\nhttps:\/\/deeplearning4j.org\/docs\/latest\/deeplearning4j-nn-recurrent.<\/a>
\nThis demo has been implemented in Scala using Jupyter with the BeakerX<\/a> kernel.<\/p>\n
Setup<\/h3>\n
We add the necessary dependencies to the classpath and we import the classes which we subsequently plan to use.
\n– deeplearning4j-core
\n– nd4j (which is used for the underlying data model)
\n– the DeeplearningforJ UI and Logback which is needed by the the UI<\/p>\n
%%classpath add mvn \norg.nd4j:nd4j-native-platform:1.0.0-beta2\norg.deeplearning4j:deeplearning4j-core:1.0.0-beta2\norg.deeplearning4j:deeplearning4j-ui_2.11:1.0.0-beta2\nch.qos.logback:logback-classic:1.2.3\n\n<\/code><\/pre>\nimport org.deeplearning4j.nn.conf.MultiLayerConfiguration;\nimport org.deeplearning4j.nn.conf.NeuralNetConfiguration;\nimport org.deeplearning4j.nn.conf.NeuralNetConfiguration.ListBuilder;\nimport org.deeplearning4j.nn.conf.layers.LSTM;\nimport org.deeplearning4j.nn.conf.layers.RnnOutputLayer;\nimport org.deeplearning4j.nn.conf.layers.recurrent.SimpleRnn;\nimport org.deeplearning4j.nn.multilayer.MultiLayerNetwork;\nimport org.deeplearning4j.nn.weights.WeightInit;\nimport org.deeplearning4j.optimize.listeners.ScoreIterationListener;\n\nimport org.deeplearning4j.ui.api.UIServer\nimport org.deeplearning4j.ui.storage.InMemoryStatsStorage\nimport org.deeplearning4j.ui.stats.StatsListener\n\nimport org.nd4j.linalg.activations.Activation;\nimport org.nd4j.linalg.api.ndarray.INDArray;\nimport org.nd4j.linalg.api.ops.impl.indexaccum.IMax;\nimport org.nd4j.linalg.dataset.DataSet;\nimport org.nd4j.linalg.factory.Nd4j;\nimport org.nd4j.linalg.learning.config.AMSGrad;\nimport org.nd4j.linalg.learning.config.RmsProp;\nimport org.nd4j.linalg.lossfunctions.LossFunctions.LossFunction;\n\nimport java.util.ArrayList;\nimport java.util.LinkedHashSet;\nimport java.util.List;\nimport java.util.Random;\nimport java.util.Arrays\n\n<\/code><\/pre>\nimport org.deeplearning4j.nn.conf.MultiLayerConfiguration\nimport org.deeplearning4j.nn.conf.NeuralNetConfiguration\nimport org.deeplearning4j.nn.conf.NeuralNetConfiguration.ListBuilder\nimport org.deeplearning4j.nn.conf.layers.LSTM\nimport org.deeplearning4j.nn.conf.layers.RnnOutputLayer\nimport org.deeplearning4j.nn.conf.layers.recurrent.SimpleRnn\nimport org.deeplearning4j.nn.multilayer.MultiLayerNetwork\nimport org.deeplearning4j.nn.weights.WeightInit\nimport org.deeplearning4j.optimize.listeners.ScoreIterationListener\nimport org.deeplearning4j.ui.api.UIServer\nimport org.deeplearning4j.ui.storage.InMemoryStatsStorage\nimport org.deeplearning4j.ui.stats.StatsListener\nimport org.nd4j.linalg.activations.Activation\nimport org.nd4j.linalg.api.ndarray.INDArray\nimport org.nd4j.linalg.api.ops.impl.i...\n<\/code><\/pre>\nData Model – Character Encoding<\/h3>\n
We define the sentence to learn as String. A special character is added at the beginning so the RNN learns the complete string and ends with the marker.<\/p>\n
We also create a dedicated List of possible chars in the charList variable<\/p>\n
val learnString = \"*A computer will do what you tell it to do, but that may be much different from what you had in mind.\";\n\nval charList = learnString.toSet.toList\n\n<\/code><\/pre>\n[[e, *, n, ., y, t, u, f, A, a, m, i, , ,, b, l, p, c, h, r, w, o, d]]\n<\/code><\/pre>\nWe use this character list to encode the characters in an ND4 array:<\/p>\n
