Machine Learning with cuML and Feature Engineering
Free NVIDIA-Certified Associate: Accelerated Data Science practice — 6 questions on Machine Learning with cuML and Feature Engineering, with explanations. No sign-up. Full 12-question mixed test →
A dataset contains clusters with noise points and significantly varying densities across clusters — some tight and dense, others sparse and spread out. The team needs GPU-accelerated clustering that does not require a single global density threshold and can still identify noise. Which cuML algorithm is the BEST fit?
cuML's HDBSCAN builds a hierarchy of clusters using varying density thresholds and extracts stable clusters automatically, making it robust to clusters of different densities while still labeling noise points, unlike single-eps DBSCAN.
In cuML's RandomForestClassifier, which parameter directly controls the number of CUDA streams used to build multiple trees concurrently on the GPU, most directly impacting training-time GPU utilization?
n_streams sets how many CUDA streams cuML uses to build trees in parallel on the GPU; increasing it can improve GPU utilization during forest construction (subject to available GPU resources).
A data science team needs to train a logistic regression model on a dataset containing 800 rows and 12 features. They are debating whether to use cuML's LogisticRegression on a GPU instance or scikit-learn's LogisticRegression on CPU. Which statement BEST describes the expected outcome?
cuML's performance advantage scales with data size; for very small datasets (hundreds of rows), host-to-device transfer and kernel launch overhead can dominate, making cuML comparable to or even slower than scikit-learn on CPU. The exam tests recognizing that GPU acceleration benefits large-scale data, not small workloads.
A feature engineering pipeline includes a categorical column (postal code) with roughly 30,000 unique values. The team wants to encode this feature for a GPU-accelerated model without causing a massive increase in dimensionality. Which cuML preprocessing tool is the BEST choice?
cuML's TargetEncoder replaces high-cardinality categorical values with statistics derived from the target variable, producing a single informative numeric column per feature without exploding dimensionality — the standard recommended approach for very high-cardinality categoricals.
By default, what determines the data type (e.g., NumPy array, cuDF DataFrame, CuPy array) of predictions returned by a fitted cuML estimator's .predict() method?
cuML estimators track the type of data passed into fit/predict and, by default, return outputs in that same type (input-mirroring behavior), unless the estimator's output_type parameter or cuml.set_global_output_type() is explicitly set to force a different type.
A team wants to perform hyperparameter tuning for a cuML model at scale, distributing the search across multiple GPUs in a cluster rather than running it on a single GPU. Which approach correctly achieves this within the RAPIDS ecosystem?
Dask-ML integrates with dask_cudf and cuML's Dask-enabled estimators to distribute hyperparameter search (e.g., grid or randomized search) and training workloads across multiple GPUs/nodes, which is the standard RAPIDS pattern for scaling tuning beyond a single GPU.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$99.99 $34.99 with code FREETEST33 — valid through September 7.
Get my $34.99 deal →