AI/ML 研究领域方法分类知识

综述选题时参考此文件设计分类框架。

1. Computer Vision (cs.CV)

Image Classification

CNN 系列: ResNet, EfficientNet, ConvNeXt
Vision Transformer: ViT, DeiT, Swin Transformer
MLP 系列: MLP-Mixer, gMLP
高效架构: MobileNet, ShuffleNet, EfficientFormer

Object Detection

Anchor-based: Faster R-CNN, RetinaNet, YOLO 系列
Anchor-free: CenterNet, FCOS, CornerNet
Transformer-based: DETR, Deformable DETR, DINO
开放集检测: OWL-ViT, Grounding DINO

Image Segmentation

语义分割: FCN, DeepLab, SegFormer
实例分割: Mask R-CNN, SOLOv2, Mask2Former
全景分割: Panoptic FPN, kMaX-DeepLab
通用分割: SAM, Segment Everything Everywhere

Image Generation

GAN 系列: StyleGAN, BigGAN, GigaGAN
Diffusion Models: DDPM, LDM/Stable Diffusion, DALL-E
Autoregressive: VQGAN, Parti, LlamaGen
Flow Matching: Rectified Flow, Stable Diffusion 3

Video Understanding

Action Recognition: SlowFast, TimeSformer, VideoMAE
Video Generation: Sora, CogVideo, Gen-3
Video-Language: VideoCLIP, VideoLLaVA

2. Natural Language Processing (cs.CL)

Language Models

Encoder: BERT, RoBERTa, DeBERTa
Decoder: GPT 系列, LLaMA, Mistral, Qwen
Encoder-Decoder: T5, BART, UL2
Mixture of Experts: Mixtral, Switch Transformer

Reasoning & Agents

Chain-of-Thought: CoT, Tree of Thoughts, Graph of Thoughts
Tool Use: Toolformer, Gorilla, ToolLLM
Agent: ReAct, AutoGPT, Voyager
Planning: SayCan, Inner Monologue

Retrieval-Augmented Generation (RAG)

Dense Retrieval: DPR, Contriever, E5
RAG 架构: RAG, RETRO, Atlas
长上下文: RoPE 扩展, Ring Attention

Alignment & Safety

RLHF: InstructGPT, PPO, DPO
Constitutional AI
Red Teaming & Jailbreak Defense

3. Vision-Language Models (cs.CV + cs.CL)

Image-Text Pretraining

Contrastive: CLIP, ALIGN, SigLIP
Generative: Flamingo, BLIP-2, LLaVA
Unified: CoCa, PaLI, Qwen-VL

Visual Question Answering

VQA 模型: mPLUG, InstructBLIP
Chart/Doc Understanding: Pix2Struct, DocPedia

Image Captioning & Generation

Caption: CoCa, GIT, BLIP-2
Text-to-Image: DALL-E, Stable Diffusion, Midjourney

4. Graph Neural Networks (cs.LG)

基础架构

Message Passing: GCN, GAT, GraphSAGE
Spectral: ChebNet, GNN-FiLM
Expressivity: GIN, k-WL GNN

应用领域

分子性质预测: SchNet, DimeNet, GemNet
推荐系统: LightGCN, PinSage
知识图谱: R-GCN, CompGCN

可扩展性

采样: GraphSAGE, ClusterGCN
分布式: DistDGL, P3

5. Reinforcement Learning (cs.LG + cs.AI)

基础方法

Value-based: DQN, Rainbow, C51
Policy Gradient: PPO, SAC, TD3
Model-based: Dreamer, MBPO, TD-MPC

离线 RL

Conservative: CQL, IQL
Decision Transformer 系列

Multi-Agent RL

CTDE: QMIX, MAPPO
Communication: CommNet, TarMAC

6. Generative Models (cs.LG + cs.CV)

Diffusion Models

基础: DDPM, Score Matching
加速: DDIM, DPM-Solver, Consistency Models
条件生成: Classifier-Free Guidance
架构: U-Net, DiT, U-ViT

Flow Models

Normalizing Flows: RealNVP, Glow
Continuous: FFJORD, Flow Matching
Rectified Flow

VAE 系列

基础 VAE, β-VAE
VQ-VAE, dVAE
Hierarchical VAE

7. Self-Supervised Learning (cs.LG + cs.CV)

Contrastive Learning

SimCLR, MoCo, BYOL, DINO
Barlow Twins, VICReg

Masked Prediction

视觉: MAE, BEiT, iBOT
语言: MLM (BERT), CLM (GPT)
多模态: MultiMAE

Foundation Models

视觉: DINOv2, SAM, SegGPT
语言: GPT-4, Claude, Gemini
多模态: GPT-4V, Gemini, LLaVA

8. Efficient AI (cs.LG + cs.AR)

模型压缩

剪枝: Structured/Unstructured Pruning
量化: INT8, INT4, GPTQ, AWQ
蒸馏: Knowledge Distillation

高效架构

Linear Attention: Linformer, Performer
State Space: Mamba, S4, H3
Mixture of Experts

高效训练

LoRA, QLoRA, Adapter
Mixed Precision, Gradient Checkpointing
Distributed: FSDP, DeepSpeed, Megatron

9. AI for Science (cs.LG + physics/bio/chem)

蛋白质

AlphaFold, ESMFold, RFdiffusion
Protein Language Models: ESM-2, ProtTrans

药物发现

分子生成: MolGPT, REINVENT
对接: DiffDock, EquiBind

天气/气候

FourCastNet, Pangu-Weather, GenCast

数学推理

AlphaProof, Lean, Isabelle
Mathematical Reasoning in LLMs

10. Robotics & Embodied AI (cs.RO + cs.AI)

操控

模仿学习: ACT, Diffusion Policy
强化学习: SAC, PPO for Manipulation

导航

Visual Navigation: CLIP-Nav
Embodied QA: EQA, MP3D

Foundation Models for Robotics

RT-2, Octo, OpenVLA
Language-Conditioned Policies

通用评估指标

领域	常用指标
分类	Accuracy, Top-5 Acc, F1
检测	AP, AP50, AP75, mAP
分割	Dice, IoU, HD95, ASSD
生成	FID, IS, CLIP Score
NLP	BLEU, ROUGE, BERTScore
RL	Cumulative Reward, Success Rate