var nd4Array = Nd4j.zeros(1, charList.size, 1)\n\n<\/code><\/pre>\n[[0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0]]\n<\/code><\/pre>\nWe determine the index position of the character which we want to represent<\/p>\n
var pos:Int = charList.indexOf('A')\n\n<\/code><\/pre>\n8\n<\/code><\/pre>\nAnd then we set the corresponding index position in the array to 1<\/p>\n
nd4Array.putScalar(pos, 1)\n<\/code><\/pre>\n[[0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 1.0000, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0, \n 0]]\n<\/code><\/pre>\nDefinition of Neural Network<\/h3>\n
We define the network architecture: It consists of subsequent LSTM (Long\/Short Term Memory) Layers followed by a RnnOutputLayer<\/p>\n
import org.nd4j.linalg.learning.config._\n\n\/\/ RNN dimensions\nval HIDDEN_LAYER_WIDTH = 100;\n\n\/\/ some common parameters\nvar conf = new NeuralNetConfiguration.Builder()\n .seed(123)\n .biasInit(0)\n .miniBatch(false)\n .updater(new RmsProp(0.001))\n .weightInit(WeightInit.XAVIER)\n .list()\n .layer(0, new LSTM.Builder()\n .nIn(charList.size)\n .nOut(HIDDEN_LAYER_WIDTH)\n .activation(Activation.TANH)\n .build())\n .layer(1, new LSTM.Builder()\n .nIn(HIDDEN_LAYER_WIDTH)\n .nOut(HIDDEN_LAYER_WIDTH)\n .activation(Activation.TANH)\n .build())\n .layer(2, new LSTM.Builder()\n .nIn(HIDDEN_LAYER_WIDTH)\n .nOut(HIDDEN_LAYER_WIDTH)\n .activation(Activation.TANH)\n .build())\n .layer(3, new RnnOutputLayer.Builder(LossFunction.MCXENT)\n .activation(Activation.SOFTMAX)\n .nIn(HIDDEN_LAYER_WIDTH)\n .nOut(charList.size)\n .build())\n .pretrain(false)\n .backprop(true)\n .build()\n\n\/\/ create network\nvar net = new MultiLayerNetwork(conf)\n\nconf\n<\/code><\/pre>\n{\n \"backprop\" : true,\n \"backpropType\" : \"Standard\",\n \"cacheMode\" : \"NONE\",\n \"confs\" : [ {\n \"cacheMode\" : \"NONE\",\n \"epochCount\" : 0,\n \"iterationCount\" : 0,\n \"layer\" : {\n \"@class\" : \"org.deeplearning4j.nn.conf.layers.LSTM\",\n \"activationFn\" : {\n \"@class\" : \"org.nd4j.linalg.activations.impl.ActivationTanH\"\n },\n \"biasInit\" : 0.0,\n \"biasUpdater\" : null,\n \"constraints\" : null,\n \"dist\" : null,\n \"distRecurrent\" : null,\n \"forgetGateBiasInit\" : 1.0,\n \"gateActivationFn\" : {\n \"@class\" : \"org.nd4j.linalg.activations.impl.ActivationSigmoid\"\n },\n \"gradientNormalization\" : \"None\",\n \"gradientNormalizationThreshold\" : 1.0,\n \"idropout\" : null,\n \"iupdater\" : {\n \"@class\" : \"org.nd4j.linalg.learning.config.RmsProp\",\n \"epsilon\" : 1.0E-8,\n \"learningRate\" : 0.001,\n \"rmsDecay\" : 0.95\n },\n \"l1\" : 0.0,\n \"l1Bias\" : 0.0,\n \"l2\" : 0.0,\n \"l2Bias\" : 0.0,\n \"layerName\" : \"layer0\",\n \"nin\" : 23,\n \"nout\" : 100,\n \"pretrain\" : false,\n \"weightInit\" : \"XAVIER\",\n \"weightInitRecurrent\" : null,\n \"weightNoise\" : null\n },\n \"maxNumLineSearchIterations\" : 5,\n \"miniBatch\" : false,\n \"minimize\" : true,\n \"optimizationAlgo\" : \"STOCHASTIC_GRADIENT_DESCENT\",\n \"pretrain\" : false,\n \"seed\" : 123,\n \"stepFunction\" : null,\n \"variables\" : [ ]\n }, {\n \"cacheMode\" : \"NONE\",\n \"epochCount\" : 0,\n \"iterationCount\" : 0,\n \"layer\" : {\n \"@class\" : \"org.deeplearning4j.nn.conf.layers.LSTM\",\n \"activationFn\" : {\n \"@class\" : \"org.nd4j.linalg.activations.impl.ActivationTanH\"\n },\n \"biasInit\" : 0.0,\n \"biasUpdater\" : null,\n \"constraints\" : null,\n \"dist\" : null,\n \"distRecurrent\" : null,\n \"forgetGateBiasInit\" : 1.0,\n \"gateActivationFn\" : {\n \"@class\" : \"org.nd4j.linalg.activations.impl.ActivationSigmoid\"\n },\n \"gradientNormalization\" : \"None\",\n \"gradientNormalizationThreshold\" : 1.0,\n \"idropout\" : null,\n \"iupdater\" : {\n \"@class\" : \"org.nd4j.linalg.learning.config.RmsProp\",\n \"epsilon\" : 1.0E-8,\n \"learningRate\" : 0.001,\n \"rmsDecay\" : 0.95\n },\n \"l1\" : 0.0,\n \"l1Bias\" : 0.0,\n \"l2\" : 0.0,\n \"l2Bias\" : 0.0,\n \"layerName\" : \"layer1\",\n \"nin\" : 100,\n \"nout\" : 100,\n \"pretrain\" : false,\n \"weightInit\" : \"XAVIER\",\n \"weightInitRecurrent\" : null,\n \"weightNoise\" : null\n },\n \"maxNumLineSearchIterations\" : 5,\n \"miniBatch\" : false,\n \"minimize\" : true,\n \"optimizationAlgo\" : \"STOCHASTIC_GRADIENT_DESCENT\",\n \"pretrain\" : false,\n \"seed\" : 123,\n \"stepFunction\" : null,\n \"variables\" : [ ]\n }, {\n \"cacheMode\" : \"NONE\",\n \"epochCount\" : 0,\n \"iterationCount\" : 0,\n \"layer\" : {\n \"@class\" : \"org.deeplearning4j.nn.conf.layers.LSTM\",\n \"activationFn\" : {\n \"@class\" : \"org.nd4j.linalg.activations.impl.ActivationTanH\"\n },\n \"biasInit\" : 0.0,\n \"biasUpdater\" : null,\n \"constraints\" : null,\n \"dist\" : null,\n \"distRecurrent\" : null,\n \"forgetGateBiasInit\" : 1.0,\n \"gateActivationFn\" : {\n \"@class\" : \"org.nd4j.linalg.activations.impl.ActivationSigmoid\"\n },\n \"gradientNormalization\" : \"None\",\n \"gradientNormalizationThreshold\" : 1.0,\n \"idropout\" : null,\n \"iupdater\" : {\n \"@class\" : \"org.nd4j.linalg.learning.config.RmsProp\",\n \"epsilon\" : 1.0E-8,\n \"learningRate\" : 0.001,\n \"rmsDecay\" : 0.95\n },\n \"l1\" : 0.0,\n \"l1Bias\" : 0.0,\n \"l2\" : 0.0,\n \"l2Bias\" : 0.0,\n \"layerName\" : \"layer2\",\n \"nin\" : 100,\n \"nout\" : 100,\n \"pretrain\" : false,\n \"weightInit\" : \"XAVIER\",\n \"weightInitRecurrent\" : null,\n \"weightNoise\" : null\n },\n \"maxNumLineSearchIterations\" : 5,\n \"miniBatch\" : false,\n \"minimize\" : true,\n \"optimizationAlgo\" : \"STOCHASTIC_GRADIENT_DESCENT\",\n \"pretrain\" : false,\n \"seed\" : 123,\n \"stepFunction\" : null,\n \"variables\" : [ ]\n }, {\n \"cacheMode\" : \"NONE\",\n \"epochCount\" : 0,\n \"iterationCount\" : 0,\n \"layer\" : {\n \"@class\" : \"org.deeplearning4j.nn.conf.layers.RnnOutputLayer\",\n \"activationFn\" : {\n \"@class\" : \"org.nd4j.linalg.activations.impl.ActivationSoftmax\"\n },\n \"biasInit\" : 0.0,\n \"biasUpdater\" : null,\n \"constraints\" : null,\n \"dist\" : null,\n \"gradientNormalization\" : \"None\",\n \"gradientNormalizationThreshold\" : 1.0,\n \"hasBias\" : true,\n \"idropout\" : null,\n \"iupdater\" : {\n \"@class\" : \"org.nd4j.linalg.learning.config.RmsProp\",\n \"epsilon\" : 1.0E-8,\n \"learningRate\" : 0.001,\n \"rmsDecay\" : 0.95\n },\n \"l1\" : 0.0,\n \"l1Bias\" : 0.0,\n \"l2\" : 0.0,\n \"l2Bias\" : 0.0,\n \"layerName\" : \"layer3\",\n \"lossFn\" : {\n \"@class\" : \"org.nd4j.linalg.lossfunctions.impl.LossMCXENT\",\n \"softmaxClipEps\" : 1.0E-10,\n \"configProperties\" : false,\n \"numOutputs\" : -1\n },\n \"nin\" : 100,\n \"nout\" : 23,\n \"pretrain\" : false,\n \"weightInit\" : \"XAVIER\",\n \"weightNoise\" : null\n },\n \"maxNumLineSearchIterations\" : 5,\n \"miniBatch\" : false,\n \"minimize\" : true,\n \"optimizationAlgo\" : \"STOCHASTIC_GRADIENT_DESCENT\",\n \"pretrain\" : false,\n \"seed\" : 123,\n \"stepFunction\" : null,\n \"variables\" : [ ]\n } ],\n \"epochCount\" : 0,\n \"inferenceWorkspaceMode\" : \"ENABLED\",\n \"inputPreProcessors\" : { },\n \"iterationCount\" : 0,\n \"pretrain\" : false,\n \"tbpttBackLength\" : 20,\n \"tbpttFwdLength\" : 20,\n \"trainingWorkspaceMode\" : \"ENABLED\"\n}\n<\/code><\/pre>\nTraining Data<\/h2>\n
We generate the training data as array of all encoded input characters with their corrsponding encoded output which is just the subsequent character. E.g for the inptut ‘A’ we use the output (label) ‘ ‘.<\/p>\n
\/\/ create input and output arrays: SAMPLE_INDEX, INPUT_NEURON,\n\/\/ learnString\nvar input = Nd4j.zeros(1, charList.size, learnString.size);\nvar labels = Nd4j.zeros(1, charList.size, learnString.size);\n\/\/ loop through our sample-sentence\n\nfor (samplePos <- 0 to learnString.size - 1) {\n \/\/ small hack: when currentChar is the last, take the first char as\n \/\/ nextChar - not really required. Added to this hack by adding a starter first character.\n var currentChar = learnString(samplePos);\n \/\/ On the last character we point back to the first character as next position \n var nextChar = learnString((samplePos + 1) % (learnString.length));\n \/\/ input neuron for current-char is 1 at \"samplePos\"\n input.putScalar(Array[Int](0, charList.indexOf(currentChar), samplePos ), 1);\n \/\/ output neuron for next-char is 1 at \"samplePos\"\n labels.putScalar(Array[Int](0, charList.indexOf(nextChar), samplePos ), 1);\n}\n\nvar trainingData = new DataSet(input, labels);\n\n<\/code><\/pre>\n===========INPUT===================\n[[[ 0, 0, 0 ... 0 0, 0], \n [ 1.0000, 0, 0 ... 0 0, 0], \n [ 0, 0, 0 ... 1.0000 0, 0], \n ..., \n [ 0, 0, 0 ... 0 0, 0], \n [ 0, 0, 0 ... 0 0, 0], \n [ 0, 0, 0 ... 0 1.0000, 0]]]\n=================OUTPUT==================\n[[[ 0, 0, 0 ... 0 0, 0], \n [ 0, 0, 0 ... 0 0, 1.0000], \n [ 0, 0, 0 ... 0 0, 0], \n ..., \n [ 0, 0, 0 ... 0 0, 0], \n [ 0, 0, 0 ... 0 0, 0], \n [ 0, 0, 0 ... 1.0000 0, 0]]]\n<\/code><\/pre>\nPlease note that both the features and labels are stored in a 3 dimensions tensor with the following size:<\/p>\n
println(\"Size of Dimension 0 (time series): \"+trainingData.getFeatures().size(0))\nprintln(\"Size of Dimension 1 (values per time step): \"+trainingData.getFeatures().size(1))\nprintln(\"Size of Dimension 2 (time steps): \"+trainingData.getFeatures().size(2))\n<\/code><\/pre>\nSize of Dimension 0 (time series): 1\nSize of Dimension 1 (values per time step): 23\nSize of Dimension 2 (time steps): 101\n\nnull\n<\/code><\/pre>\nTraining UI<\/h3>\n
So that we can follow the leaning in the GUI we do the following:
\n– we set up the UIServer
\n– and show the UIServer in an IFrame
\n– we define the StatsListener to update the UI
\n– and finally start the learing<\/p>\n
