Deep Learning With H2O - Webpages.uidaho.edu

H2O’s Deep Learning Architecture 115.1Summary of FeaturesH2O’s Deep Learning functionalities include: supervised training protocol for regression and classification tasks fast and memory-efficient Java implementations based on columnar compression and fine-grain MapReduce multi-threaded and distributed parallel computation that can be run on asingle or a multi-node cluster automatic, per-neuron, adaptive learning rate for fast convergence optional specification of learning rate, annealing, and momentum options regularization options such as L1, L2, dropout, Hogwild!, and modelaveraging to prevent model overfitting elegant and intuitive web interface (Flow) fully scriptable R API from H2O’s CRAN package fully scriptable Python API grid search for hyperparameter optimization and model selection automatic early stopping based on convergence of user-specified metricsto user-specified tolerance model checkpointing for reduced run times and model tuning automatic pre- and post-processing for categorical and numerical data automatic imputation of missing values (optional) automatic tuning of communication vs computation for best performance model export in plain Java code for deployment in production environments additional expert parameters for model tuning deep autoencoders for unsupervised feature learning and anomaly detection

12 H2O’s Deep Learning Architecture5.2Training ProtocolThe training protocol described below follows many of the ideas and advancesdiscussed in recent Deep Learning literature.5.2.1InitializationVarious Deep Learning architectures employ a combination of unsupervisedpre-training followed by supervised training, but H2O uses a purely supervisedtraining protocol. The default initialization scheme is the uniform adaptiveoption, which is an optimized initialization based on the size of the network.Deep Learning can also be started using a random initialization drawn fromeither a uniform or normal distribution, optionally specifying a scaling parameter.5.2.2Activation and Loss FunctionsThe choices for the nonlinear activation function f described in the introductionare summarized in Table 1 below. xi and wi represent the firing neuron’s inputvaluesPand their weights, respectively; α denotes the weighted combinationα i wi xi b.Table 1: Activation FunctionsFunctionTanhRectified LinearMaxoutFormulaeα e αeα e αf (α) f (α) max(0, α)f (α1 , α2 ) max(α1 , α2 )Rangef (·) [ 1, 1]f (·) R f (·) RThe tanh function is a rescaled and shifted logistic function; its symmetryaround 0 allows the training algorithm to converge faster. The rectified linearactivation function has demonstrated high performance on image recognitiontasks and is a more biologically accurate model of neuron activations (LeCunet al, 1998).Maxout is a generalization of the Rectifiied Linear activation, whereeach neuron picks the largest output of k separate channels, where each channelhas its own weights and bias values. The current implementation supports onlyk 2. Maxout activation works particularly well with dropout (Goodfellow etal, 2013). For more information, refer to Regularization.

H2O’s Deep Learning Architecture 13The Rectifier is the special case of Maxout where the output of one channel isalways 0. It is difficult to determine a “best” activation function to use; eachmay outperform the others in separate scenarios, but grid search models can helpto compare activation functions and other parameters. For more information,refer to Grid Search for Model Comparison. The default activation functionis the Rectifier. Each of these activation functions can be operated with dropoutregularization. For more information, refer to Regularization.Specify the one of the following distribution functions for the response variableusing the distribution argument: Poisson Gamma Tweedie AUTO Bernoulli Multinomial LaplaceQuantileHuberGaussianEach distribution has a primary association with a particular loss function, butsome distributions allow users to specify a non-default loss function from thegroup of loss functions specified in Table 2. Bernoulli and multinomial areprimarily associated with cross-entropy (also known as log-loss), Gaussian withMean Squared Error, Laplace with Absolute loss (a special case of Quantile withquantile alpha 0.5) and Huber with Huber loss. For Poisson, Gamma,and Tweedie distributions, the loss function cannot be changed, so loss mustbe set to AUTO.The system default enforces the table’s typical use rule based on whetherregression or classification is being performed. Note here that t(j) and o(j) arethe predicted (also known as target) output and actual output, respectively, fortraining example j; further, let y represent the output units and O the outputlayer.Table 2: Loss functionsFunctionFormulaTypical useMean Squared ErrorAbsoluteL(W, B j) 12 kt(j) o(j) k22(j)(j)(L(W, B j) kt o k11 (j)(j) 2kt okforkt(j) o(j) k1 1,2L(W, B j) 2 (j)1(j)kt o k1 2 otherwise. P (j)(j)(j)(j)L(W, B j) ln(oy ) · ty ln(1 oy ) · (1 ty )RegressionRegressionHuberCross EntropyRegressionClassificationy OTo predict the 80-th percentile of the petal length of the Iris dataset in R, usethe following:

