|
| 1 | +\documentclass{beamer} |
| 2 | + |
| 3 | +\usepackage{amsmath,amsfonts,amssymb,bm} |
| 4 | +\usepackage{physics} |
| 5 | + |
| 6 | +\usetheme{Madrid} |
| 7 | + |
| 8 | +\title{Diffusion Models and Normalizing Flows} |
| 9 | +\author{MHJ} |
| 10 | +\date{March 29, 2025} |
| 11 | + |
| 12 | +\begin{document} |
| 13 | + |
| 14 | +\frame{\titlepage} |
| 15 | + |
| 16 | +%================================================ |
| 17 | +\section{Motivation} |
| 18 | +%================================================ |
| 19 | + |
| 20 | +\begin{frame}{Generative Modeling} |
| 21 | +Goal: learn a distribution |
| 22 | +\[ |
| 23 | +p(x), \quad x \in \mathbb{R}^d |
| 24 | +\] |
| 25 | + |
| 26 | +Two major paradigms: |
| 27 | +\begin{itemize} |
| 28 | +\item Normalizing flows (deterministic) |
| 29 | +\item Diffusion models (stochastic) |
| 30 | +\end{itemize} |
| 31 | +\end{frame} |
| 32 | + |
| 33 | +%================================================ |
| 34 | +\section{Normalizing Flows} |
| 35 | +%================================================ |
| 36 | + |
| 37 | +\begin{frame}{Basic Idea} |
| 38 | +\[ |
| 39 | +z \sim p_0(z), \quad x = f(z) |
| 40 | +\] |
| 41 | + |
| 42 | +\[ |
| 43 | +p(x) = p_0(z)\left|\det \frac{\partial z}{\partial x}\right| |
| 44 | +\] |
| 45 | +\end{frame} |
| 46 | + |
| 47 | +%------------------------------------------------ |
| 48 | + |
| 49 | +\begin{frame}{Change of Variables} |
| 50 | +\[ |
| 51 | +\log p(x) = \log p_0(z) - \log |\det J| |
| 52 | +\] |
| 53 | +\end{frame} |
| 54 | + |
| 55 | +%------------------------------------------------ |
| 56 | + |
| 57 | +\begin{frame}{Composition} |
| 58 | +\[ |
| 59 | +x = f_K \circ \cdots \circ f_1(z) |
| 60 | +\] |
| 61 | +\end{frame} |
| 62 | + |
| 63 | +%------------------------------------------------ |
| 64 | + |
| 65 | +\begin{frame}{Continuous Flow} |
| 66 | +\[ |
| 67 | +\frac{dx}{dt} = v(x,t) |
| 68 | +\] |
| 69 | + |
| 70 | +\[ |
| 71 | +\frac{d}{dt}\log p(x) = -\nabla \cdot v |
| 72 | +\] |
| 73 | +\end{frame} |
| 74 | + |
| 75 | +%------------------------------------------------ |
| 76 | + |
| 77 | +\begin{frame}{Interpretation} |
| 78 | +\begin{itemize} |
| 79 | +\item deterministic transport |
| 80 | +\item invertible mapping |
| 81 | +\end{itemize} |
| 82 | +\end{frame} |
| 83 | + |
| 84 | +%================================================ |
| 85 | +\section{Diffusion Models} |
| 86 | +%================================================ |
| 87 | + |
| 88 | +\begin{frame}{Forward Process} |
| 89 | +\[ |
| 90 | +dx = -\frac{1}{2}\beta(t)x\,dt + \sqrt{\beta(t)}\,dW_t |
| 91 | +\] |
| 92 | +\end{frame} |
| 93 | + |
| 94 | +%------------------------------------------------ |
| 95 | + |
| 96 | +\begin{frame}{Fokker–Planck Equation} |
| 97 | +\[ |
| 98 | +\partial_t p = -\nabla \cdot (f p) + \frac{1}{2}\nabla^2 (\sigma^2 p) |
| 99 | +\] |
| 100 | +\end{frame} |
| 101 | + |
| 102 | +%------------------------------------------------ |
| 103 | + |
| 104 | +\begin{frame}{Reverse Process} |
| 105 | +\[ |
| 106 | +dx = [f(x,t) - \sigma^2 \nabla \log p(x,t)]dt + \sigma d\bar{W}_t |
| 107 | +\] |
| 108 | +\end{frame} |
| 109 | + |
| 110 | +%------------------------------------------------ |
