Skip to content

ml-coding.md

Latest commit

 

History

History
90 lines (66 loc) · 7.24 KB

ml-coding.md

File metadata and controls

90 lines (66 loc) · 7.24 KB
Raw

2. ML/Data Coding 🤖

ML coding rounds vary by company. Some focus on implementing classical algorithms from scratch, while others test practical Python and PyTorch skills such as tensor operations, preprocessing, metrics, and training loops. In either format, interviewers evaluate correctness, numerical stability, code quality, edge-case handling, complexity, and your ability to explain design choices.

How to use this chapter

  • Use solutions/ml_algorithms.py as the canonical, executable NumPy reference for core interview problems.
  • Run solutions/test_ml_algorithms.py to verify the implementations and study useful edge cases.
  • Use the older notebooks as supplementary, exploratory material. Some predate the canonical solutions and may be less complete.
  • Practice writing each priority problem without looking at the reference, then compare correctness, complexity, and edge-case handling.

Difficulty and company tags

Every coding question uses a LeetCode-style difficulty label:

  • Easy: Usually one core operation, limited state, and a focused implementation that should take about 15 minutes.
  • Medium: Multiple steps or non-trivial edge cases; a strong interview implementation usually takes 20–35 minutes.
  • Hard: A 40–60+ minute problem involving several interacting components, advanced debugging, or systems-level trade-offs.

Difficulty is an editorial judgment about the complete prompt in this repository, not a claim that every source rates an equivalent problem identically. Exact matches were calibrated against TorchLeet and Deep-ML; unmatched prompts were rated with the same rubric. When sources disagree, the scope and required edge cases of this repository’s prompt determine the final label.

Company tags are added only when a reference associates a company with the same implementation problem. They are historical preparation signals, not a guarantee that the company currently asks the question. Only metadata facts are used; no third-party problem statements or solutions are copied.

Run the reference tests from the repository root:

uv run --with numpy python src/MLC/solutions/test_ml_algorithms.py

PyTorch ML Coding

Modern ML coding interviews may test practical PyTorch skills rather than only asking candidates to implement algorithms from scratch. The PyTorch ML Coding Problems guide includes:

  • A 60-minute mock interview covering tensors, preprocessing, metrics, training/evaluation loops, and debugging
  • Python utility and clean-code questions for ML workflows
  • Tensor operations, datasets, batching, device handling, and autograd
  • Training, optimization, mixed precision, checkpointing, and reproducibility
  • Testing, debugging, deployment, and advanced PyTorch questions
  • Coding challenges and concise reference responses

Priority ML coding problems

The following set combines classic questions that remain common with modern primitives increasingly expected in AI/ML interviews.

Problem Difficulty Company tags Canonical solution Supplemental notebook What a strong solution should cover
Numerically stable softmax and cross-entropy Easy Apple, Meta, Google, Amazon softmax, cross_entropy_from_logits — Max subtraction, log-sum-exp, shapes, class-index validation
Linear regression with gradient descent Medium — linear_regression_gradient_descent Linear regression Vectorized gradients, bias, MSE scaling, convergence
Logistic regression with gradient descent Hard Google, Meta, Amazon logistic_regression_gradient_descent Logistic regression Stable sigmoid, binary cross-entropy gradient, thresholds
k-nearest neighbors Medium Uber, LinkedIn, Meta knn_predict k-NN Pairwise distances, top-k selection, ties, complexity
k-means clustering Medium Uber, LinkedIn, Google, Amazon kmeans k-means Initialization, vectorized assignment, convergence, empty clusters
Decision-tree split Medium — gini_impurity, best_gini_split Decision tree Candidate thresholds, weighted impurity, stopping conditions
Principal component analysis Medium — principal_component_analysis — Centering, SVD/eigendecomposition, component ordering, variance
2D convolution Medium — conv2d_valid Convolution Output shape, stride, cross-correlation vs convolution
Scaled dot-product attention Medium — scaled_dot_product_attention — Q/K/V shapes, 1/sqrt(d_k), masking before stable softmax
Binary metrics and ROC-AUC Medium — binary_classification_metrics, roc_auc — Zero denominators, class imbalance, ties, rank interpretation
Reservoir sampling Medium — reservoir_sample — Unknown stream length, uniform probability, O(k) memory
TF-IDF Medium — tfidf — Token counts, document frequency, smoothing, sparse scaling

All canonical functions are in solutions/ml_algorithms.py.

Additional classic algorithms

These are useful follow-up exercises, especially when they match the target team's domain:

  • Linear SVM and hinge loss (notebook) — Difficulty: Medium
  • Perceptron learning rule (notebook) — Difficulty: Easy
  • Feedforward neural network and backpropagation (notebook) — Difficulty: Hard
  • Multiclass or multilabel extensions of metrics and losses — Difficulty: Medium
  • Naive Bayes for text classification — Difficulty: Medium
  • Matrix factorization for recommendation systems — Difficulty: Medium
  • Gradient boosting: explain the training loop and implement a simple residual-fitting step — Difficulty: Hard

Data and sampling questions

  • Implement train/validation/test splitting without leakage — Difficulty: Easy
  • Standardize features using training-only statistics — Difficulty: Easy
  • Handle missing values and unseen categories consistently — Difficulty: Medium
  • Implement uniform, stratified, weighted, and reservoir sampling — Difficulty: Medium
  • Build mini-batches and pad variable-length sequences — Difficulty: Medium
  • Aggregate sample-weighted losses and streaming metrics correctly — Difficulty: Medium

What to explain during the interview

  1. State input shapes, dtypes, assumptions, and expected outputs before coding.
  2. Start with a correct baseline, then vectorize or optimize the bottleneck.
  3. Discuss time and space complexity, including the cost of pairwise matrices.
  4. Handle numerical stability, empty inputs, ties, constant features, and invalid labels.
  5. Write small tests for normal cases and at least one failure or boundary case.
  6. Explain how the implementation would change for large datasets, GPUs, distributed training, or production libraries.