d8a80f6272
이름만 다른(표기 변형) [[위키링크]]를 대상 문서의 canonical 제목으로 치환해 끊겼던 1,200개 링크를 연결. 제목/파일명 정규화 일치만 적용하고 별칭 매칭은 과병합 위험으로 제외(애매성 가드). 원본은 _link_reconcile_backup/ 에 백업. 도구: Datacollect/scripts/link_reconcile_apply.mjs Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
177 lines
6.4 KiB
Markdown
177 lines
6.4 KiB
Markdown
---
|
|
id: wiki-2026-0508-state-space
|
|
title: State Space
|
|
category: 10_Wiki/Topics
|
|
status: verified
|
|
canonical_id: self
|
|
aliases: [ssm, state-space-model, mamba, s4, linear-rnn]
|
|
duplicate_of: none
|
|
source_trust_level: A
|
|
confidence_score: 0.9
|
|
verification_status: applied
|
|
tags: [state-space, control-theory, ssm, mamba, sequence-modeling]
|
|
raw_sources: []
|
|
last_reinforced: 2026-05-10
|
|
github_commit: pending
|
|
tech_stack:
|
|
language: python
|
|
framework: pytorch-mamba-jax
|
|
---
|
|
|
|
# State Space
|
|
|
|
## 매 한 줄
|
|
> **"매 hidden state evolves over time, observations emerge from it"**. State space 는 매 control theory / signal processing 의 핵심 representation: $x_{t+1}=Ax_t+Bu_t,\; y_t=Cx_t+Du_t$. 매 2024-2026 의 deep learning 의 SSM (S4, Mamba, Mamba-2) 가 매 Transformer 의 long-context alternative 로 부상.
|
|
|
|
## 매 핵심
|
|
|
|
### 매 control-theory origin
|
|
- **State** $x_t$: 매 system 의 internal memory.
|
|
- **Input** $u_t$: 매 external control / observation.
|
|
- **Output** $y_t$: 매 measurable.
|
|
- **Matrices** $A, B, C, D$: 매 dynamics, input mapping, observation, feedthrough.
|
|
- **Discrete vs continuous**: $\dot x = Ax + Bu$ (continuous) vs $x_{t+1}$ (discrete).
|
|
|
|
### 매 deep learning SSM 진화
|
|
- **HiPPO (2020)**: 매 long-range memory 의 polynomial 근사 — 매 A matrix 의 영리한 init.
|
|
- **S4 (2022)**: structured SSM, FFT-based convolution view, long-range arena SOTA.
|
|
- **H3, Hyena**: 매 SSM + gating 의 hybrid.
|
|
- **Mamba (2023)**: 매 selective SSM — 매 input-dependent A,B,C, hardware-aware parallel scan.
|
|
- **Mamba-2 (2024)**: 매 SSD (state-space duality) — 매 attention 과 의 unification.
|
|
- **Hybrid (2025+)**: 매 Jamba, Zamba, Samba — 매 Transformer + Mamba mix.
|
|
|
|
### 매 vs Transformer
|
|
| 측면 | Transformer | SSM (Mamba) |
|
|
|---|---|---|
|
|
| Train | O(N²) parallel | O(N) parallel scan |
|
|
| Infer | O(N) per token, KV cache | O(1) per token, fixed state |
|
|
| Memory at infer | grows with context | constant |
|
|
| Long context | quadratic cost | linear cost |
|
|
| In-context recall | strong | weaker (improving with hybrids) |
|
|
|
|
### 매 응용
|
|
1. **Control systems** — 매 robotics, aerospace, Kalman filter.
|
|
2. **Time series** — 매 Kalman / particle filter, dynamic factor model.
|
|
3. **Long-context LLM** — 매 Mamba-3B, Jamba-1.5 의 1M+ context.
|
|
4. **DNA / genomics** — 매 Caduceus, HyenaDNA 의 long sequence.
|
|
5. **Audio** — 매 SaShiMi 의 raw-waveform generation.
