Day 1: Morning Track Selection

This quiz helps you pick the right morning track. It mixes knowledge questions with experience questions to give a balanced recommendation. Takes ~3 minutes.

1. What does loss.backward() do in PyTorch?

Decreases the loss value by one optimization step Computes gradients of the loss w.r.t. all parameters with requires_grad=True Updates the model weights using the computed gradients Resets the computation graph for the next forward pass

2. Drag the steps of a training loop into the correct order.

3. What problem does the Adam optimizer address compared to vanilla SGD?

Adam uses less memory than SGD Adam guarantees convergence to the global minimum Adam adapts the learning rate per parameter using running estimates of gradient moments Adam removes the need to set a learning rate entirely

4. Training loss keeps decreasing but validation loss starts going up after epoch 10. What is happening?

The learning rate is too low The model is underfitting The model is overfitting The optimizer has a bug

5. Match each architecture name to its PyTorch implementation.

Click a name, then click the code block it belongs to. Click a placed name to remove it.

6. Have you rented and worked on a remote GPU (RunPod, Lambda, AWS, HPC cluster, etc.)?

Yes, I'm comfortable with SSH, environment setup, and remote development I've done it but it was painful / I followed a tutorial No, only local / Colab

7. Have you used experiment tracking (WandB, TensorBoard, MLflow) or Python packaging tools (pip, uv, pyproject.toml)?

Yes, I regularly use experiment tracking and manage project dependencies I've used one or both but not extensively No experience with either

8. Have you dealt with multi-GPU training, distributed computing, or concurrent/multithreaded Python code?

Yes, I've used torch.distributed, multiprocessing, or similar tools I've encountered these topics but haven't used them hands-on No experience

9. Have you published a dataset or model weights (e.g. on HuggingFace, GitHub, or a similar platform)?

Yes, I've published artifacts with documentation for others to use I've shared code/weights informally but not as a proper release No

This is a recommendation, not a rule. If you feel strongly that the other track is a better fit, go for it.