CCM Blog

GAN & DCGAN 讨论

GAN的两篇入门paper Generative adversarial nets(Ian Goodfellow, 2014)，简称为GAN Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks(Alec Radford, etc, 2015)，简称...

Posted by c cm on July 23, 2017

Boosting方法中的特征重要性

XGBOOST source def get_score(self, fmap='', importance_type='weight'): """Get feature importance of each feature. Importance type can be defined as: 'weight' - the n...

Posted by c cm on July 15, 2017

Machine Learning Yearning

测试/验证集 1 应该根据未来数据的长相，和想要预测好的数据来选择的测试/验证集 2 测试/验证集应该来自同一分布 3 验证集的大小应该要至少能计算出不同算法在评价指标上我们需要的精确度；测试集的大小应该对算法给出足够的置信度，1w内用30%，1w以上可减少 4 优化指标要唯一 5 多个优化目标时，选择N-1个可接受目标，在满足这些目标的基础上，优化最重要的目标 6 用测试/验证...

Posted by c cm on July 8, 2017

Improved GAN

Convergent GAN 1. feature matching ~= maximum mean discrepancy Feature matching addresses the instability of GANs by specifying a new objective for the generator that prevents it from overtrainin...

Posted by c cm on June 30, 2017

Mac上用iTerm使用rz

安装lrzsz brew install lrzsz 下载iterm2-zmodem git clone https://github.com/mmastrac/iterm2-zmodem.git iTerm里面设置Trigger，路径：Preferences-Profiles-Advanced-Trigger Edit ...

Posted by c cm on June 12, 2017

Generative Adversarial Networks

生成模型：用从分布$p_{data}$采样得到的训练集，学习估计出该分布的模型。形式可以是估计出分布$p_{model}$，也可以是学会生成$p_{model}$的样本。 1 生成模型优点生成模型和强化学习关系密切。model-based强化学习中就包含生成模型。比如可用来学习根据现在state和action，生成未来state的分布；使用生成学习来模拟，还能减少因为错误acti...

Posted by c cm on March 15, 2017

Factorization Machines

1. 模型定义其中，k为描述分解维度的超参。优点：特征稀疏时可用线性复杂度 O(kn) 2. 学习直接优化primal 3. 模型对比 SVM 线性核SVM一致比Polynomial核，权重间是相关的比其他Factorization方法如Matrix and Tensor Factori...

Posted by c cm on March 3, 2017

Ad Click Prediction

1. FTRL $\textbf{g}t$ 第t个样本向量 $g{t, i}$ 第t个样本向量的第i个变量 $\textbf{g}_{1:t} = \sum_1^t \textbf{g}_i$ OGD(Online Gradient Descent): 其中$\eta$为学习速率。缺点为稀疏性不足。 FTRL: 其中$\sigma$为学习率，$\sigma_{1:t} ...

Posted by c cm on February 22, 2017

Dynamic Topic Models

1. Dynamic Topic Model 1. Static Topic Model eg. LDA $\beta_{1:K}$ K topics, multinomial distribution of fixed volcabulary Generative Process: Choose topic proportions θ from a distributio...

Posted by c cm on February 15, 2017

Recent Advances in Convolutional Neural Networks

1. Introduction 基本结构： input layer first layer convolutional layer aims to learn feature representations of the inputs. The activation function introduces nonline...

Posted by c cm on February 9, 2017