\/\/Initialize the user interface backend\nvar uiServer = UIServer.getInstance();\n\n\/\/Configure where the network information (gradients, score vs. time etc) is to be stored. Here: store in memory.\nvar statsStorage = new InMemoryStatsStorage(); \/\/Alternative: new FileStatsStorage(File), for saving and loading later\n\n\/\/Attach the StatsStorage instance to the UI: this allows the contents of the StatsStorage to be visualized\nuiServer.attach(statsStorage)\n\n\"The server is available at http:\/\"+uiServer.getAddress()\n<\/code><\/pre>\nThe server is available at http:\/\/0.0.0.0:9000\n<\/code><\/pre>\n![](\"https:\/\/www.pschatzmann.ch\/wp-content\/uploads\/2018\/09\/TrainingUI.png\")
<\/p>\n
The score is the output of the loss function. The goal of the learning is to minimize the loss: so a smaller number is better as a bigger. We are looking for a situation the Model Score vs Iteration shows a steadily decreasing falling line.<\/p>\n
net.init();\n\/\/ add the StatsListener to collect this information to display in the UI\nnet.setListeners(new StatsListener(statsStorage))\n\nval epochs = 1000\nfor (epoch <- 0 to epochs) {\n \/\/ train the data\n net.fit(trainingData);\n}\n\n\"Training done!\"\n<\/code><\/pre>\nTraining done!\n<\/code><\/pre>\nDisplaying the Result<\/h3>\n
The result can be determined with the help of the rnnTimeStep method. For the first call we pass the encoded start character * and get a matrix as result.<\/p>\n
\/\/ clear current stance from the last example\nnet.rnnClearPreviousState();\n\n\/\/ put the first character into the rrn as an initialisation\nvar testInit = Nd4j.zeros(1, charList.size, 1)\nvar pos:Int = charList.indexOf(learnString(0))\ntestInit.putScalar(pos, 1)\n\n\/\/ run one step -> IMPORTANT: rnnTimeStep() must be called, not\n\/\/ output(). The output shows what the net thinks what should come next\nvar output = net.rnnTimeStep(testInit);\n\n\n<\/code><\/pre>\n[[0.0008, \n 0.0013, \n 0.0007, \n 0.0005, \n 0.0012, \n 0.0029, \n 0.0010, \n 0.0007, \n 0.8118, \n 0.0011, \n 0.0022, \n 0.0007, \n 0.1530, \n 0.0009, \n 0.0007, \n 0.0013, \n 0.0004, \n 0.0122, \n 0.0012, \n 0.0007, \n 0.0015, \n 0.0023, \n 0.0010]]\n<\/code><\/pre>\nThe resulting character can be determined from the maximum value in the array. Here it is the 9.th position with the value of 0.8118. If we check in our charList – the 9.th position contains the character ‘A’ !<\/p>\n
The result matrix is then the input for the next call….<\/p>\n
\/\/ now the net should guess LEARNSTRING.length more characters\nvar result = \"\"\nfor (char <- learnString) {\n \/\/ first process the last output of the network to a concrete\n \/\/ neuron, the neuron with the highest output has the highest\n \/\/ chance to get chosen\n var sampledCharacterIdx = Nd4j.getExecutioner().exec(new IMax(output), 1).getInt(0);\n\n \/\/ concatenate the chosen output\n result += charList(sampledCharacterIdx);\n\n \/\/ use the last output as input\n var nextInput = Nd4j.zeros(1, charList.size, 1);\n nextInput.putScalar(sampledCharacterIdx, 1);\n output = net.rnnTimeStep(nextInput);\n}\n\nresult\n<\/code><\/pre>\nA computer will do what you tell it to do, but that may be much different from what you had in mind.