http://neuralnetworksanddeeplearning.com/chap1.html

. Sigmoid neurons are similar to perceptrons, but modified so that small changes in their weights and bias cause only a small change in their output.

http://neuralnetworksanddeeplearning.com/chap3.html

// This is a paper.js widget to show a single neuron learning.  In

// particular, the widget is used to show the learning slowdown that

// occurs when the output is saturated.

//

// The same basic widget is used several times, in slightly different

// configurations.  paper.js makes it somewhat complex to reuse the

// code, so I have simply duplicated the code.  This can give rise to

// bugs if one is not careful to keep the code in sync, so I have

// separated the code into two pieces.

//

// The first piece is the header code.  This changes between widgets.

// It sets up things like the starting weight, bias, cost function,

// and so on -- things which may vary betweens widgets.

//

// The second piece is the body code.  This is almost exactly the same

// for the different widgets.  Note, however, that the costGraphX and

// epochX variables change name, due to a bug in the way paperjs

// handles scope.

//

// We can make these changes by searching on costGraph1 and replacing

// with costGraph2, costGraph3 etc, by replacing epoch1 with epoch2,

// epoch3 etc, and by replcacing cost1 with cost2, cost3 etc.

//

// This separation makes it easy to maintain the duplicated code.

// HEADER CODE

var startingWeight = 0.6;

var startingBias = 0.9;

var eta = 0.15;

var numFrames = 300;

quadratic_cost = {

    fn: function(a) {return a*a/2;},

    derivative: function(a) {return a*a*(1-a);},

    scaling: 240 // used to scale on the graph

}

cross_entropy_cost = {

    fn: function(a) {return -Math.log(1-a);},

    derivative: function(a) {return 1/(1-a);},

    scaling: 30

}

cost1 = quadratic_cost;

// A path for the graph.

costGraph1 = new Path();

costGraph1.strokeColor = "#2A6EA6";

// BODY CODE

// STATIC ELEMENTS

//

// Note that this includes some paper.js items which will later be

// modified, e.g., the variables output and weightText.  This section

// merely sets the static parts of the elements.

var input = new PointText(new Point(8, 40));

input.fontSize = 18;

input.content = "Input: 1.0";

arrow(new Point(100, 35), new Point(230, 35), 0.8); // input arrow

var neuron = new Path.Circle(new Point(260, 35), 30);

neuron.strokeColor = "black";

arrow(new Point(290, 35), new Point(380, 35), 0.8); // output arrow

// The output text's content will be set dynamically, later

var output = new PointText(new Point(390, 40));

output.fontSize = 18;

// The weight text and bar

var weightText = new PointText(new Point(120, 52));

weightText.fontSize=14;

var weightBar = new Path.Rectangle(new Rectangle(120, 57, 90, 9));

weightBar.strokeColor = "grey";

weightBar.strokeWidth = 1;

var weightTick = new Path(new Point(165, 57), new Point(165, 71));

weightTick.strokeColor = "black";

var weightSlider = new Path.Line(

    new Point(165, 61.5), new Point(165+weight*20, 61.5));

weightSlider.strokeColor = "#2A6EA6";

weightSlider.strokeWidth = 9;

// The bias text and bar

var biasText = new PointText(new Point(230, 82));

biasText.fontSize = 14;

var biasBar = new Path.Rectangle(new Rectangle(230, 88, 90, 9));

biasBar.strokeColor = "grey";

biasBar.strokeWidth = 1;

var biasTick = new Path(new Point(275, 88), new Point(275, 102));

biasTick.strokeColor = "black";

var biasSlider = new Path.Line(

    new Point(275, 92.5), new Point(275+bias*20, 92.5));

biasSlider.strokeColor = "#2A6EA6";

biasSlider.strokeWidth = 9;

// Axes for the graph

arrow(new Point(100, 250), new Point(100, 120));

arrow(new Point(100, 250), new Point(130+numFrames/2, 250));

// Labels on the axes

var costText = new PointText(new Point(60, 145));

costText.fontSize = 18;

costText.content = "Cost";

var epoch1LabelText = new PointText(new Point(140+numFrames/2, 255));

epoch1LabelText.fontSize = 18;

epoch1LabelText.content = "Epoch";

