Advanced interview questions in AI & ML, page 6
Why Transformers use multi-head attention
A single head averages into one representation subspace; multiple heads attend in parallel to different subspaces, letting the model capture diverse relations like syntax and coreference at once, then…
Mode collapse in GANs and how to fix it
Mode collapse is the generator producing few similar outputs, missing data diversity to fool the discriminator; fixes include Wasserstein loss, minibatch discrimination, unrolled GANs, and feature…
How does MAML's inner and outer loop work?
Inner loop does task-specific gradient steps from shared init, outer loop updates the init for fast adaptability via second-order gradients.
How do BYOL and Barlow Twins avoid representation collapse?
Collapse is embeddings shrinking to a constant or low-rank subspace; BYOL uses predictor plus momentum target plus stop-gradient, Barlow Twins decorrelates feature dimensions.
Three techniques to cut LLM inference latency
Quantization shrinks weights with small accuracy risk, KV-cache plus continuous batching boost throughput, speculative decoding drafts tokens for lossless speedup.
Design a cryptographically verifiable ML audit trail from dataset to deployment
Tests cryptographic provenance and tamper-evident ML pipelines. Strong answers cover content-addressed datasets, signed training logs linking code and hyperparameters to model hashes, and deployment signature checks.
Design a defense-in-depth strategy against adversarial evasion on a deployed image classifier
Proactive: adversarial training, preprocessing, ensembles.
Diagnosing Spark executor OutOfMemoryError
Check the Spark UI for skew and spills, inspect executor memory and partition count, find culprits like wide collect, huge shuffles, or skewed keys, and fix via more partitions, memory tuning, or…
How does a plenoptic camera enable post-capture refocus?
Light field records ray direction plus position via a microlens array, refocusing shears and integrates the 4D data to choose a virtual focal plane.
Why merge a burst instead of one long low-light exposure?
Short frames avoid motion blur and clipping while averaging cuts noise; align frames and merge robustly, rejecting misaligned regions to avoid ghosting.
How does focus stacking build an all-in-focus image?
Align frames, measure local sharpness as a focus metric, select or blend the sharpest source per region, then composite.
Design a showback or chargeback system for ML infrastructure costs
Tag workloads to cost centers; define shared-resource formulas; automate reconciliation; use showback.
Describe an ML workflow with massive egress fees and re-architecture to mitigate
Tests whether you recognize egress spikes when storage and compute cross cloud or region boundaries. Great answers sketch a multi-cloud training pipeline, cite per-GB rates, and propose caching or compute placement. Red flag: suggesting compression alone.
Design a cashierless retail checkout vision system.
Multi-camera coverage, person tracking and re-ID, product detection plus action recognition for take or return, sensor fusion, edge-cloud split, and failure handling.
Design drift detection and retraining for a CV model.
Monitor input and prediction distributions plus delayed labels, detect data and concept drift, trigger an automated retrain-validate-deploy pipeline with versioning and rollback.
Zero-downtime model updates with blue-green or canary
Blue-green swaps full traffic after validation, canary ramps a small slice; both need health, latency, and quality monitoring plus instant rollback.
Minimizing model prediction latency end to end
Shrink the model via quantization, pruning, distillation, compilation; cut infra latency via caching, edge, faster runtimes.
Differential privacy, epsilon, and noisy aggregates
Define DP as bounded output change when one record is added or removed, explain epsilon as the privacy-accuracy knob, add calibrated noise scaled to sensitivity.
Federated learning architecture, risks, and defenses
Devices train locally and send updates not data, a server aggregates; gradients still leak, enabling inversion or membership inference; defend with secure aggregation and DP.
How do agents use tool-calling and what can go wrong?
This tests your grasp of practical agentic architectures and their real-world trade-offs. A great answer distinguishes between predefined "workflows" and dynamic "agents," explains how an augmented LLM selects tools, and then details failure modes like framework obfuscation, debugging complexity, and the high latency/cost of multi-step processes. A red flag is vaguely describing agents without separating these patterns or ignoring the significant debugging and cost challenges.
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