Free NCA-ADS practice test: 30 questions with answers

A large gap between training and validation accuracy indicates overfitting. Underfitting shows poor training performance, while encryption and network congestion are unrelated.

Randomized or Bayesian search with parallel evaluation covers the space efficiently. Exhaustive grids are costly, manual daily changes are slow, and single random configurations are not tuned.

RMSE squares errors and penalizes large deviations more. Accuracy, precision and silhouette scores apply to classification or clustering.

The silhouette score measures cluster cohesion and separation without labels. The other metrics require labelled targets.

Precision and recall focus on the rare positive class, and precision-recall curves reflect performance well on imbalanced data. Overall accuracy can be misleading, feature counts and training time do not measure detection quality.

Feature importance and attribution methods explain influence on predictions. Hiding inputs, random reordering and removing features do not explain the model.

Time-based splits respect temporal order and prevent leakage. Random shuffling, reversed time splits and training on test periods leak or distort evaluation.

Validation-based early stopping and regularization reduce overfitting. Training regardless of validation, using the test set for decisions and duplicating rows make overfitting more likely.

Dask partitions dataframes across workers, including GPU workers with cuDF. Laptops and deleting data are not scaling approaches, and converting to one number discards information.

Consistent imputation such as median fills missing values for both training and inference. Removing columns without checking, zero replacement regardless of meaning and assuming models handle missing values all cause problems.

One-hot or ordinal encoding represents categories numerically. The other options lose or distort information.

Fitting preprocessing only on training data prevents leakage. Fitting on all data before splitting and using the test set for tuning leak information.

Vectorized operations reduce launch overhead and use the GPU efficiently. More small operations, text editors and zero memory are not improvements.

Compatible key types and sensible partitioning improve join efficiency. Random key types, unrelated sorting and text exports slow the operation.

GPU speed and a pandas-like API ease adoption. They do not remove validation, always reduce memory, or make data more accurate.

GPU groupby operations on properly typed columns are efficient. Python loops, long text keys and spreadsheets are slower or impractical.

cuDF provides a pandas-like dataframe API that executes on GPUs. cuBLAS is a linear algebra library, TensorRT optimises inference, and Triton serves models.

cuML offers GPU-accelerated classical ML algorithms with a scikit-learn-like API. cuDNN accelerates deep learning primitives, NCCL handles collective communication, and NVENC is a video encoder.

MIG partitions supported GPUs into isolated instances. NVLink connects GPUs, a monitor and keyboard do not partition GPUs.

GPUDirect Storage moves data more directly between storage and GPUs, reducing CPU involvement. The other options are obsolete or unsuitable.

Nsight tools profile GPU activity and kernels. Riva is for speech, NeMo Guardrails for LLM rules, and Clara Parabricks for genomics.

Frequent transfers and many tiny kernel launches reduce GPU efficiency. Large vectorized operations and well-sized batches typically improve performance.

Contiguous column access improves memory throughput on GPUs. Random cell access, images and row text files are slower for this purpose.

Histograms and density plots summarise distributions efficiently. Pie charts with thousands of slices, full row tables and maps without geography are inappropriate.

A multi-series line chart compares trends across groups over time. A single pie chart, a single-observation box plot and a word cloud do not show trends effectively.

Aggregating or sampling reduces the points rendered while preserving patterns. Rendering every point, printing and enlarging the chart do not solve performance.

Accessible colours, clear labels and text summaries improve visualizations. Meaningless rainbow colours and overlapping labels reduce clarity.

A KPI tile with target comparison and trend communicates status quickly. Raw tables, 3D pie charts and word clouds do not show performance against target clearly.

A sorted bar chart compares values across categories clearly. Lines between unrelated categories, many-slice pies and scatter plots of names are poorly suited.

Transparency and density binning reveal concentrations in dense plots. Larger markers worsen overlap, removing labels reduces clarity, and plotting only the first hundred points discards most data.