Перейти к содержанию

Deep Learning: Подготовка к интервью

~3 минуты чтения

Предварительно: Учебные материалы DL | Математика для ML

150+ вопросов с развернутыми ответами, покрывающих deep learning от BNN до Mamba. По статистике ML-интервью 2025 (Blind, levels.fyi), 70% DL-вопросов приходятся на: backpropagation, attention/transformers, optimizers, normalization и CNN/RNN. Остальные 30% -- emerging topics (MoE, SSM, distillation, quantization). Формат: Q/A с формулами и кодом.

Типичные вопросы с собеседований 2025-2026 Обновлено: 2026-02-14

Тематические разделы

1. Основы DL

BNN & Uncertainty, Loss Functions, Backpropagation, Optimizers, Weight Initialization, Normalization.

Ключевые вопросы: Vanishing/exploding gradients, Adam vs SGD, BatchNorm vs LayerNorm, Xavier/He init, KL divergence, focal loss.

2. Архитектуры

CNN, RNN/LSTM, Attention & Transformers, Positional Encodings, Senior+ вопросы (KV-Cache, Flash Attention, Gradient Checkpointing).

Ключевые вопросы: Depthwise separable convolutions, LSTM gating, multi-head attention, RoPE vs ALiBi, MoE routing, Flash Attention IO-awareness.

3. Vision & Detection

ViT, Swin Transformer, Object Detection (YOLO, Faster R-CNN, FPN), Contrastive & Self-Supervised Learning (SimCLR, CLIP, MoCo, BYOL, MAE).

Ключевые вопросы: ViT patch embedding, shifted window attention, anchor-free vs anchor-based, FPN multi-scale, contrastive loss, InfoNCE.

4. Генеративные модели

Diffusion Models (DDPM, DDIM, CFG, Latent Diffusion, Consistency Models, U-Net, DiT), VAE (ELBO, Reparameterization, Posterior Collapse, beta-VAE), GAN (Mode Collapse, WGAN, DCGAN).

Ключевые вопросы: Forward/reverse diffusion, classifier-free guidance, latent vs pixel diffusion, ELBO derivation, reparameterization trick, Wasserstein distance.

5. GNN & Reinforcement Learning

Graph Neural Networks (Message Passing, GCN/GAT/GraphSAGE, Over-smoothing, GIN, Heterogeneous Graphs), Reinforcement Learning (Q-Learning, DQN, Policy Gradient, PPO, SAC, Actor-Critic).

Ключевые вопросы: Message passing framework, WL test expressiveness, over-smoothing solutions, Bellman equation, policy gradient theorem, PPO clipping.

6. Обучение & Оптимизация

Second-Order Optimization, Mixed Precision Training, Gradient Debugging, Gradient Flow Analysis, Gradient Checkpointing, Advanced Regularization (DropPath, Mixup, Label Smoothing).

Ключевые вопросы: L-BFGS vs Adam, FP16/BF16 tradeoffs, loss scaling, gradient clipping strategies, activation recomputation overhead, stochastic depth.

7. Компрессия & Transfer Learning

Model Pruning (Magnitude, Structured, Lottery Ticket), Knowledge Distillation, Transfer Learning & Fine-tuning (LoRA, Adapters), Weight Tying.

Ключевые вопросы: Structured vs unstructured pruning, temperature scaling in distillation, catastrophic forgetting, LoRA rank selection, pseudo-inverse tying.

8. Специальные темы

Time Series DL (DeepAR, TFT), Uncertainty Quantification (MC Dropout, Ensembles, Conformal Prediction), t-SNE & UMAP, Speculative Decoding, Mamba & SSM, TCN & WaveNet.

Ключевые вопросы: Temporal Fusion Transformer, aleatoric vs epistemic uncertainty, t-SNE perplexity, speculative decoding acceptance, S4 vs Mamba selectivity, dilated causal convolutions.

См. также