Files

T

Antigravity Agent f8b21af4be Wiki cleanup: error-doc removal, dedup merge, link normalization

10_Wiki/Topics 대규모 정리:
- 오류 캡처/미완성 stub 문서 227개 제거
- 교차폴더 중복 43클러스터 병합 (63파일 → redirect)
- 링크명 정규화: 깨진 링크 수정·redirect 직결·개념 매핑 ~2,400건
- 카테고리 MOC 6개 신규 생성
- Graph 섹션 미해결 related-keyword 링크 10,058건 제거

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-20 23:52:15 +09:00

3.6 KiB

Raw Blame History

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

title

Noise

매 한 줄

"매 signal은 noise와 함께 살아간다". Noise는 measurement·channel·process에 섞이는 unwanted variation으로, 매 statistical structure (Gaussian, Poisson, 1/f 등)을 가진다. 2026 ML 시대에서도 매 denoising diffusion model의 핵심 도구로 부활.

매 핵심

매 분류 (by spectrum)

White noise: flat power spectrum, 매 사실상 i.i.d.
Pink (1/f) noise: 매 자연계 보편 — neural firing, music, finance.
Brownian (1/f²): 매 random walk integral.
Shot noise (Poisson): 매 photon counting, low-light imaging.
Quantization noise: ADC bit depth 한계.

매 noise model

Additive: y = x + n (대부분 가정).
Multiplicative: y = x · n (speckle, fading).
Convolutive: 매 reverberation.

매 응용

Denoising diffusion (Stable Diffusion 3, FLUX) — noise를 학습 시그널로 사용.
Differential privacy — Laplace/Gaussian noise 추가.
Stochastic optimization — SGD의 noise가 generalization 도움.

💻 패턴

Gaussian noise 추가

import numpy as np
def add_gaussian(x, sigma=0.1):
    return x + np.random.normal(0, sigma, x.shape)

Pink noise 생성 (Voss-McCartney)

def pink_noise(n, num_sources=16):
    array = np.zeros((num_sources, n))
    for i in range(num_sources):
        step = 2 ** i
        array[i, ::step] = np.random.randn((n + step - 1) // step)
    return array.sum(axis=0)

SNR 계산

def snr_db(signal, noise):
    return 10 * np.log10(np.var(signal) / np.var(noise))

Wiener filter (optimal linear denoise)

from scipy.signal import wiener
denoised = wiener(noisy, mysize=5)

DP-noise (differential privacy)

def laplace_dp(value, sensitivity, epsilon):
    return value + np.random.laplace(0, sensitivity / epsilon)

Diffusion forward process

def forward_diffuse(x0, t, betas):
    alpha_bar = np.cumprod(1 - betas)[t]
    eps = np.random.randn(*x0.shape)
    return np.sqrt(alpha_bar) * x0 + np.sqrt(1 - alpha_bar) * eps, eps

매 결정 기준

상황	Approach
Sensor 측정	Gaussian assumption + Kalman
Photon-limited	Poisson MLE
Privacy preserve	Laplace/Gaussian DP
Generative model	Diffusion (DDPM/EDM)

기본값: Additive Gaussian (most analyzable).

🔗 Graph

부모: Entropy in Information Theory · Statistics
응용: Kalman-Filter-and-State-Tracking · Differential-Privacy

🤖 LLM 활용

언제: Data augmentation, robustness training, generative modeling, privacy. 언제 X: Deterministic exact computation 필요 시.

❌ 안티패턴

Noise blindness: noise model 가정 없이 deterministic 처리.
SNR 무시: low-SNR 데이터로 high-precision claim.
Whiteness 가정: 매 실제는 colored noise인데 white로 모델링.

🧪 검증 / 중복

Verified (Papoulis "Probability, Random Variables, and Stochastic Processes").
신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — Noise taxonomy + DP/diffusion patterns

3.6 KiB Raw Blame History