Interview Q&A — NumPy & High-Performance Array Computing¶
30 High-Yield Questions & Model Answers for AI/ML and Data Science Technical Interviews.
Q1. Explain NumPy array broadcasting rules with mathematical conditions.¶
Answer: Two dimensions are compatible for broadcasting when: (1) they are equal, or (2) one of them is 1. Starting from the trailing (rightmost) dimensions, NumPy compares dimensions. A dimension of size 1 is virtually stretched to match the other array's shape without allocating extra memory.
Q2. What are NumPy strides and how do they determine memory indexing?¶
Answer: Strides are tuples of bytes to step in each dimension when moving to the next element. For a 2D float64 array of shape (M, N), row stride is N*8 bytes, and column stride is 8 bytes. Transposing an array simply swaps the strides without copying underlying memory buffer.
Q3. Explain the difference between C-contiguous and Fortran-contiguous array memory layouts.¶
Answer: C-contiguous (row-major) stores row elements consecutively in memory; advancing row index requires a large jump in memory address. Fortran-contiguous (column-major) stores column elements consecutively. C-contiguous aligns with C/Python indexing; Fortran with BLAS/Fortran libraries.
Q4. What is the difference between a view and a copy in NumPy?¶
Answer: A view shares the exact same data buffer as the original array (checked via arr.base is not None); modifying a view mutates the original. A copy allocates an entirely new independent memory buffer.
Q5. How does basic slicing differ from advanced (fancy) indexing in NumPy?¶
Answer: Basic slicing (e.g. arr[1:5, ::2]) ALWAYS returns a view. Advanced indexing (indexing with integer arrays, lists, or boolean masks like arr[[0, 2]]) ALWAYS returns a copy.
Q6. Explain the difference between 'np.dot()', 'np.matmul()', and the '@' operator.¶
Answer: 'np.dot()' performs matrix multiplication for 2D arrays but calculates inner products for higher-dimensional tensors. 'np.matmul()' and '@' treat higher dimensions as batches of 2D matrices, which is essential for deep learning mini-batch tensor multiplications.
Q7. What are universal functions (ufuncs) in NumPy?¶
Answer: Ufuncs are vectorized functions operating element-wise in compiled C, supporting broadcasting, type casting, and reduction methods like np.add.reduce() and np.add.accumulate().
Q8. Explain how boolean masking works in NumPy and how to combine masks.¶
Answer: Boolean masking evaluates element-wise conditions, returning a boolean array of identical shape. Combine multiple conditions using bitwise operators (& for AND, | for OR, ~ for NOT), wrapping each condition in parentheses: '(arr > 0) & (arr < 10)'.
Q9. What is the difference between 'np.concatenate()', 'np.stack()', 'np.vstack()', and 'np.hstack()'?¶
Answer: 'np.concatenate()' joins existing arrays along an existing axis. 'np.stack()' joins arrays along a NEW axis. 'np.vstack()' stacks along vertical row axis (axis 0). 'np.hstack()' stacks along horizontal column axis (axis 1).
Q10. What does 'np.vectorize' do and is it truly as fast as C-level vectorization?¶
Answer: 'np.vectorize' is a convenience wrapper providing a vectorized interface over a Python function. It does NOT compile Python code into C or accelerate execution; it simply runs a Python for-loop internally.
Q11. How does NumPy handle floating-point special values: 'np.nan', 'np.inf', and '-np.inf'?¶
Answer: They adhere to IEEE 754 standards. 'np.nan' represents Not-a-Number (np.nan == np.nan evaluates to False; use np.isnan()). Arithmetic with nan results in nan. 'np.inf' represents positive infinity from division by zero.
Q12. What is 'np.where()' and how does its three-argument form work?¶
Answer: 'np.where(condition, x, y)' is a vectorized ternary operator: where condition is True, yield x; where False, yield y. When called with only condition, it returns the indices where condition is True.
Q13. Explain the concept of memory alignment and SIMD operations in NumPy.¶
Answer: Modern CPUs use SIMD (Single Instruction, Multiple Data) instructions (AVX2, AVX-512) to execute mathematical operations on 4 to 8 floating-point numbers simultaneously in parallel hardware registers. NumPy contiguous memory buffers align with cache lines to maximize SIMD throughput.
Q14. What is 'np.einsum()' and why is it powerful in linear algebra and machine learning?¶
Answer: Einstein summation ('np.einsum') computes complex tensor contractions, matrix multiplications, traces, and batched dot products using index notation (e.g. 'ik,kj->ij' for matrix multiplication) without allocating large intermediate memory arrays.
