Explain deployment, retrieval, and regularization

You are interviewing for a machine-learning position at a large short-video platform. Respond to the conceptual questions below.

1. Given strict limits on GPU compute and VRAM, how would you serve a multimodal model for video retrieval or ranking? Cover architecture decisions, compression, batching, caching, and the trade-offs among quality, p99 latency, throughput, and serving cost.
2. Assume captions and video embeddings are already precomputed and persisted. How would you speed up online video retrieval? Address indexing, approximate nearest neighbor search, hybrid text-plus-vector retrieval, reranking, memory usage, and freshness.
3. Define overfitting, explain how you would detect it, and describe how you would reduce it in deep learning systems.
4. Explain the intuition and math behind Dropout, including why inverted Dropout preserves the expected activation scale from training to inference.
5. Compare common normalization approaches such as BatchNorm, LayerNorm, GroupNorm, and RMSNorm. When is each suitable, and how are their statistics handled during inference?
6. Explain how reinforcement learning is applied in LLM post-training, particularly RLHF. Describe the roles of supervised fine-tuning, preference data, reward modeling, policy optimization, KL regularization, and typical failure modes.

Overview: This question assesses machine learning systems engineering skills, including multimodal model deployment under GPU/VRAM limits, scalable video retrieval and indexing, regularization and normalization techniques, and reinforcement learning–based post-training for language models.