42 lines
1.3 KiB
Markdown
42 lines
1.3 KiB
Markdown
---
|
||
title: "Composable Base Model Architecture"
|
||
created: 2026-05-15
|
||
updated: 2026-05-15
|
||
type: concept
|
||
tags: [machine-learning, architecture, modular-design]
|
||
sources: [raw/papers/zeng-neurida-2025.md]
|
||
---
|
||
|
||
# Composable Base Model Architecture
|
||
|
||
**可组合基础模型架构**是 NeurIDA 实现 [[dynamic-in-database-modeling|动态建模]] 的架构基础。
|
||
|
||
## 构成
|
||
|
||
```
|
||
基础模型池 M = {m₁, m₂, ..., mₖ}
|
||
+
|
||
共享模型组件
|
||
├── 统一元组编码器（Unified Tuple Encoder）
|
||
├── 关系感知消息传递模块（Relation-Aware Message Passing）
|
||
└── 上下文感知融合模块（Context-Aware Fusion）
|
||
```
|
||
|
||
## 基础模型池
|
||
|
||
涵盖四类异构模型：
|
||
- **Traditional ML**：RF, CatBoost, LightGBM, Logistic Regression
|
||
- **Tuple Representation Models (TRM)**：FT-Transformer, ARM-Net, TabM, ResNet, DNN, DeepFM
|
||
- **Tabular Foundation Models**：TabPFN, TabICL
|
||
- **Large Tabular Models (LTM)**：TP-BERTa, Nomic, BGE
|
||
|
||
## 设计原则
|
||
|
||
1. **异构性**：不同架构的模型互补，适应不同数据分布
|
||
2. **可组合性**：共享组件的接口统一，可任意与基础模型组合
|
||
3. **查询条件化**：[[conditional-model-dispatcher|Dispatcher]] 根据任务选择合适的 m*，DIME 按需装配
|
||
|
||
## 来源
|
||
|
||
- [[zeng-neurida-2025|NeurIDA 论文]]
|