sp26-cs639 cs639
Practice Midterm 3 Practice Midterm 3
Content locked
The catalog is public. Reading this content requires approved access.
Sign in to readPractice Midterm 3
1. [ML Foundations] Why Multitask Learning Can Fail
2. [ML Foundations] Cross-Entropy Loss Computation
3. [ML Foundations] Test Data Contamination
4. [Deep Learning: CNN] Output Dimensions After Convolution
5. [Deep Learning: RNN] Vanishing Gradient and LSTMs
6. [Deep Learning: GNN] Graph Mixing and Semi-Supervised Learning
7. [Self-Supervised Learning] Teacher-Student EMA Update
8. [Scaling Laws] Skill-Dependent Compute Optima
9. [Transformers: Attention] Scaling by \sqrt{d_k}: Computation
10. [Transformers: Architecture] Division of Labor Between Self-Attention and FFN
11. [Transformers: Inference] Why Autoregressive Inference Is Sequential
12. [BERT] Role of the [CLS] Token
13. [BERT] Why NSP Is Not Strictly Necessary
14. [GPT] Autoregressive Probability and Per-Token Loss
15. [GPT-3 / Llama] Memory Requirements for Large Models
16. [Attention Variants] Linear Attention Trade-off
17. [FlashAttention] Recomputation Speeds Up Training
18. [SSMs] S4 vs Mamba for Selective Tasks
19. [Multimodal: ViT] Why Patch Tokenization
20. [Multimodal: CLIP] Zero-Shot Classification Mechanism
21. [Multimodal: Flamingo vs LLaVA] Architectural Comparison
22. [Prompting] Recency Bias in In-Context Learning
23. [Prompting] Hard vs Soft Prompting
24. [Prompting: Chain-of-Thought] Why Zero-Shot CoT Works
25. [Integrated: Multi-Lecture] Failure Mode in RAG Pipeline