Katkat
Back

Research Deep Dive: Model Details, Coding Habits and Diffusion Architecture

Wells FargoData Scientist · Technical Screen
HardData ScientistTechnical ScreenMachine Learning

Last updated

A research deep dive asks you to present a project under repeated follow-ups about the exact model you used and your coding habits, then explain how a diffusion model is built, trained and sampled. It tests ownership of your work, depth under probing, and command of noising, denoising networks and conditioning.

This interview for an applied AI research internship opened with two behavioral questions about your research, and both turned into technical deep dives. In the first, the interviewer followed up three times, asking about the specific model you used and about your coding habits. In the second, the interviewer returned to your research and asked about the architecture of a diffusion model.

Practice both: present one research project so that it holds up under repeated probing, then explain how a diffusion model is built, trained and sampled.

Clarifying Questions

  • Should the project come from the resume, and how long should the overview be before the questions start?
  • For the diffusion question, does the interviewer want the architecture of the model in your own project, or of a standard image diffusion model?
  • How much mathematical detail is expected: the training objective and noise schedule, or a component-level description?

Part 1 — Your research under repeated follow-ups

Walk through one research project: the problem, your contribution, the model you used, and the results. Expect up to three rounds of follow-up questions, each going one level deeper into the specific model (why this one, how it was configured, what you tried instead) and into how you wrote the code (structure, reproducibility, testing).

What This Part Should Cover

  • A crisp problem statement, your specific contribution, and measurable results
  • Model specifics: architecture choice, alternatives considered, key hyperparameters and ablations
  • The coding practices behind the work: structure, configuration, reproducibility, testing
  • Honest limitations and the next step

Part 2 — Diffusion model architecture

Explain the architecture of a diffusion model: how data is corrupted during training, what the network predicts, how the network is built and conditioned, and how samples are generated.

What This Part Should Cover

  • The forward noising process and its closed form
  • The training objective and what the network predicts
  • The denoising network: backbone, step embedding and conditioning
  • Sampling, and the trade-off between speed and quality

Hints

Hint 1

Prepare three levels down: For every claim in your summary, ask yourself "why?" three times in a row, and make sure you can answer each level from your own work rather than from the paper you built on.

Hint 2

Code is part of the research: Think about how someone else would rerun your experiment from your repository, and what in your code makes that possible or impossible.

Hint 3

Two processes: Separate the fixed process that adds noise from the learned process that removes it, and say which of the two has parameters.

Hint 4

What the network sees: List every input the denoising network receives at each step, and think about how a single number such as the step index can be fed into a convolutional or transformer network.

Follow-up questions

  • Why might a diffusion model be preferable to a GAN or a VAE for your problem, and what does it cost at inference time?
  • How does classifier-free guidance work, and what happens when the guidance scale is set too high?
  • How would you reduce the number of sampling steps without retraining from scratch?
  • A collaborator cannot reproduce your main result from your repository. How do you find out why?