Xuan's blog

So, one more dr!nk?

On the existence of optimal policy

$\newcommand{\mc}{\mathcal} \newcommand{\mb}{\mathbb}$ Setting We are considering in the setting of: Discrete actions Discrete states Bounded rewards Stationary policy Infinite horizo...

ML notes: Latent variable models

General latent variable model Learning problem: Given incomplete dataset $\{ x_i\}$, estimate $\theta^\ast$, for instance by MLE: $\hat \theta = \arg \max_\theta p(x;\theta)$ Inference problem: ...

ML notes: Kernel methods and SVM

$\newcommand{\ip}[2]{\langle #1, #2 \rangle}​$ The Representer Theorum For a general problem where $f: \mathbb{R}^m \rightarrow \mathbb{R}​$ is usually the cost, and $R: \mathbb{R}_+ \rightarrow...

ML notes: Statistical learning theory

$\newcommand{\mc}{\mathcal} \newcommand{\mb}{\mathbb}$ Learning theory Three spaces $\mathcal{X}$ : Input space $\mathcal{A}$ : Action space $ \mathcal{Y}$ : Outcome space Note Outc...

ML notes: Bayesian

$\newcommand{\mc}{\mathcal} \newcommand{\mb}{\mathbb}$ Classic statistics notions The world is in a parametric (density) family $\mc{P}_{\Theta} = \{ p(y \vert\theta) \;\vert \;\theta \in \Theta\...

Convex optimization review

just to connect some dots

TODO: 1) Convergence analysis 2) interior point primal dual method Convexity Separating hyperplane theorum If two convex sets, $A​$ and $B​$ are disjoint, then there exists a hyperplane separa...

About explaining away

experiments and thoughts

Explore the phenomenon of “explaining away” First, let’s consider a simple example, where a guy’s happiness (H) can be caused by either a raise in his company (R) or a sunny day (S), and assuming t...

Elements of GPU overflow programming:P

an exploded Integer whitens all the data

An exploded Integer whitens all the data -1. Update! So, I finally figured things out after turning to Professor… So when overflow happens, the kernel got killed because GPU may have some internal ...

How to use TensorFlow models in Java

A skeleton example

A Skeleton Showing How to Use TF Models in JAVA 0. Requirements Tensorflow >= 1.3 Maven Note: Maven is not necessary, you could as well add raw TensorFlow’s Java binding JAR into your pro...

自动客服系统

Build QA bot from historical dump dialog

Build QA Bot from Dumped Dialog author: Xuan Wang 1. Intro This is a summary from my work in Laiye, Inc. I will be mainly looking at the processes not specific algorithems or models. 1...