| 111 | + |
| 112 | +\begin{frame}{Score Function} |
| 113 | +\[ |
| 114 | +s(x,t) = \nabla \log p(x,t) |
| 115 | +\] |
| 116 | +\end{frame} |
| 117 | + |
| 118 | +%------------------------------------------------ |
| 119 | + |
| 120 | +\begin{frame}{Training Objective} |
| 121 | +\[ |
| 122 | +\mathbb{E}\|s_\theta(x,t) - \nabla \log p(x,t)\|^2 |
| 123 | +\] |
| 124 | +\end{frame} |
| 125 | + |
| 126 | +%================================================ |
| 127 | +\section{Connection Between the Two} |
| 128 | +%================================================ |
| 129 | + |
| 130 | +\begin{frame}{Common Structure} |
| 131 | +Both aim to construct: |
| 132 | +\[ |
| 133 | +T: p_0 \rightarrow p(x) |
| 134 | +\] |
| 135 | +\end{frame} |
| 136 | + |
| 137 | +%------------------------------------------------ |
| 138 | + |
| 139 | +\begin{frame}{Flows vs Diffusion} |
| 140 | +Flows: |
| 141 | +\[ |
| 142 | +x = f(z) |
| 143 | +\] |
| 144 | + |
| 145 | +Diffusion: |
| 146 | +\[ |
| 147 | +dx = v(x,t)dt + \sigma dW_t |
| 148 | +\] |
| 149 | +\end{frame} |
| 150 | + |
| 151 | +%------------------------------------------------ |
| 152 | + |
| 153 | +\begin{frame}{Deterministic vs Stochastic} |
| 154 | +\begin{itemize} |
| 155 | +\item flows: deterministic transport |
| 156 | +\item diffusion: stochastic evolution |
| 157 | +\end{itemize} |
| 158 | +\end{frame} |
| 159 | + |
| 160 | +%================================================ |
| 161 | +\section{PDE Perspective} |
| 162 | +%================================================ |
| 163 | + |
| 164 | +\begin{frame}{Continuity Equation (Flows)} |
| 165 | +\[ |
| 166 | +\partial_t p + \nabla \cdot (v p) = 0 |
| 167 | +\] |
| 168 | +\end{frame} |
| 169 | + |
| 170 | +%------------------------------------------------ |
| 171 | + |
| 172 | +\begin{frame}{Fokker–Planck (Diffusion)} |
| 173 | +\[ |
| 174 | +\partial_t p = -\nabla \cdot (v p) + \frac{1}{2}\nabla^2 (\sigma^2 p) |
| 175 | +\] |
| 176 | +\end{frame} |
| 177 | + |
| 178 | +%------------------------------------------------ |
| 179 | + |
| 180 | +\begin{frame}{Key Difference} |
| 181 | +\begin{itemize} |
| 182 | +\item flows: transport only |
| 183 | +\item diffusion: transport + noise |
| 184 | +\end{itemize} |
| 185 | +\end{frame} |
| 186 | + |
| 187 | +%================================================ |
| 188 | +\section{Statistical Physics Interpretation} |
| 189 | +%================================================ |
| 190 | + |
| 191 | +\begin{frame}{Boltzmann Distribution} |
| 192 | +\[ |
| 193 | +p(x) \propto e^{-E(x)} |
| 194 | +\] |
| 195 | +\end{frame} |
| 196 | + |
| 197 | +%------------------------------------------------ |
| 198 | + |
| 199 | +\begin{frame}{Diffusion as Langevin Dynamics} |
| 200 | +\[ |
| 201 | +dx = -\nabla E(x)dt + \sqrt{2}dW_t |
| 202 | +\] |
| 203 | +\end{frame} |
| 204 | + |
| 205 | +%------------------------------------------------ |
| 206 | + |
| 207 | +\begin{frame}{Flows as Deterministic Maps} |
| 208 | +\begin{itemize} |
| 209 | +\item map Gaussian to Boltzmann |
| 210 | +\end{itemize} |
| 211 | +\end{frame} |
| 212 | + |
| 213 | +%================================================ |
| 214 | +\section{Likelihood vs Score Matching} |
