Interview questions in AI & ML, page 15
What RLHF is and the safety problem it solves
Collect human preference rankings, train a reward model, fine-tune the policy with PPO; it aligns outputs with human intent the loss function cannot specify.
GAN architecture: generator and discriminator roles
Generator maps noise to fake samples, discriminator classifies real versus fake, they train as a two-player game until samples fool the discriminator.
Design an automated system to diagnose model performance drop root causes
Tests causal attribution between pipeline bugs and drift. Strong answers sequence schema/null audits, feature drift via PSI/KS, then concept drift via holdout decay. Red flag: skipping pipeline checks to retrain immediately.
P-value vs confidence interval in an A/B test
A p-value gauges evidence against the null; a 95% CI shows plausible effect sizes and precision.
Red teaming LLMs and harmful output categories
Red teaming deliberately probes a model to elicit harmful behavior; categories include toxic or hateful content and dangerous instructions like weapons or malware.
Diffusion forward and reverse processes
Forward process gradually adds Gaussian noise until data is pure noise; reverse process learns to denoise step by step; the network predicts the noise added at each timestep.
How do you monitor thousands of per-customer models as a fleet?
Tests fleet-level statistical aggregation versus per-instance alerting. Strong answers propose tiered telemetry, cohort baselining for drift, and hierarchical alerting to prevent fatigue.
A/B test: 0.1% lift. Statistical vs practical significance?
Statistical significance says the 0.1% is real; practical significance asks if revenue exceeds engineering cost. Frame with CIs and ROI.
Stages of an end-to-end ML pipeline
Ingest, validate, preprocess and feature engineer, train, evaluate against a baseline, then register and package for deployment, with gates between stages.
Designing input and output guardrails for a chatbot
Input guardrails filter or classify user prompts (injection, off-topic, PII) before the model; output guardrails validate responses for toxicity, leakage, and policy before sending.
Mode collapse in GAN training
Generator produces few outputs ignoring data diversity, caused by chasing whatever fools the current discriminator; mitigate with minibatch discrimination, unrolled GANs, or Wasserstein loss.
Why not stop an A/B test when it looks significant early?
Tests whether you understand repeated looks inflate false positives. The term is peeking: checking daily can turn a 5% Type I error rate into roughly 15% by day 3. Red flag: citing "low sample size" without stating that early stopping invalidates the p-value.
Prompt injection versus jailbreak, and defenses
Injection hijacks the model via untrusted data overriding developer instructions; jailbreak coaxes a model past its safety policy. Defense: separate trusted instructions from untrusted data and filter.
Evaluating generative models with FID versus IS
FID compares Inception feature distributions of real and fake images via Frechet distance between two Gaussians; it uses real data as reference and detects diversity issues, unlike IS which uses no real…
What triggers automatic full retraining in an ML pipeline?
A strong answer lists four triggers: fresh data, code changes, model drift, and scheduled cadence.
How do network effects violate A/B tests and how to mitigate them?
Tests SUTVA violations and network experiment design. Answers note treated users alter control outcomes, then propose social-graph cluster randomization to isolate spillovers. Red flag: ignoring peer-to-peer spillover and using user-level randomization.
Model quantization benefits and risks
Quantization stores weights and activations in lower precision like INT8 or INT4; benefits are smaller memory and faster, cheaper inference; risk is accuracy loss.
Analyzing skewed revenue-per-user experiments
Heavy tails inflate variance and slow significance, and the mean is dominated by whales; mitigate via winsorization or capping, log transforms, CUPED variance reduction, or bootstrap and rank tests.
Why U-Net skip connections matter for denoising
Skips carry high-resolution spatial detail from encoder to decoder, preserving fine structure lost in downsampling and easing gradient flow, which lets the model restore detail while removing…
How would you implement automated data validation before training?
Enforce schema contracts, halt training on failure, quarantine bad batches, and alert owners.
We are hiring for this. Every open role lists the topics its interview covers, so you can prepare for the real thing rather than guessing.
See open roles