ML Modeling Fundamentals: Overview and Study Plan

ML Modeling Fundamentals: Overview and Study Plan

This section is built for a hybrid ML Modeling & Fundamentals interview: explain theory from first principles, reason about modeling choices, and implement small NumPy/PyTorch snippets.

The target domain is autonomous driving simulation, but the ideas also transfer to robotics, ranking, perception, prediction, and model debugging.

What you should be able to do

After this section, you should be able to:

• Choose the right metric instead of optimizing blindly.
• Explain precision, recall, F1, PR-AUC, calibration, ADE/FDE, and ranking metrics.
• Explain why loss functions create incentives.
• Use weighted cross entropy, focal loss, and multi-modal trajectory losses.
• Explain diffusion from first principles: forward noise, reverse denoising, epsilon prediction.
• Explain why diffusion helps generate diverse future scenes.
• Evaluate simulation outputs for realism, safety, map validity, and physical feasibility.
• Debug a model that is not learning using a systematic checklist.
• Write small PyTorch snippets in CoderPad.

Recommended study order

1. Start with metrics

Read: ML Modeling Metrics

Why first: metrics tell you what “good” means. Without them, losses and models are easy to misuse.

Focus on:

• confusion matrix,
• precision vs recall,
• thresholding,
• PR-AUC for rare events,
• calibration,
• ADE/FDE limitations.

Notebook: ml_modeling_metrics_colab.ipynb

2. Learn custom losses

Read: Custom Losses

Why next: losses are how you make the model care about the right mistakes.

Focus on:

• weighted cross entropy,
• focal loss,
• why hard examples are not always good examples,
• why MSE fails for multi-modal futures,
• multi-modal trajectory loss.

Notebook: custom_losses_colab.ipynb

3. Learn diffusion basics

Read: Diffusion Basics

Why next: diffusion is easier once you understand loss design and multi-modal prediction.

Focus on:

• forward noising,
• reverse denoising,
• epsilon prediction,
• timestep conditioning,
• why sampling gives multiple plausible outputs.

Notebook: diffusion_basics_colab.ipynb

4. Apply diffusion to simulation

Read: Diffusion for Simulation

Why next: this connects the math to autonomous driving and robotics.

Focus on:

• conditioning on map/history/lights/route/intent,
• controllability,
• rare scenario generation,
• realism vs safety tradeoffs.

Notebook: diffusion_for_simulation_colab.ipynb

5. Evaluate generated scenarios

Read: Simulation Metrics

Why next: generated scenarios need different evaluation than ordinary prediction.

Focus on:

• collision,
• offroad,
• wrong-way,
• kinematic infeasibility,
• log divergence,
• realism vs safety.

Notebook: simulation_metrics_colab.ipynb

6. Finish with debugging

Read: Debugging a Model That Is Not Learning

Why last: debugging requires knowing what the model, loss, and metrics are supposed to do.

Focus on:

• overfit-one-batch test,
• data skew,
• bad labels,
• normalization bugs,
• learning rate problems,
• leakage,
• gradient inspection.

Notebook: debugging_model_not_learning_colab.ipynb

One-week study plan

Day 1: Metrics

Goal: never say “accuracy” blindly.

Tasks:

• Read the metrics article.
• Run the notebook.
• Explain precision and recall out loud using pedestrian crossing.
• Practice threshold tradeoffs.

Day 2: Custom losses

Goal: understand how losses change model incentives.

Tasks:

• Read the custom losses article.
• Run the notebook.
• Implement focal loss from memory.
• Explain why MSE fails for multi-modal futures.

Day 3: Diffusion basics

Goal: explain DDPM without paper notation overload.

Tasks:

• Read diffusion basics.
• Run the noising notebook.
• Write the equation for $x_t$ from memory.
• Explain why predicting epsilon is convenient.

Day 4: Diffusion for simulation

Goal: connect diffusion to future-scene generation.

Tasks:

• Read diffusion for simulation.
• Run the conditional trajectory notebook.
• Explain map/history/light/route conditioning.
• Discuss rare scenario generation tradeoffs.

Day 5: Simulation metrics

Goal: evaluate generated driving scenarios rigorously.

Tasks:

• Read simulation metrics.
• Run the metrics notebook.
• Implement ADE/FDE and kinematic checks.
• Explain realism vs safety.

Day 6: Debugging

Goal: have a reliable debugging playbook.

Tasks:

• Read debugging article.
• Run the overfit-one-batch notebook.
• Practice diagnosing train/eval loss patterns.
• List autonomous-driving-specific data bugs.

Day 7: Mock interview loop

Goal: combine theory, modeling judgment, and code.

Practice prompts:

• “Why is accuracy bad for rare event detection?”
• “Implement focal loss.”
• “Why does MSE fail for multi-modal trajectories?”
• “Explain diffusion forward and reverse processes.”
• “How would you condition a diffusion model on map and traffic lights?”
• “How do you evaluate generated scenarios?”
• “Your model is not learning. Walk me through your debugging process.”

60-minute cram plan

If you only have one hour:

0-10 min   Metrics: precision, recall, PR-AUC, calibration
10-20 min  Losses: weighted CE, focal loss
20-30 min  Multi-modal trajectory loss and MSE failure
30-40 min  Diffusion basics: x_t equation and epsilon prediction
40-50 min  Simulation: conditioning and scenario metrics
50-60 min  Debugging: overfit-one-batch and gradient inspection

CoderPad checklist

Be ready to implement:

• confusion matrix metrics,
• precision/recall/F1,
• focal loss,
• weighted cross entropy call,
• multi-modal trajectory loss,
• ADE/FDE,
• kinematic speed/acceleration checks,
• overfit-one-batch training loop,
• gradient norm helper,
• diffusion q_sample.

Interview stance

A strong answer usually has this shape:

1. Define the task and the cost of mistakes.
2. Choose metrics that reflect that cost.
3. Choose a loss that optimizes toward the metric.
4. Discuss tradeoffs and failure modes.
5. Give a small implementation.
6. Say how you would debug it.

That structure works for most ML modeling questions in autonomous driving simulation.