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. JavaScript Best Practices

    原文: https://www.w3schools.com/js/js_best_practices.asp --------------------------------------------- ...

  2. zabbix自定义监控项二

    为zabbix增加支持传参的自定义监控项 例如使用zabbix来监控tcp的12种状态 tcp的12种状态可以通过man netstat来找到,即 LISTEN:等待从任何远端TCP 和端口的连接请求 ...

  3. 倍福TwinCAT(贝福Beckhoff)基础教程5.1 TwinCAT如何执行系统命令

    TwinCAT提供了一系列的执行Windows系统命令的方法 Name 描述 NT_Shutdown 关机操作系统 NT_AbortShutdown 取消关机操作系统命令 NT_Reboot 重启操作 ...

  4. nginx静态文件缓存

    open_file_cache max=65535 inactive=30s; open_file_cache 打开缓存的同时也指定了缓存最大数目,以及缓存的时间 open_file_cache_va ...

  5. 24.JAVA编程思想——违例差错控制

    24.JAVA编程思想--违例差错控制 Java 的基本原理就是"形式错误的代码不会执行". 与C++类似,捕获错误最理想的是在编译期间,最好在试图执行程序曾经.然而.并不是全部错 ...

  6. Wamp访问缓慢、本地主机访问缓慢解决方案

    Wamp访问缓慢.本地主机访问缓慢解决方案 Wamp访问速度缓慢可能的原因:          1.一般原因: Wamp的日志文件太大.处理办法,将位于wamp/logs/下的日志文件清空.     ...

  7. 安卓---JNI使用

    交叉编译 系统平台:Windows.Mac OS.Linux CPU平台:ARM.X86.MIPS 概念:在一个平台下打包编译出还有一个平台能够运行的程序#交叉编译原理 源代码->预编译-> ...

  8. angular中通过CSS使下拉列表默认值变灰

    angular版本:angular5 先看效果图: drop down的样式是我用CSS样式控制的,没有用插件.想要改变Drop Down List里的默认值的颜色,我的思路是这样的. 在<se ...

  9. 【HttpClient4.5中文教程】【第一章 ：基础】1.1运行请求（二）

    很多其它HttpClient4.5中文教程请查看:点击打开链接 ==================================================================== ...

  10. zabbix监控xenserver

    xenserver是基于redhat的,可以在zabbix官网下载对应的redhat zabbix安装包,直接安装即可 http://repo.zabbix.com/zabbix/3.0/rhel/5 ...