Q15. How do you find unique elements and their counts in a NumPy array?¶
Answer: Use 'np.unique(arr, return_counts=True)'. It sorts the array and returns unique values and their frequencies in O(N log N) time.
Q16. What is the difference between 'arr.reshape()' and 'arr.resize()'?¶
Answer: 'arr.reshape()' returns a new view with the requested shape if compatible with total elements (or raises ValueError). 'arr.resize()' modifies the array in-place, truncating or padding with zeros if size changes.
Q17. How does 'np.pad()' work in computer vision and signal processing?¶
Answer: 'np.pad()' pads array boundaries with constants, reflection, or edge replication, widely used in convolutional layers to maintain spatial dimensions across image borders.
Q18. What is structured array in NumPy?¶
Answer: A structured array allows defining compound datatypes (like C structs) with named fields of differing types (e.g. ('age', 'i4'), ('salary', 'f8')) stored contiguously in memory.
Q19. Explain 'np.argsort()' and how it is used in ranking algorithms.¶
Answer: 'np.argsort()' returns the indices that would sort an array along a given axis, essential for ranking top-k predictions, nearest neighbors, and sorting parallel arrays simultaneously.
Q20. How do you compute pairwise Euclidean distance matrix using NumPy broadcasting without for-loops?¶
Answer: Given matrix X of shape (N, D), expand dimensions: 'diff = X[:, np.newaxis, :] - X[np.newaxis, :, :]' (shape N, N, D). The distance matrix is 'np.sqrt(np.sum(diff**2, axis=-1))'.
Q21. What is 'np.clip()' and when is it used in deep learning?¶
Answer: 'np.clip(arr, a_min, a_max)' constrains values within an interval. In deep learning, gradient clipping prevents exploding gradients, and probability clipping (e.g. clip to [1e-15, 1-1e-15]) prevents log(0) in cross-entropy loss.
Q22. How does 'np.tile()' differ from 'np.repeat()'?¶
Answer: 'np.repeat()' repeats individual elements consecutively ('[1, 2]' -> '[1, 1, 2, 2]'). 'np.tile()' replicates the entire array structure ('[1, 2]' -> '[1, 2, 1, 2]').
Q23. Explain 'np.argmax()' along an axis in classification models.¶
Answer: 'np.argmax(arr, axis=1)' returns the column index containing the maximum predicted probability for each sample row, converting probability distributions into discrete class label predictions.
Q24. What is memory leakage in NumPy arrays due to small views of huge arrays?¶
Answer: If you create a tiny slice from a massive 10 GB array ('small = huge[:5]'), the small view holds a reference to the entire 10 GB buffer ('small.base is huge'), preventing the 10 GB array from being garbage-collected. Fixed by forcing a copy: 'small = huge[:5].copy()'.
Q25. How do you perform element-wise random sampling in NumPy?¶
Answer: Use the modern 'np.random.default_rng()' generator: '.integers()' for discrete uniform, '.random()' for continuous [0, 1), and '.normal(mean, std, size)' for Gaussian sampling.
Q26. What is 'np.bincount()' and why is it preferred over np.unique for counting non-negative integers?¶
Answer: 'np.bincount(arr)' counts occurrences of each integer value from 0 up to max(arr) directly using an index-based bucket array, running in O(N) linear time compared to O(N log N) for np.unique.
Q27. How does 'np.memmap' handle datasets larger than physical RAM?¶
Answer: 'np.memmap' maps large binary files on disk directly into an ndarray view; the operating system page cache transparently loads slices into RAM on-demand, enabling manipulation of terabyte-scale tensors without Out-Of-Memory errors.
Q28. What is the difference between 'np.percentile()' and 'np.quantile()'?¶
Answer: 'np.percentile()' takes percentages in the range [0, 100]. 'np.quantile()' takes probabilities in the range [0, 1]. Both compute sample percentiles using linear interpolation.
Q29. How do you invert a square matrix in NumPy and when should you avoid it?¶
Answer: Invert using 'np.linalg.inv(A)'. In production, never invert matrices to solve Ax = b; solving 'np.linalg.solve(A, b)' using LU or Cholesky decomposition is twice as fast and mathematically much more numerically stable.
Q30. What is condition number of a matrix ('np.linalg.cond')?¶
Answer: The condition number measures how sensitive linear systems are to small errors in input data. A high condition number (ill-conditioned matrix) indicates that tiny perturbations in features produce huge fluctuations in linear regression weights.