“Classifier-Free Diffusion Guidance”는 조건부 생성(conditional generation) Diffusion Model을 개선한 논문입니다. 이 논문이 발표되기 이전, 특정 클래스의 이미지를 생성하도록 모델을 유도하는 일반적인 방법은 별도의 이미지 분류기(Classifier)를 활용하는 ‘Classifier Guidance’였습니다. 이 방식은 효과적이었으나, Diffusion Model 외에 노이즈 낀 이...
2020년 발표된 “Denoising Diffusion Probabilistic Models” (DDPM)은 생성 모델 연구에 중요한 전환점을 제시한 논문입니다. 당시 Generative Adversarial Networks (GANs)는 훈련 불안정성과 모드 붕괴(mode collapse)와 같은 문제점에도 불구하고 높은 성능으로 인해 고품질 이미지 생성 분야에서 널리 사용되고 있었습니다. DDPM은 이러한 상황 속에서 Diffusion...
Explains the principles and characteristics of L1 and L2 regularization, their connection to Maximum a Posteriori (MAP) estimation, and how to apply regularization to prevent overfitting in deep learning models.
A statistical method used to evaluate whether differences between two sample means are significant or due to chance.
A detailed exploration of Cross-Entropy and KL Divergence, deriving their formulas step-by-step from the principles of probability and information theory.
This article explores the basics of discrete-time signals, including periodic signals and key types like impulse, step, and exponential signals. It also explains LTI systems and how to analyze their stability.
This article will explore the fundamental concepts that arise in Digital Signal Processing (DSP). It will cover concepts such as signals and systems, sinusoids, energy, and power. Examples will be provided to reinforce understanding.
Discover how the Discrete Time Fourier Transform (DTFT) connects discrete signals to their frequency domain representation. By exploring its mathematical foundations, you'll understand its crucial role in digital signal processing.
This article simplifies the concept of Fourier Transform for those without prior exposure to signals and systems. It covers Fourier Series, Continuous Time Fourier Transform (CTFT), and its inverse with step-by-step explanations.
Discover how linear transformations are represented in different bases and why similar matrices preserve essential properties. Understand the mathematics behind change-of-basis operations.
Explains the principles and characteristics of L1 and L2 regularization, their connection to Maximum a Posteriori (MAP) estimation, and how to apply regularization to prevent overfitting in deep learning models.
