Machine learning interviews punish shallow textbook answers

Someone asks how you would detect overfitting. You mention the training and validation gap, then the interviewer gives you a small noisy table and asks which regularizer you would try first. The holdout looks fine, live precision falls after a week and the neat textbook answer suddenly has to survive contact with a product.

The AI/ML track lets you practice the parts that match your role. The catalog starts with linear algebra, probability, bias-variance, supervised learning, trees, feature work and evaluation. Later sessions cover neural networks, CNNs, sequence models, transformers, NLP, recommendation systems, reinforcement learning, generative and multimodal models, and graph learning. One onsite will cover a slice of that list, so use the posting to choose your slice.

Start with the data and the baseline

Model choice should follow the shape of the problem. A small tabular dataset with noisy labels may favor a regularized linear model. Interactions in a wide table can make trees attractive. Images bring spatial structure, while sequence length and timing change the options for text, audio or time series.

Interviewers often ask for a baseline because it exposes whether you can measure progress. Define the simplest useful model, the split and the metric before proposing a larger architecture. Then discuss leakage, label quality, imbalance and the way examples arrive in production. Those details decide whether a benchmark improvement means anything.

Bias and variance become easier to discuss once you use a concrete dataset. Say what error you expect from an overly simple model and what instability you expect from a flexible one trained on limited data. Those choices form the center of the Bias-Variance and Regularization interview. Dropout on a 2-layer network with 800 rows is a strange reflex when L2, early stopping or a simpler model can address the problem with less ceremony.

A metric is an operating decision

A fraud model can predict the majority class all day and still look accurate. Precision and recall expose different costs, and calibration matters when a score drives a threshold or a human review queue. Explain who pays for each kind of error. The Model Evaluation interview follows that cost into the threshold choice. The metric becomes defensible once those costs are attached to an action.

PR curves and ROC curves answer different questions under imbalance. F1 compresses precision and recall into one number, which can hide a preference the product needs to state. In an interview, pick a metric, explain its limits and show how you would choose a threshold. Naming every metric in scikit-learn sounds busy while leaving the decision untouched.

Production evaluation adds drift and feedback. A random holdout can look excellent when the deployed data arrives by time, geography or customer cohort. Discuss temporal splits, subgroup checks and the delay between prediction and label. If the live metric moves, inspect data quality and population change before assuming the model forgot how to predict.

Deep learning questions still need a purpose

For vision roles, expect filters, receptive fields, pooling and the difference between classification, detection and segmentation. Explain how transfer learning changes the plan when labeled images are scarce. A sentence about fine-tuning ResNet needs the next sentence about augmentation, frozen layers, evaluation and the failure cases you would inspect.

Transformer questions often connect attention, positional information and memory costs. Place Q, K and V clearly, then explain what changes as the context grows. The Transformer Internals interview stays with that explanation until the memory cost is explicit. Repeating the title of a famous paper will not answer why positional encoding exists or why attention becomes expensive.

Generative-model questions depend heavily on the role. GANs bring adversarial training and mode collapse. VAEs trade reconstruction against a structured latent distribution, while diffusion and autoregressive models use different generation procedures. You can discuss these faithfully without pretending to have trained a foundation model last quarter.

Specialized systems change the interview

Recommendation interviews usually separate candidate generation from ranking. Discuss implicit feedback, cold start, offline metrics and feedback loops. The product can distort its future training data by repeatedly showing the same popular items, so an evaluation plan has to account for exposure.

Reinforcement learning questions begin with the state, action, reward and policy. A poorly specified reward can produce behavior that technically scores well while violating the product’s intent. Work through who designed the reward, how exploration is constrained and how off-policy data affects evaluation.

Multimodal and graph roles deserve focused preparation only when the posting asks for them. For a vision-language role, be ready to discuss encoders, fusion, paired data and evaluation across modalities. For graph work, connect message passing and neighborhood structure to the problem. Reciting model families leaves the design question unanswered. Breadth has a cost, and shallow answers across 12 topics rarely beat a careful answer in the 4 the team uses.

Practice the follow-up chain

Take one dataset you have touched and rebuild the decision sequence. What was the target? How was the split made? Which baseline ran first? Why did the chosen metric fit the product, and what changed when the model met new data? A class project counts if you can discuss its weak labels and awkward scraping code honestly.

Have someone interrupt with constraints. Cut the labels in half. Raise the cost of false positives. Add a group that was absent from training or require inference on a phone. Each constraint should force a change in the model, metric or data plan. This is much closer to an interview than another evening of algorithm flashcards.

Pick the relevant session in the AI/ML track and keep following each answer with a constraint. The goal is a choice you can defend with incomplete information. Walk in ready to draw the dataset and baseline, then name the exact metric you would trust when the clean holdout meets Monday’s production traffic.