// Marker for the current epoch

var epoch1Tick = new Path(new Point(100, 250), new Point(100, 255));

epoch1Tick.strokeColor = "black";

var epoch1Number = new PointText(new Point(100, 267));

epoch1Number.fontSize = 14;

epoch1Number.justification = "center";

// We group the epochTick and epochNumber, to make it easy to move

epoch1 = new Group([epoch1Tick, epoch1Number]);

// Initialize the dynamic elements.  It's convenient to do this in a

// function, so that function can also be called upon a (re)start of

// the widget.

var weight, bias;

initDynamicElements();

function initDynamicElements() {

    weight = startingWeight;

    bias = startingBias;

    weightText.content = paramContent("w = ", weight);

    weightSlider.segments[1].point.x = 165+weight*20;

    biasText.content = paramContent("b = ", bias);

    biasSlider.segments[1].point.x = 275+bias*20;

    output.content = outputContent(weight, bias);

    epoch1.position.x = 100;

    epoch1Number.content = "0";

    costGraph1.removeSegments();

}

function paramContent(s, x) {

    sign = (x >= 0)? "+": "";

    return s+sign+x.toFixed(2);

}

// The run button

var runBox = new Path.Rectangle(new Rectangle(430, 230, 60, 30), 5);

runBox.fillColor = "#dddddd";

var runText = new PointText(new Point(460, 250));

runText.justification = "center";

runText.fontSize = 18;

runText.content = "Run";

var runIcon = new Group([runBox, runText]);

runIcon.onMouseEnter = function(event) {

    runBox.fillColor = "#aaaaaa";

}

runIcon.onMouseLeave = function(event) {

    runBox.fillColor = "#dddddd";

}

var playing = false;

var count = 0;

runIcon.onClick = function(event) {

    initDynamicElements();

    this.visible = false;

    weight = startingWeight;

    bias = startingBias;

    playing = true;

}

// The actual procedure

function onFrame(event) {

    if (playing) {

	a = outputValue(weight, bias);

	delta = cost1.derivative(a);

	weight += -eta*delta;

	bias += -eta*delta;

	weightText.content = paramContent("w = ", weight);

	weightSlider.segments[1].point.x = 165+weight*20;

	biasText.content = paramContent("b = ", bias);

	biasSlider.segments[1].point.x = 275+bias*20;

	output.content = outputContent(weight, bias);

	if (count % 2 === 0) {epoch1.position.x += 1;}

	costGraph1.add(new Point(epoch1.position.x, 250-cost1.scaling*cost1.fn(a)));

	epoch1Number.content = count;

	count += 1;

	if (count > numFrames) {

	    count = 0;

	    runIcon.visible = true;

	    playing = false;

	}

	}

}

function outputValue(weight, bias) {

    return sigmoid(weight+bias);

}

function outputContent(weight, bias) {

    return "Output: "+outputValue(weight, bias).toFixed(2);

}

function sigmoid(z) {

    return 1/(1+Math.exp(-z));

}

function arrow(point1, point2, width, color) {

    if (typeof width === 'undefined') {width=1};

    if (typeof color === 'undefined') {color='black'};

    delta = point1 - point2;

    n = delta/delta.length;

    nperp = new Point(-n.y, n.x);

    line = new Path(point1, point2);

    line.strokeColor = color;

    line.strokeWidth = width;

    arrow_stroke_1 = new Path(point2, point2+(n+nperp)*6);

    arrow_stroke_1.strokeWidth = width;

    arrow_stroke_1.strokeColor = color;

    arrow_stroke_2 = new Path(point2, point2+(n-nperp)*6);

    arrow_stroke_2.strokeWidth = width;

    arrow_stroke_2.strokeColor = color;

}

　　

http://neuralnetworksanddeeplearning.com/js/saturation4.js

// This is a paper.js widget to show a single neuron learning.  In

// particular, the widget is used to show the learning slowdown that

// occurs when the output is saturated.

