training

Answer

Exposure bias is caused by too-large batch sizes

Answer

It only affects encoder models

Question 27

In RLHF, what is the purpose of the KL-divergence penalty against the reference policy?

Accepted Answer

It constrains the policy near the reference to prevent reward hacking/drift

Answer

It speeds up convergence of the reward model

Answer

It replaces the reward model

Answer

It increases output diversity by maximizing KL

Question 28

How does Direct Preference Optimization (DPO) differ from PPO-based RLHF?

Accepted Answer

DPO optimizes preferences directly via a supervised loss, no separate reward model or online RL

Answer

DPO requires a larger reward model than PPO

Answer

DPO is an unsupervised pretraining method

Answer

DPO and PPO are mathematically unrelated

Question 29

What is the difference between weight decay and L2 regularization in the context of Adam?

Accepted Answer

In Adam, L2 gets rescaled by adaptive rates; AdamW decouples decay to fix this

Answer

They are always identical regardless of optimizer

Answer

Weight decay only applies to bias terms

Answer

L2 regularization is stronger than weight decay in all cases

Question 30

Why do learning-rate warmup schedules help when training Transformers?

Accepted Answer

Warmup ramps LR from near-zero to stabilize unreliable early-training dynamics

Answer

Warmup reduces the final learning rate to zero immediately

Answer

Warmup is only needed for CNNs

Answer

Warmup increases batch size gradually

Test yourself