| 215 | +%================================================ |
| 216 | + |
| 217 | +\begin{frame}{Flows} |
| 218 | +\[ |
| 219 | +\max \log p(x) |
| 220 | +\] |
| 221 | +\end{frame} |
| 222 | + |
| 223 | +%------------------------------------------------ |
| 224 | + |
| 225 | +\begin{frame}{Diffusion} |
| 226 | +\[ |
| 227 | +\min \|s_\theta - \nabla \log p\|^2 |
| 228 | +\] |
| 229 | +\end{frame} |
| 230 | + |
| 231 | +%================================================ |
| 232 | +\section{Advantages and Limitations} |
| 233 | +%================================================ |
| 234 | + |
| 235 | +\begin{frame}{Normalizing Flows} |
| 236 | +\begin{itemize} |
| 237 | +\item exact likelihood |
| 238 | +\item invertible |
| 239 | +\item limited flexibility |
| 240 | +\end{itemize} |
| 241 | +\end{frame} |
| 242 | + |
| 243 | +%------------------------------------------------ |
| 244 | + |
| 245 | +\begin{frame}{Diffusion Models} |
| 246 | +\begin{itemize} |
| 247 | +\item highly expressive |
| 248 | +\item expensive sampling |
| 249 | +\item no explicit likelihood |
| 250 | +\end{itemize} |
| 251 | +\end{frame} |
| 252 | + |
| 253 | +%================================================ |
| 254 | +\section{Geometric View} |
| 255 | +%================================================ |
| 256 | + |
| 257 | +\begin{frame}{Flows} |
| 258 | +\begin{itemize} |
| 259 | +\item diffeomorphisms |
| 260 | +\item volume changes via Jacobian |
| 261 | +\end{itemize} |
| 262 | +\end{frame} |
| 263 | + |
| 264 | +%------------------------------------------------ |
| 265 | + |
| 266 | +\begin{frame}{Diffusion} |
| 267 | +\begin{itemize} |
| 268 | +\item stochastic paths |
| 269 | +\item measure evolution |
| 270 | +\end{itemize} |
| 271 | +\end{frame} |
| 272 | + |
| 273 | +%================================================ |
| 274 | +\section{Unification} |
| 275 | +%================================================ |
| 276 | + |
| 277 | +\begin{frame}{Unified View} |
| 278 | +\[ |
| 279 | +\text{Generative model} = \text{transport of probability measure} |
| 280 | +\] |
| 281 | +\end{frame} |
| 282 | + |
| 283 | +%------------------------------------------------ |
| 284 | + |
| 285 | +\begin{frame}{Operator Perspective} |
| 286 | +\begin{itemize} |
| 287 | +\item flows: pushforward operator |
| 288 | +\item diffusion: stochastic evolution operator |
| 289 | +\end{itemize} |
| 290 | +\end{frame} |
| 291 | + |
| 292 | +%------------------------------------------------ |
| 293 | + |
| 294 | +\begin{frame}{Master Equation} |
| 295 | +\[ |
| 296 | +\boxed{ |
| 297 | +\text{Flows = deterministic transport} |
| 298 | +\quad |
| 299 | +\text{Diffusion = stochastic transport} |
| 300 | +} |
| 301 | +\] |
| 302 | +\end{frame} |
| 303 | + |
| 304 | +%================================================ |
| 305 | +\section{Outlook} |
| 306 | +%================================================ |
| 307 | + |
| 308 | +\begin{frame}{Future Directions} |
| 309 | +\begin{itemize} |
| 310 | +\item flow–diffusion hybrids |
| 311 | +\item Schrödinger bridges |
| 312 | +\item connections to quantum systems |
| 313 | +\end{itemize} |
| 314 | +\end{frame} |
| 315 | + |
| 316 | +\end{document} |
0 commit comments