//

// The same basic widget is used several times, in slightly different

// configurations.  paper.js makes it somewhat complex to reuse the

// code, so I have simply duplicated the code.  This can give rise to

// bugs if one is not careful to keep the code in sync, so I have

// separated the code into two pieces.

//

// The first piece is the header code.  This changes between widgets.

// It sets up things like the starting weight, bias, cost function,

// and so on -- things which may vary betweens widgets.

//

// The second piece is the body code.  This is almost exactly the same

// for the different widgets.  Note, however, that the costGraphX and

// epochX variables change name, due to a bug in the way paperjs

// handles scope.

//

// We can make these changes by searching on costGraph1 and replacing

// with costGraph2, costGraph3 etc, and by replacing epoch1 with

// epoch2, epoch3 etc.

//

// This separation makes it easy to maintain the duplicated code.

// HEADER CODE

var startingWeight = 2.0;

var startingBias = 2.0;

var eta = 0.005;

var numFrames = 300;

quadratic_cost = {

    fn: function(a) {return a*a/2;},

    derivative: function(a) {return a*a*(1-a);},

    scaling: 240 // used to scale on the graph

}

cross_entropy_cost = {

    fn: function(a) {return -Math.log(1-a);},

    derivative: function(a) {return 1/(1-a);},

    scaling: 30

}

cost4 = cross_entropy_cost;

// A path for the graph.

costGraph4 = new Path();

costGraph4.strokeColor = "#2A6EA6";

// BODY CODE

// STATIC ELEMENTS

//

// Note that this includes some paper.js items which will later be

// modified, e.g., the variables output and weightText.  This section

// merely sets the static parts of the elements.

var input = new PointText(new Point(8, 40));

input.fontSize = 18;

input.content = "Input: 1.0";

arrow(new Point(100, 35), new Point(230, 35), 0.8); // input arrow

var neuron = new Path.Circle(new Point(260, 35), 30);

neuron.strokeColor = "black";

arrow(new Point(290, 35), new Point(380, 35), 0.8); // output arrow

// The output text's content will be set dynamically, later

var output = new PointText(new Point(390, 40));

output.fontSize = 18;

// The weight text and bar

var weightText = new PointText(new Point(120, 52));

weightText.fontSize=14;

var weightBar = new Path.Rectangle(new Rectangle(120, 57, 90, 9));

weightBar.strokeColor = "grey";

weightBar.strokeWidth = 1;

var weightTick = new Path(new Point(165, 57), new Point(165, 71));

weightTick.strokeColor = "black";

var weightSlider = new Path.Line(

    new Point(165, 61.5), new Point(165+weight*20, 61.5));

weightSlider.strokeColor = "#2A6EA6";

weightSlider.strokeWidth = 9;

// The bias text and bar

var biasText = new PointText(new Point(230, 82));

biasText.fontSize = 14;

var biasBar = new Path.Rectangle(new Rectangle(230, 88, 90, 9));

biasBar.strokeColor = "grey";

biasBar.strokeWidth = 1;

var biasTick = new Path(new Point(275, 88), new Point(275, 102));

biasTick.strokeColor = "black";

var biasSlider = new Path.Line(

    new Point(275, 92.5), new Point(275+bias*20, 92.5));

biasSlider.strokeColor = "#2A6EA6";

biasSlider.strokeWidth = 9;

// Axes for the graph

arrow(new Point(100, 250), new Point(100, 120));

arrow(new Point(100, 250), new Point(130+numFrames/2, 250));

// Labels on the axes

var costText = new PointText(new Point(60, 145));

costText.fontSize = 18;

costText.content = "Cost";

var epoch4LabelText = new PointText(new Point(140+numFrames/2, 255));

epoch4LabelText.fontSize = 18;

epoch4LabelText.content = "Epoch";

// Marker for the current epoch

var epoch4Tick = new Path(new Point(100, 250), new Point(100, 255));

epoch4Tick.strokeColor = "black";

var epoch4Number = new PointText(new Point(100, 267));

epoch4Number.fontSize = 14;

epoch4Number.justification = "center";

