Free NVIDIA NCA-GENM Practice Exam with Questions & Answers | Set: 2

Pruning removes weights, neurons, or entire filters/channels that contribute minimally to model output — identified via magnitude-based criteria (removing near-zero weights), sensitivity analysis, or more sophisticated importance scoring — producing a smaller, sparser model that requires fewer computations and less memory at inference time while aiming to preserve accuracy through careful selection and, often, a fine-tuning step after pruning to recover any lost performance. Structured pruning (removing entire filters/channels) yields hardware-friendly speedups on standard accelerators, while unstructured pruning (removing individual weights) achieves higher sparsity ratios but requires specialized sparse-computation hardware or libraries to realize actual speed gains.

The remaining options move in the wrong direction or address a different concern: adding more convolutional filters (B) increases model capacity and parameter count, which increases computational cost, the opposite of the stated goal. Increasing batch size (D) affects training throughput and memory usage per step but does not reduce the per-sample computational cost of inference — a larger batch does more total work, not less per inference call, and batch size at inference is often constrained by latency requirements rather than optimization goals. Option C's premise — selectively replacing activation functions with simpler ones in "some neurons" — is not a standard or well-defined optimization technique; activation function choice is typically uniform within a layer and driven by training dynamics, not a piecemeal inference-cost lever.

Bias evaluation for facial recognition centers on measuring whether the model's accuracy — true positive rate, false positive rate, false match rate — is consistent across demographic subgroups (race, gender, age), rather than assuming a single aggregate accuracy figure represents performance fairly for all groups. This concern is well grounded empirically: independent benchmarking, including NIST's Face Recognition Vendor Test studies, has repeatedly documented substantial accuracy disparities across demographic groups in widely deployed facial recognition systems, with error rates for some subgroups measured many times higher than for others — a direct consequence of training data underrepresentation and the representativeness-bias issue covered elsewhere in this domain.

Processing speed (A) is a performance/latency engineering concern, not a bias or fairness concern — a model could process all demographic groups at identical speed while still exhibiting severe accuracy disparities between them. Facial expression recognition capability (C) addresses a different task dimension (emotion/expression classification) than the identity-recognition bias question being asked. Operating system compatibility (D) is a software-deployment/engineering concern entirely unrelated to model fairness.

Subgroup accuracy disparities in facial recognition carry serious real-world consequences — misidentification risk in law enforcement or access-control contexts — which is why disaggregated (per-subgroup) evaluation, not just aggregate accuracy, is considered a baseline requirement under Trustworthy AI bias auditing practice.

Before deciding *how* to handle missing data, best practice requires understanding *why* it's missing — analyzing whether missingness is Missing Completely at Random (MCAR, no systematic pattern), Missing at Random (MAR, related to other observed variables but not the missing value itself), or Missing Not at Random (MNAR, related to the missing value itself, e.g., patients with severe symptoms being less likely to complete a survey field). This diagnostic step determines which downstream handling strategy is statistically appropriate: naive row deletion under MNAR conditions can introduce systematic bias into the remaining dataset, while mean/median imputation applied blindly can distort variance and correlational structure if missingness isn't actually random.

Options B, C, and D each jump directly to a specific remedial action without first establishing whether that action is appropriate for the missingness pattern present. Removing rows (B) sacrifices sample size and can bias results if missingness correlates with the outcome of interest. Filling with a default value (C) without understanding the pattern risks introducing artificial structure that doesn't reflect the true underlying data. Removing entire columns (D) may discard genuinely informative features if missingness in that column is low or non-systematic.

Only after this initial pattern analysis should you select an appropriate strategy: listwise deletion, mean/median/mode imputation, model-based imputation (e.g., MICE, k-NN imputation), or explicit missingness indicators as additional features.

Data cleansing (or data cleaning) operates on data you already have: it identifies and resolves quality issues within an existing dataset — handling missing values (A), removing duplicate records (D), correcting formatting or type inconsistencies (B), fixing structural errors, and standardizing units or encodings. Collecting additional data (C) belongs to a conceptually earlier and separate phase of the pipeline: data acquisition or data collection, which determines what data enters the pipeline in the first place, rather than what is done to improve the quality of data already collected.

This distinction matters operationally: a cleansing step is typically deterministic and reversible against the existing dataset (you can inspect, log, and audit exactly which rows were dropped or imputed), whereas collecting more data is a scoping decision that may require new labeling budgets, new consent/privacy review, or new data-source integration — a materially different workflow with different stakeholders.

That said, insufficient data volume discovered *during* cleansing (e.g., after removing corrupted records the sample size becomes too small for the target class) can trigger a decision to go back and collect more — but that action itself is not classified as a cleansing step; it is the trigger for restarting an earlier pipeline stage.

The standard ML project lifecycle proceeds: data collection first, since you need raw data before anything else can happen; data preprocessing next, to clean, transform, and prepare that raw data (handling missing values, normalization, encoding, splitting into train/validation/test sets) into a form a model can consume; model training next, where the algorithm learns patterns from the preprocessed training data; and model evaluation last, where the trained model's performance is measured on held-out data it did not see during training. Each stage depends on the output of the one before it — you cannot preprocess data you haven't collected, train on data that hasn't been cleaned and split, or evaluate a model that hasn't been trained — which is what makes B the only internally consistent ordering among the four options.

Options A, C, and D each place a downstream step before its prerequisite: A attempts preprocessing before collection (nothing to preprocess yet); C and D both place evaluation before training and, in D's case, before data even exists — evaluation requires a trained model to assess, so it cannot logically precede training or the data-collection/preprocessing steps that training itself depends on.

In practice this pipeline is iterative rather than strictly linear — evaluation results often send you back to preprocessing (feature engineering) or even data collection (targeted collection to address weak subgroups) — but the canonical forward sequence for a first pass remains collection → preprocessing → training → evaluation.

Word Error Rate is the standard evaluation metric for ASR systems. It measures the edit distance between the model's transcription and a human reference transcript, computed as (Substitutions + Deletions + Insertions) / Number of reference words, expressed as a percentage. Lower WER indicates better transcription accuracy. Its character-level analogue, Character Error Rate (CER), is used for languages without clear word boundaries or for morphologically complex languages.

The distractors target common confusions: CTC (Connectionist Temporal Classification) Loss (A) is a *training* objective used to align variable-length audio input with variable-length text output in ASR models like DeepSpeech — it optimizes the model but is not itself a post-hoc evaluation metric on held-out accuracy. F1 Score (B) evaluates classification tasks with defined positive/negative classes, such as keyword spotting or wake-word detection, not full transcription. Mean Opinion Score (C) is a subjective, human-rated metric used to evaluate speech *synthesis* quality (TTS) or perceived audio naturalness — the inverse task of ASR — not transcription accuracy.

On NVIDIA's Riva and NeMo ASR pipelines, WER is the benchmark reported against datasets like LibriSpeech, and it remains the figure typically referenced in the exam's Multimodal Data and Experimentation domains when discussing speech model evaluation.