Банк вопросов из реальных собеседований

Explain how dropout behaves during training and inference. Why does the implementation need scaling, and what is inverted dropout?

What is a receptive field? Compare one 5x5 convolution with two sequential 3x3 convolutions.

Name methods specific for neural networks to combat retraining. What methods of data augmentation can be used in Computer Vision?

How do you fight overfitting, what risks come with synthetic data, and why do you need a validation set?

What is semantic segmentation? How is it different from instance segmentation? How does the convolution work at CNN?

What is tokenization and why is it needed in text processing tasks? What are the differences between Bag of Words, TF-IDF, and embeddings like word2vec?

The interviewer asks: When considering logistic regression, how is it similar to and different from linear regression?

How does bagging differ from boosting, which uses random forest and gradient boosting, and what can we say about the tendency to overfit?

How are bagging and gradient boosting different? What happens if you remove one tree from Random Forest versus gradient boosting, and how can trees overfit rare categorical features?

Why does Random Forest usually reduce variance compared to a single tree and what trade-offs remain?

How does CatBoost encode categorical features and why does it not yield leakage?

For the task of matching food photos and categories, what should you choose: a CLIP approach or a supervised multilabel classifier?

How does the CLIP model work and how does it learn? What is the difference between SigLIP and classic CLIP?

There is a model for first-person VR/fisheye, and you need to work on flat third-person video. How do you transfer quality?

Why can feature importance cheat if features are highly correlated?

How do graph-based recommenders such as GCN and GraphSAGE use user-item graphs, and what is the high-level difference between them?

In object detection, why is loss based on IoU often used for frame regression, rather than the usual MSE for frame angle coordinates?

How to make recommendations for matching clothes: embeddings, category restrictions, nearest neighbors and reranker?

When to choose a linear model, and when to choose a tree-based model or boosting?

How to build logo detection for many brands and logo variants, if manual bbox markup is expensive?

Describe the mechanism of LoRA. If LoRA and classic adapters give the same quality, what should I choose for inference?

What is LoRA? Why is it useful for training large models? How do you usually initialize low-ranking matrices to avoid distorting the base model at the start?

How does LoRA differ from full fine-tuning, what are the limitations of multi-GPU learning LLM and how does the RLHF approach differ from the classical RL?

How to deal with model overfitting? Explain L1/L2, dropout 0.5 and what happens with dropout to inference.