// We group the epochTick and epochNumber, to make it easy to move

epoch4 = new Group([epoch4Tick, epoch4Number]);

// Initialize the dynamic elements.  It's convenient to do this in a

// function, so that function can also be called upon a (re)start of

// the widget.

var weight, bias;

initDynamicElements();

function initDynamicElements() {

    weight = startingWeight;

    bias = startingBias;

    weightText.content = paramContent("w = ", weight);

    weightSlider.segments[1].point.x = 165+weight*20;

    biasText.content = paramContent("b = ", bias);

    biasSlider.segments[1].point.x = 275+bias*20;

    output.content = outputContent(weight, bias);

    epoch4.position.x = 100;

    epoch4Number.content = "0";

    costGraph4.removeSegments();

}

function paramContent(s, x) {

    sign = (x >= 0)? "+": "";

    return s+sign+x.toFixed(2);

}

// The run button

var runBox = new Path.Rectangle(new Rectangle(430, 230, 60, 30), 5);

runBox.fillColor = "#dddddd";

var runText = new PointText(new Point(460, 250));

runText.justification = "center";

runText.fontSize = 18;

runText.content = "Run";

var runIcon = new Group([runBox, runText]);

runIcon.onMouseEnter = function(event) {

    runBox.fillColor = "#aaaaaa";

}

runIcon.onMouseLeave = function(event) {

    runBox.fillColor = "#dddddd";

}

var playing = false;

var count = 0;

runIcon.onClick = function(event) {

    initDynamicElements();

    this.visible = false;

    weight = startingWeight;

    bias = startingBias;

    playing = true;

}

// The actual procedure

function onFrame(event) {

    if (playing) {

	a = outputValue(weight, bias);

	delta = cost4.derivative(a);

	weight += -eta*delta;

	bias += -eta*delta;

	weightText.content = paramContent("w = ", weight);

	weightSlider.segments[1].point.x = 165+weight*20;

	biasText.content = paramContent("b = ", bias);

	biasSlider.segments[1].point.x = 275+bias*20;

	output.content = outputContent(weight, bias);

	if (count % 2 === 0) {epoch4.position.x += 1;}

	costGraph4.add(new Point(epoch4.position.x, 250-cost4.scaling*cost4.fn(a)));

	epoch4Number.content = count;

	count += 1;

	if (count > numFrames) {

	    count = 0;

	    runIcon.visible = true;

	    playing = false;

	}

	}

}

function outputValue(weight, bias) {

    return sigmoid(weight+bias);

}

function outputContent(weight, bias) {

    return "Output: "+outputValue(weight, bias).toFixed(2);

}

function sigmoid(z) {

    return 1/(1+Math.exp(-z));

}

function arrow(point1, point2, width, color) {

    if (typeof width === 'undefined') {width=1};

    if (typeof color === 'undefined') {color='black'};

    delta = point1 - point2;

    n = delta/delta.length;

    nperp = new Point(-n.y, n.x);

    line = new Path(point1, point2);

    line.strokeColor = color;

    line.strokeWidth = width;

    arrow_stroke_1 = new Path(point2, point2+(n+nperp)*6);

    arrow_stroke_1.strokeWidth = width;

    arrow_stroke_1.strokeColor = color;

    arrow_stroke_2 = new Path(point2, point2+(n-nperp)*6);

    arrow_stroke_2.strokeWidth = width;

    arrow_stroke_2.strokeColor = color;

}

　　

output value . Sigmoid neurons are similar to perceptrons, but modified so that small changes in their weights and bias cause only a small change in their output.的更多相关文章

  1. 使用神经网络识别手写数字Using neural nets to recognize handwritten digits

    The human visual system is one of the wonders of the world. Consider the following sequence of handw ...

  2. chapter1:using neural nets to recognize handwritten digits

    two important types of artificial neuron :the perceptron and the sigmoid neuron Perceptrons 感知机的输入个数 ...

  3. 提高神经网络的学习方式Improving the way neural networks learn

    When a golf player is first learning to play golf, they usually spend most of their time developing ...