*\n<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"A recurrent neural network (RNN) is a class of artificial neural network where connections between nodes form a directed graph along a sequence. This allows it to exhibit temporal dynamic behavior for a time sequence. Unlike feedforward neural networks, RNNs can use their internal state (memory) to process sequences of […]<\/p>\n","protected":false},"author":1,"featured_media":567,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_crdt_document":"","_import_markdown_pro_load_document_selector":0,"_import_markdown_pro_submit_text_textarea":"","_exactmetrics_skip_tracking":false,"_exactmetrics_sitenote_active":false,"_exactmetrics_sitenote_note":"","_exactmetrics_sitenote_category":0,"footnotes":""},"categories":[14],"tags":[],"class_list":["post-565","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-machine-learning"],"yoast_head":"\n
Deeplearning4j - Recurrent Neural Networks (RNN) - Phil Schatzmann<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Deeplearning4j - Recurrent Neural Networks (RNN) - Phil Schatzmann\" \/>\n<meta property=\"og:description\" content=\"A recurrent neural network (RNN) is a class of artificial neural network where connections between nodes form a directed graph along a sequence. This allows it to exhibit temporal dynamic behavior for a time sequence. Unlike feedforward neural networks, RNNs can use their internal state (memory) to process sequences of […]\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/\" \/>\n<meta property=\"og:site_name\" content=\"Phil Schatzmann\" \/>\n<meta property=\"article:published_time\" content=\"2018-09-24T18:24:14+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2020-11-21T21:22:51+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.pschatzmann.ch\/wp-content\/uploads\/2018\/09\/score.png\" \/>\n\t<meta property=\"og:image:width\" content=\"544\" \/>\n\t<meta property=\"og:image:height\" content=\"389\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"author\" content=\"pschatzmann\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"pschatzmann\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"9 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/\"},\"author\":{\"name\":\"pschatzmann\",\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/#\\\/schema\\\/person\\\/73a53638a4e34e8373405fd737dac9b1\"},\"headline\":\"Deeplearning4j – Recurrent Neural Networks (RNN)\",\"datePublished\":\"2018-09-24T18:24:14+00:00\",\"dateModified\":\"2020-11-21T21:22:51+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/\"},\"wordCount\":474,\"commentCount\":1,\"publisher\":{\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/#\\\/schema\\\/person\\\/73a53638a4e34e8373405fd737dac9b1\"},\"image\":{\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/www.pschatzmann.ch\\\/wp-content\\\/uploads\\\/2018\\\/09\\\/score.png\",\"articleSection\":[\"Machine Learning\"],\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/\",\"url\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/\",\"name\":\"Deeplearning4j - Recurrent Neural Networks (RNN) - Phil Schatzmann\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/www.pschatzmann.ch\\\/wp-content\\\/uploads\\\/2018\\\/09\\\/score.png\",\"datePublished\":\"2018-09-24T18:24:14+00:00\",\"dateModified\":\"2020-11-21T21:22:51+00:00\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/#primaryimage\",\"url\":\"https:\\\/\\\/www.pschatzmann.ch\\\/wp-content\\\/uploads\\\/2018\\\/09\\\/score.png\",\"contentUrl\":\"https:\\\/\\\/www.pschatzmann.ch\\\/wp-content\\\/uploads\\\/2018\\\/09\\\/score.png\",\"width\":544,\"height\":389},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/2018\\\/09\\\/24\\\/deeplearningforj-recurrent-neural-networks-rnn\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Deeplearning4j – Recurrent Neural Networks (RNN)\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/#website\",\"url\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/\",\"name\":\"Phil Schatzmann Consulting\",\"description\":\"\",\"publisher\":{\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/#\\\/schema\\\/person\\\/73a53638a4e34e8373405fd737dac9b1\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":[\"Person\",\"Organization\"],\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/home\\\/#\\\/schema\\\/person\\\/73a53638a4e34e8373405fd737dac9b1\",\"name\":\"pschatzmann\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/wp-content\\\/uploads\\\/2022\\\/08\\\/pschatzmann.png\",\"url\":\"https:\\\/\\\/www.pschatzmann.ch\\\/wp-content\\\/uploads\\\/2022\\\/08\\\/pschatzmann.png\",\"contentUrl\":\"https:\\\/\\\/www.pschatzmann.ch\\\/wp-content\\\/uploads\\\/2022\\\/08\\\/pschatzmann.png\",\"width\":305,\"height\":305,\"caption\":\"pschatzmann\"},\"logo\":{\"@id\":\"https:\\\/\\\/www.pschatzmann.ch\\\/wp-content\\\/uploads\\\/2022\\\/08\\\/pschatzmann.png\"}}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Deeplearning4j - Recurrent Neural Networks (RNN) - Phil Schatzmann","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/","og_locale":"en_US","og_type":"article","og_title":"Deeplearning4j - Recurrent Neural Networks (RNN) - Phil Schatzmann","og_description":"A recurrent neural network (RNN) is a class of artificial neural network where connections between nodes form a directed graph along a sequence. This allows it to exhibit temporal dynamic behavior for a time sequence. Unlike feedforward neural networks, RNNs can use their internal state (memory) to process sequences of […]","og_url":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/","og_site_name":"Phil Schatzmann","article_published_time":"2018-09-24T18:24:14+00:00","article_modified_time":"2020-11-21T21:22:51+00:00","og_image":[{"width":544,"height":389,"url":"https:\/\/www.pschatzmann.ch\/wp-content\/uploads\/2018\/09\/score.png","type":"image\/png"}],"author":"pschatzmann","twitter_card":"summary_large_image","twitter_misc":{"Written by":"pschatzmann","Est. reading time":"9 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/#article","isPartOf":{"@id":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/"},"author":{"name":"pschatzmann","@id":"https:\/\/www.pschatzmann.ch\/home\/#\/schema\/person\/73a53638a4e34e8373405fd737dac9b1"},"headline":"Deeplearning4j – Recurrent Neural Networks (RNN)","datePublished":"2018-09-24T18:24:14+00:00","dateModified":"2020-11-21T21:22:51+00:00","mainEntityOfPage":{"@id":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/"},"wordCount":474,"commentCount":1,"publisher":{"@id":"https:\/\/www.pschatzmann.ch\/home\/#\/schema\/person\/73a53638a4e34e8373405fd737dac9b1"},"image":{"@id":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/#primaryimage"},"thumbnailUrl":"https:\/\/www.pschatzmann.ch\/wp-content\/uploads\/2018\/09\/score.png","articleSection":["Machine