Papers
Paper note는 공개 논문을 요약하고, 논문 안의 claim을 재사용 가능한 concept로 다시 연결하는 공간입니다. 이 영역은 workflow, checklist, template을 모두 쌓아둔 덤프가 아니라 선별된 reading shelf처럼 읽히는 것을 목표로 합니다.
매일 모이는 raw paper candidate는 먼저 Inbox에 두고, 실제로 읽고 남길 가치가 있는 논문만 여기로 승격합니다.
선별 영역
| 영역 | 용도 |
|---|---|
| Architecture papers | Transformer, GNN, SSM/Mamba, MoE, geometric architecture design, language-model backbone papers |
| Generative model papers | diffusion, flow, molecule/protein generation |
| Computational Biology papers | structure-based modeling, protein modeling, molecular generation, domain evaluation |
Essential AI Reading Queue
논문 노트로 바로 만들기 전, 먼저 Essential AI reading queue에서 큰 축별 후보만 관리합니다. 이 목록은 세부 리뷰가 아니라 앞으로 채울 paper note의 후보 선반입니다.
Paper Note Boundary
Paper note는 논문 하나의 주장, 방법, 실험, 한계를 읽기 위한 페이지입니다. 논문을 핑계로 전체 개념을 설명하려면 해당 설명은 Concepts로 분리합니다.
논문은 여러 주제와 겹칠 수 있지만 canonical note는 가장 강한 claim을 기준으로 한 선반에만 둡니다. 다른 선반에서는 중복 문서를 만들지 않고 cross-link합니다. 따라서 Architecture에 있는 구조 논문이 Computational Biology나 Generative Models에서 다시 등장해도 이는 중복 보관이 아니라 독서 경로 연결입니다.
| Field | Write |
|---|---|
| Citation | paper title, year, venue/arXiv if public |
| Claim | 논문이 실제로 주장하는 것 |
| Method | architecture, data, objective, evaluation setup |
| Evidence | table, benchmark, ablation, theorem, qualitative result |
| Limits | dataset, metric, baseline, leakage, scaling, reproducibility issue |
| Links | reusable concepts, related papers, possible projects |
Unsupported personal interpretation은 open question이나 to verify로 표시합니다.
Promotion and Routing
Raw paper candidates should not all become paper notes. The promotion path is:
| If the paper mainly contributes | Route |
|---|---|
| architecture or inductive bias | Architecture papers, then Architectures |
| learning objective or supervision setup | Learning method papers, then Learning methods |
| molecular/protein/structure modeling | Computational Biology papers |
| benchmark or evaluation protocol | Paper analysis, Evaluation |
| reproducible implementation target | Paper reproducibility, then Projects |
| agent/tool/workflow behavior | Agents and paper workflow notes |
Reading Depth
Not every paper needs a long review.
| Depth | Use when | Output |
|---|---|---|
| candidate | maybe relevant | inbox item or queue row |
| skim | useful but not central | short note with claim and route |
| full note | central to wiki topic | paper note with evidence and links |
| reproduction | implementation matters | readiness, plan, result pages |
| synthesis | multiple papers form a pattern | post or concept update |
현재 논문 노트
| 논문 노트 | 영역 |
|---|---|
| Gradient-Based Learning Applied to Document Recognition | LeNet-5 and early CNN document recognition |
| ImageNet Classification with Deep CNNs | large-scale CNN vision architecture |
| Very Deep Convolutional Networks | deep small-filter CNN architecture |
| Network In Network | convolution and global average pooling |
| Going Deeper with Convolutions | Inception CNN module |
| Densely Connected Convolutional Networks | dense CNN connectivity |
| EfficientNet | CNN compound scaling |
| Attention Is All You Need | Transformer architecture |
| Pointer Networks | attention-based variable-output pointer decoder |
| Dynamic Routing Between Capsules | vector-valued entities and routing-by-agreement |
| Human-level control through deep reinforcement learning | CNN observation encoder and action-value control architecture |
| World Models | latent visual state, probabilistic dynamics, and imagined control |
| Dream to Control: Learning Behaviors by Latent Imagination | RSSM, latent imagination, and actor-critic learning through imagined trajectories |
| Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model | MuZero latent model, reward/policy/value prediction, and MCTS planning |
| Universal Transformers | recurrent-depth Transformer with adaptive computation |
| FaceNet | shared-weight embedding architecture and triplet-distance geometry |
| Siamese Neural Networks for One-shot Image Recognition | tied branches, pairwise verification, and one-shot support/query transfer |
| Neural Architecture Search with Reinforcement Learning | controller-generated architectures and search-budget contract |
| Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm | policy/value network, MCTS, and self-play agent architecture |
| BERT | encoder-only Transformer language representation |
| Language Models are Unsupervised Multitask Learners | decoder-only Transformer language model |
| Long Short-Term Memory | gated recurrent memory |
| Neural Turing Machines | differentiable external memory and soft read/write heads |
| Memory Networks | modular long-term memory for multi-step inference |
| End-To-End Memory Networks | recurrent multi-hop attention over external memory |
| Hybrid Computing Using a Neural Network with Dynamic External Memory | dynamic memory allocation and temporal-link traversal |
| Hopfield Networks is All You Need | continuous associative memory and attention-like retrieval |
| Learning Phrase Representations using RNN Encoder-Decoder | GRU and encoder-decoder architecture |
| Sequence to Sequence Learning with Neural Networks | LSTM encoder-decoder sequence transduction |
| Gated Linear Attention | hardware-aware gated linear attention |
| DeltaNet | delta-rule linear Transformer |
| Griffin | gated linear recurrence and local attention hybrid |
| Dropout | stochastic activation masking and regularization |
| Maxout Networks | learned piecewise-linear activation block |
| Batch Normalization | normalization architecture block |
| Highway Networks | gated skip paths for very deep networks |
| Deep Residual Learning | residual CNN architecture |
| U-Net | dense prediction architecture |
| An Image is Worth 16x16 Words | Vision Transformer |
| Swin Transformer | hierarchical vision Transformer |
| Layer Normalization | sequence-model normalization |
| Semi-Supervised Classification with GCNs | graph neural network |
| Graph Attention Networks | graph attention network |
| Equivariant Sheaf Neural Networks | directed matrix-valued geometric transport with exact Euclidean equivariance |
| Deep Sets | permutation-invariant set architecture |
| Set Transformer | attention-based set architecture |
| Perceiver IO | structured input/output attention architecture |
| Neural Ordinary Differential Equations | continuous-depth architecture |
| Mamba | selective state-space sequence model |
| Jamba | hybrid Transformer-Mamba-MoE language model |
| Switch Transformer | sparse mixture-of-experts architecture |
| SurfSpec | off-target-agnostic specificity via pocket-ligand geometric mismatch bounds |
| AdaptiveFlow | adaptive compute allocation for ultra-large virtual screening |
| NEAT-POCKET | pocket-conditioned autoregressive 3D molecular generation with local flow matching |
| BackFlip-2 | directional protein flexibility and dynamic coupling from static structures |
| Controllable Molecular Generation with Fine-Tuned Flow Matching | reward-guided post-training for controllable 3D molecular flow models |
| PoseBusters | structure-based evaluation |
| Multi-scale ML for Antibody-Antigen Binding | antibody-antigen binding |
| MEET | equivariant peptide modeling |
| Molexar | molecular generation |
보조 묶음
아래 묶음은 좁은 논문군을 잃지 않기 위한 보조 선반입니다. 처음 읽을 때는 더 큰 paper shelf에서 시작하고, 특정 분야로 좁힐 때 들어갑니다.
| 묶음 | 상위 선반 |
|---|---|
| Structure-based modeling papers | Computational Biology papers |
| Protein modeling papers | Computational Biology papers |
| Learning method papers | AI learning methods, Architecture papers, Agents, or Generative model papers by strongest claim |
| Systems papers | AI systems, Infra, or Agents by strongest claim |
읽기 도구
아래 페이지들은 논문 묶음이 아니라 논문 읽기를 돕는 support material입니다. 실제 논문 리뷰를 쓰거나 후보를 승격할 때 필요한 기준을 확인하는 데 사용합니다.
| 도구 | 용도 |
|---|---|
| Paper workflows | triage, reading state, note format |
| Paper analysis | claims, benchmarks, ablations, limitations |
| Paper reproducibility | artifacts, readiness, plans, results |
함께 봐야 할 개념
| 맥락 | 링크 |
|---|---|
| Data and benchmark | Benchmark |
| Inputs and tasks | Modalities, Tasks, Multimodal learning |
| Models and evidence | Architectures, Evaluation, Coverage matrix |
| Paper creation | Longform paper review guide, AI-Molecular-Math paper template, Paper review workflow |
Related
| 영역 | 링크 |
|---|---|
| Research | Research |
| SBDD | Scoring function |