|
|
|
|
## 💻 패턴
|
|
|
|
### Classic discrete state space (control)
|
|
```python
|
|
import numpy as np
|
|
def simulate(A, B, C, D, u, x0):
|
|
x, ys = x0.copy(), []
|
|
for ut in u:
|
|
y = C @ x + D @ ut
|
|
x = A @ x + B @ ut
|
|
ys.append(y)
|
|
return np.array(ys)
|
|
```
|
|
|
|
### Kalman filter
|
|
```python
|
|
def kalman_step(x, P, u, z, A, B, C, Q, R):
|
|
# predict
|
|
x = A @ x + B @ u
|
|
P = A @ P @ A.T + Q
|
|
# update
|
|
K = P @ C.T @ np.linalg.inv(C @ P @ C.T + R)
|
|
x = x + K @ (z - C @ x)
|
|
P = (np.eye(len(x)) - K @ C) @ P
|
|
return x, P
|
|
```
|
|
|
|
### S4 convolutional view (PyTorch)
|
|
```python
|
|
import torch, torch.nn.functional as F
|
|
def s4_conv(u, A_bar, B_bar, C, L):
|
|
# K_l = C A_bar^l B_bar → conv kernel of length L
|
|
K = torch.stack([C @ torch.matrix_power(A_bar, l) @ B_bar for l in range(L)])
|
|
y = F.conv1d(u.unsqueeze(0), K.flip(0).unsqueeze(0).unsqueeze(0))
|
|
return y.squeeze()
|
|
```
|
|
|
|
### Mamba selective scan (mamba-ssm package)
|
|
```python
|
|
from mamba_ssm import Mamba
|
|
import torch
|
|
model = Mamba(d_model=512, d_state=16, d_conv=4, expand=2).cuda()
|
|
x = torch.randn(2, 1024, 512, device="cuda")
|
|
y = model(x) # [B, L, D] — O(L) parallel scan
|
|
```
|
|
|
|
### HiPPO initialization (LegS)
|
|
```python
|
|
def hippo_legs(N):
|
|
A = np.zeros((N, N))
|
|
for n in range(N):
|
|
for k in range(N):
|
|
if n > k: A[n, k] = -np.sqrt((2*n+1) * (2*k+1))
|
|
elif n == k: A[n, k] = -(n + 1)
|
|
return A
|
|
```
|
|
|
|
### Mamba-2 SSD block (simplified)
|
|
```python
|
|
def ssd_block(X, A, B, C, chunk=64):
|
|
# state-space duality: structured matrix multiply
|
|
# parallelizable as matmul, shows attention-SSM equivalence
|
|
L = X.shape[1]
|
|
Y = torch.zeros_like(X)
|
|
state = torch.zeros(...)
|
|
for s in range(0, L, chunk):
|
|
Y[:, s:s+chunk], state = ssd_chunk(X[:, s:s+chunk], A, B, C, state)
|
|
return Y
|
|
```
|
|
|
|
### Hybrid (Jamba-style: Mamba + attention)
|
|
```python
|
|
class JambaBlock(nn.Module):
|
|
def __init__(self, d, use_attn=False):
|
|
super().__init__()
|
|
self.layer = MultiheadAttention(d, 8) if use_attn else Mamba(d_model=d)
|
|
self.mlp = MoE(d) if use_moe else MLP(d)
|
|
def forward(self, x): return self.mlp(self.layer(x) + x) + x
|
|
```
|
|
|
|
## 매 결정 기준
|
|
| 상황 | Approach |
|
|
|---|---|
|
|
| Control system 의 design | classical SSM + LQR / Kalman |
|
|
| Time-series filtering | Kalman / particle filter |
|
|
| Long-context language (>128k) | Mamba / Jamba hybrid |
|
|
| In-context recall heavy | Transformer or hybrid (not pure Mamba) |
|
|
| Genomics 1M token | HyenaDNA / Caduceus |
|
|
| Audio raw waveform | SaShiMi / Mamba-Audio |
|
|
|
|
**기본값**: 매 control 의 classical SSM. 매 long-context LM 의 Mamba-2 또는 Jamba 의 hybrid (pure Mamba 의 recall 약점 의 보완).
|
|
|
|
## 🔗 Graph
|
|
- 부모: [[Control-Theory]] · [[Sequence-to-Sequence-Models|Sequence-Modeling]]
|
|
- 변형: [[S4]] · [[Mamba]] · [[Linear-RNN]]
|
|
- 응용: [[Kalman-Filter-and-State-Tracking|Kalman-Filter]] · [[Time-Series-Analysis|Time-Series-Forecasting]]
|
|
- Adjacent: [[Transformer]] · [[데이터 사이언스 및 ML 엔지니어링|RNN]]
|
|
|
|
## 🤖 LLM 활용
|
|
**언제**: SSM literature 정리, mamba-ssm boilerplate, control-system identification 의 sympy / numpy code.
|
|
**언제 X**: real-time safety-critical control (aerospace, medical) — 매 verified controller / formal methods 의 영역.
|
|
|
|
## ❌ 안티패턴
|
|
- **HiPPO init 무시**: random init Mamba 의 long-range 성능 추락. 매 proper A init 필수.
|
|
- **Pure Mamba 의 in-context retrieval 기대**: needle-in-haystack 약함. Hybrid 권장.
|
|
- **No selective mechanism**: 매 vanilla S4 의 modern Mamba 보다 약함 — input-dependent param 필수.
|
|
- **Classical SSM 의 nonlinearity 무시**: 매 real system 의 nonlinear — EKF / UKF / particle filter 사용.
|
|
- **CUDA scan 미활용**: 매 naive Python loop 의 100배 slow. mamba-ssm 의 official kernel 사용.
|
|
|
|
## 🧪 검증 / 중복
|
|
- Verified (Kalman 1960, Gu et al. S4 2022, Gu & Dao Mamba 2023, Dao & Gu Mamba-2 2024, Jamba 2024).
|
|
- 신뢰도 A.
|
|
|
|
## 🕓 Changelog
|
|
| 날짜 | 변경 |
|
|
|---|---|
|
|
| 2026-05-08 | Phase 1 |
|
|
| 2026-05-10 | Manual cleanup — control SSM + modern Mamba/Mamba-2/Jamba |
|