  4. 神经网络和Deep Learning

    参考资料: 在线免费书籍 http://neuralnetworksanddeeplearning.com/chap1.html Chapter 1 1.  perceptron 感知机 it's a ...

  5. （六）6.16 Neurons Networks linear decoders and its implements

    Sparse AutoEncoder是一个三层结构的网络,分别为输入输出与隐层,前边自编码器的描述可知,神经网络中的神经元都采用相同的激励函数,Linear Decoders 修改了自编码器的定义,对 ...

  6. [LeetCode] Similar RGB Color 相似的红绿蓝颜色

    In the following, every capital letter represents some hexadecimal digit from 0 to f. The red-green- ...

  7. CS229 6.16 Neurons Networks linear decoders and its implements

    Sparse AutoEncoder是一个三层结构的网络,分别为输入输出与隐层,前边自编码器的描述可知,神经网络中的神经元都采用相同的激励函数,Linear Decoders 修改了自编码器的定义,对 ...

  8. Digital Adjustment of DC-DC Converter Output Voltage in Portable Applications

    http://pdfserv.maximintegrated.com/en/an/AN818.pdf http://www.maximintegrated.com/app-notes/index.mv ...

  9. Simple Addition Permits Voltage Control Of DC-DC Converter's Output

    http://electronicdesign.com/power/simple-addition-permits-voltage-control-dc-dc-converters-output In ...

随机推荐

  1. SparkSQL的3种Join实现

    引言 Join是SQL语句中的常用操作,良好的表结构能够将数据分散在不同的表中,使其符合某种范式,减少表冗余.更新容错等.而建立表和表之间关系的最佳方式就是Join操作. 对于Spark来说有3中Jo ...

  2. Android中IntentService详解

    简单说,IntentService是继承于Service并处理异步请求的一个类,在IntentService内有一个工作线程来处理耗时操作,启动IntentService的方式和启动传统Service ...

  3. Windows网页上碰到无法完全显示的图片怎么办

    如图所示,下一幅图片只能显示一半. 我们选中并在新标签中打开 一般即可正常显示,如果还不行,留意地址栏,这就是这个图片的真实地址,我们完全可以用迅雷直接把这幅图片下载下来. 我们甚至可以猜测,去掉后缀 ...

  4. 端口监听与telnet

    例一: 明明端口已经监听了,为什么远程的telnet连接不上. 远程机器 telnet 10.10.1.85 53 就是进不去. Linux 防火墙一关,ok service iptables sto ...

  5. 计算两个有序数组的第K大数（转）

    传统解法,最直观的解法是O(m+n).直接merge两个数组,然后求第K大的数字. 如果想要时间复杂度将为O(log(m+n)).我们可以考虑从K入手.如果我们每次能够删除一个一定在第K个元素之前的元 ...

  6. Android 冷兵器 之 tools

    代码地址如下:http://www.demodashi.com/demo/12612.html 前言 Android开发在所难免的就是UI的预览和调整,一般情况下都是直接run看效果,或者是使用AS的 ...

  7. java检测http请求的ip地址 Java问题通用解决代码

    以后再用到的话,至少能起个参考作用 java实现,struts2的Action中,依赖HttpServletRequest     package net.dookoo.web.action;   i ...

  8. asp.net core mvc视频A：笔记2-1.控制器定义

    方式一:以Controller结尾 方式二:不以Controller结尾 思考 默认路由规则为 运行示例(这里不需要写testcontroller,只写test就可以了) 同理测试test类中的控制器 ...

  9. python 读写 json文件

    json的优势: 1. 数据体积方面. JSON相对于XML来讲,数据的体积小,传递的速度更快些. 2. 传输速度方面. JSON的速度要远远快于XML 3. 数据格式 数据格式比较简单, 易于读写, ...

  10. matlab-非线性方程求根函数及函数曲线绘制

    Matlab中提供了很多求解非线性方程(y=f(x))的函数,刚開始使用,真的很困惑.全部.这里依据matlab的help文档对这些函数做一些小小的总结 fsolve函数 用来求解非线性方程组:F(x ...