CNN_Structure_2

这篇文章还是一篇在学习中整理的笔记,不想做优化,就是这么任性

CNN如何做到的可视化?

在第一层,将权重可视化,因为他们直接和图像相关。那我们如何将后面的卷积层可视化?这些可视化方法都是对神经元的反应做出的可视化,而不是对滤波器本身,实际上目前并没有好的方法对滤波器本身做可视化,所以在ConventnetJS中你可以找到权重,但是并不能直接对他们做出很好的解释。因为权重自己没有很好的解释意义

Pooling Layer 的真正目的以及这么处理的原因

More

MapReduce VS RDBMs

Now, the RDBMs – Rational Database Management Systems, is good for updating a small portion of a big database. RDBMs uses a traditional B-Tree, which is highly dependent on the time required to perform seek operations.

Compared to this, MapReduce is good for updating all, or a majority of a big database. MapReduce uses sort and merge to rebuild the database which depends more on transfer operations.

RDBMs is good for applications that required the data sets if the database to be very frequently updated such as in point queries or small dataset updates. MapReduce is better for WORM, which is a write once and read many times based data applications.

More

Cloud Service Models

We are going to look into further details of the cloud service models.

First of all, look at this picture. At the bottom which is the infrastructure, the IaaS. And in here, virtual machines, server storage, network, is basically what is provides in the infrastructure. Above that is PaaS, where this is a platform as a service, and it includes database, web server and deployment tools. The next one on the top is SaaS and this is software as a Service. Here, CRM, email, games, virtual desktop and what is actually in the software application domain.

More

Cloud Introduction

At first, what does cloud computing do?

Well. It provides online data storage. It enables configuration and accessing of online applications. Apart from that, it provides a variety of software usage. And also it provides computing platform and computing infrastructure.

Now, that may not have lightened up what the meaning of cloud computing is, so here comes an example.

More

MapReduce & Hadoop

Hadoop 是什么?Hadoop 又与 MapReduce 有什么关系?本篇文章会简单介绍两者之间的关系,算是自己在学习过程中做的笔记吧。

我们在练习过程中可能就使用几M的数据量,但是真正在实践中我们接触的数据可比这大得多,这是如果还使用一台电脑进行运算的话就显得十分笨拙,并且实际上远远超出了我们的计算能力。幸运的是一些开源的软件项目提供了处理海量数据的方案,其中一个项目就是 Hadoop,它采用JAVA语言编写,支持在大量机器上分布处理数据。

Hadoop 是 MapReduce 框架的一个开源实现。本篇笔记简单介绍 MapReduce 和 Hadoop 项目。

More

模拟退火

模拟退火是一种通用概率算法,用来在固定时间内寻求一个大的寻找空间内找到的最优解。

简介

模拟退火来自冶金学的专有名词退火。退火是将材料加热后再经特定速率冷却,目的是增大晶粒体积,并且减少晶格中的缺陷。材料中的原子原来会停留在使内能有局部最小值的位置,而随机在其他位置中移动。退火冷却时速度较慢,使得原子有较多可能可以找到内能比原来更低的位置。

模拟退火的原理也和金属退火的原理近似:我们将热力学的理论套用到统计学上,将搜寻空间内每一点想像成空气内的分子;分子的能量,就是它本身的动能;而搜寻空间内的每一点,也像空气分子一样带有“能量”,以表示该点对命题的合适程度。算法先以搜寻空间内一个任意点作起始:每一步先选择一个“邻居”,然后再计算从现有位置到达“邻居”的概率。

More