MLE ML Knowledge Orals (Set A)

MLE / AR loop opener. 60-minute verbal round covering ML fundamentals: bias-variance, parameter vs hyperparameter, batch vs real-time inference, large-data training, feature engineering, end-to-end ML lifecycle. Closes with the candidate walking through one of their own ML projects. Requirements The interviewer cycles through 5-8 ML-fundamentals questions, then asks the candidate to walk an end-to-end project. Recurring question themes: Bias-variance trade-off — define both, give an example of high-bias and high-variance failure, discuss how regularisation moves a model along the trade-off. Parameter vs hyperparameter — distinction by who sets the value (learning algorithm vs human) and how each is tuned. Batch vs real-time inference — latency, throughput, infrastructure, when to use each. Handling large training data — distributed training (data parallelism, model parallelism), gradient accumulation, mixed precision, sharding. Feature engineering — why it matters even in the deep-learning era, examples of high-impact features in financial fraud / credit risk. ML lifecycle end-to-end — problem framing → data collection → labelling → feature engineering → model selection → training → evaluation → deployment → monitoring → retraining trigger. Walk through an end-to-end project of your own — depth matters more than breadth; the interviewer probes one stage in detail. Notes The round runs as a conversation, not a quiz; the interviewer waits for short, structured answers (90 seconds each), then asks one follow-up. Bias-variance: the cleanest answer ties it to model capacity — high bias = underfit (linear model on a curved relationship), high variance = overfit (high-degree polynomial on noisy data). Mention the U-shape of test error vs capacity and the bias² + variance + noise decomposition. Large-data handling: the standard answer covers data parallelism with all-reduce (synchronous) or parameter server (asynchronous), gradient accumulation when single-batch memory is tight, mixed precision (fp16 / bf16) for throughput, and sharding the dataset across workers. Mention Horovod, PyTorch DDP, or DeepSpeed by name to signal hands-on experience. Feature engineering: have one concrete example ready, ideally from financial-services context (target encoding for high-cardinality categoricals, time-windowed aggregations for transaction-level features, ratio features for credit utilisation). The interviewer cares about practical examples, not generic platitudes. ML lifecycle: name all stages briefly, then pause for the interviewer to pick one to deep-dive. Common choices for the deep-dive: feature engineering, evaluation metric selection, deployment / monitoring. For the project walk, pick a project where you can defend every architectural decision. The interviewer will probe one stage hard — often the eval metric ("why AUC and not precision at top-K?") or the failure mode ("what was your model's worst slice?"). Preparation Build a 90-second canonical answer for each of the recurring themes. Memorise the structure (definition, trade-off, example, when-to-use), not the words. Practise the project walk against a friend. Force them to probe one stage at random; rehearse the deep-dive on at least three different stages. Have one concrete financial-services flavored example for feature engineering and one for evaluation metric choice. These differentiate from generic ML interviewing prep.