Learning"],"inLanguage":"en-US","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/#respond"]}]},{"@type":"WebPage","@id":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/","url":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/","name":"Deeplearning4j - Recurrent Neural Networks (RNN) - Phil Schatzmann","isPartOf":{"@id":"https:\/\/www.pschatzmann.ch\/home\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/#primaryimage"},"image":{"@id":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/#primaryimage"},"thumbnailUrl":"https:\/\/www.pschatzmann.ch\/wp-content\/uploads\/2018\/09\/score.png","datePublished":"2018-09-24T18:24:14+00:00","dateModified":"2020-11-21T21:22:51+00:00","breadcrumb":{"@id":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/#primaryimage","url":"https:\/\/www.pschatzmann.ch\/wp-content\/uploads\/2018\/09\/score.png","contentUrl":"https:\/\/www.pschatzmann.ch\/wp-content\/uploads\/2018\/09\/score.png","width":544,"height":389},{"@type":"BreadcrumbList","@id":"https:\/\/www.pschatzmann.ch\/home\/2018\/09\/24\/deeplearningforj-recurrent-neural-networks-rnn\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.pschatzmann.ch\/home\/"},{"@type":"ListItem","position":2,"name":"Deeplearning4j – Recurrent Neural Networks (RNN)"}]},{"@type":"WebSite","@id":"https:\/\/www.pschatzmann.ch\/home\/#website","url":"https:\/\/www.pschatzmann.ch\/home\/","name":"Phil Schatzmann Consulting","description":"","publisher":{"@id":"https:\/\/www.pschatzmann.ch\/home\/#\/schema\/person\/73a53638a4e34e8373405fd737dac9b1"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.pschatzmann.ch\/home\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":["Person","Organization"],"@id":"https:\/\/www.pschatzmann.ch\/home\/#\/schema\/person\/73a53638a4e34e8373405fd737dac9b1","name":"pschatzmann","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.pschatzmann.ch\/wp-content\/uploads\/2022\/08\/pschatzmann.png","url":"https:\/\/www.pschatzmann.ch\/wp-content\/uploads\/2022\/08\/pschatzmann.png","contentUrl":"https:\/\/www.pschatzmann.ch\/wp-content\/uploads\/2022\/08\/pschatzmann.png","width":305,"height":305,"caption":"pschatzmann"},"logo":{"@id":"https:\/\/www.pschatzmann.ch\/wp-content\/uploads\/2022\/08\/pschatzmann.png"}}]}},"post_mailing_queue_ids":[],"_links":{"self":[{"href":"https:\/\/www.pschatzmann.ch\/home\/wp-json\/wp\/v2\/posts\/565","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.pschatzmann.ch\/home\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.pschatzmann.ch\/home\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.pschatzmann.ch\/home\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.pschatzmann.ch\/home\/wp-json\/wp\/v2\/comments?post=565"}],"version-history":[{"count":1,"href":"https:\/\/www.pschatzmann.ch\/home\/wp-json\/wp\/v2\/posts\/565\/revisions"}],"predecessor-version":[{"id":2221,"href":"https:\/\/www.pschatzmann.ch\/home\/wp-json\/wp\/v2\/posts\/565\/revisions\/2221"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.pschatzmann.ch\/home\/wp-json\/wp\/v2\/media\/567"}],"wp:attachment":[{"href":"https:\/\/www.pschatzmann.ch\/home\/wp-json\/wp\/v2\/media?parent=565"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.pschatzmann.ch\/home\/wp-json\/wp\/v2\/categories?post=565"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.pschatzmann.ch\/home\/wp-json\/wp\/v2\/tags?post=565"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}