14 H2O’s Deep Learning ArchitectureExample in R12345678library(h2o)h2o.init(nthreads -1)train.hex - zonaws.com/smalldata/iris/iris wheader.csv")splits - h2o.splitFrame(train.hex, 0.75, seed 1234)dl - h2o.deeplearning(x 1:3, y "petal len",training frame splits[[1]],distribution "quantile", quantile alpha 0.8)h2o.predict(dl, splits[[2]])To predict the 80-th percentile of the petal length of the Iris dataset in Python,use the following:Example in Python12345678import h2ofrom h2o.estimators.deeplearning importH2ODeepLearningEstimatorh2o.init()train h2o.import m/smalldata/iris/iris wheader.csv")splits train.split frame(ratios [0.75], seed 1234)dl H2ODeepLearningEstimator(distribution "quantile",quantile alpha 0.8)dl.train(x range(0,2), y "petal len", training frame splits[0])print(dl.predict(splits[1]))

H2O’s Deep Learning Architecture 155.2.3Parallel Distributed Network TrainingThe process of minimizing the loss function L(W, B j) is a parallelized versionof stochastic gradient descent (SGD). A summary of standard SGD providedbelow, with the gradient L(W, B j) computed via backpropagation (LeCunet al, 1998). The constant α is the learning rate, which controls the step sizesduring gradient descent.Standard stochastic gradient descent1. Initialize W, B2. Iterate until convergence criterion reached:a. Get training example ib. Update all weights wjk W , biases bjk Bwjk : wjk α L(W,B j) wjkbjk : bjk α L(W,B j) bjkStochastic gradient descent is fast and memory-efficient but not easily parallelizable without becoming slow. We utilize Hogwild!, the recently developedlock-free parallelization scheme from Niu et al, 2011, to address this issue.Hogwild! follows a shared memory model where multiple cores (whereeach core handles separate subsets or all of the training data) are able to makeindependent contributions to the gradient updates L(W, B j) asynchronously.In a multi-node system, this parallelization scheme works on top of H2O’sdistributed setup that distributes the training data across the cluster. Eachnode operates in parallel on its local data until the final parameters W, B areobtained by averaging.

16 H2O’s Deep Learning ArchitectureParallel distributed and multi-threaded training with SGD in H2O Deep Learning1. Initialize global model parameters W, B2. Distribute training data T across nodes (can be disjoint or replicated)3. Iterate until convergence criterion reached:3.1. For nodes n with training subset Tn , do in parallel:a. Obtain copy of the global model parameters Wn , Bnb. Select active subset Tna Tn(user-given number of samples per iteration)c. Partition Tna into Tnac by cores ncd. For cores nc on node n, do in parallel:i. Get training example i Tnacii. Update all weights wjk Wn , biases bjk Bnwjk : wjk α L(W,B j) wjkbjk : bjk α L(W,B j) bjk3.2. Set W, B : Avgn Wn , Avgn Bn3.3. Optionally score the model on (potentially sampled)train/validation scoring setsHere, the weights and bias updates follow the asynchronous Hogwild! procedure to incrementally adjust each node’s parameters Wn , Bn after seeing theexample i. The Avgn notation represents the final averaging of these localparameters across all nodes to obtain the global model parameters and completetraining.

H2O’s Deep Learning Architecture 175.2.4Specifying the Number of Training SamplesH2O Deep Learning is scalable and can take advantage of large clusters ofcompute nodes. There are three operating modes. The default behavior allowsevery node to train on the entire (replicated) dataset but automatically shuffling(and/or using a subset of) the training examples for each iteration locally.For datasets that don’t fit into each node’s memory (depending on the amountof heap memory specified by the -Xmx Java option), it might not be possibleto replicate the data, so each compute node can be specified to train only withlocal data. An experimental single node mode is available for cases where finalconvergence is slow due to the presence of too many nodes, but this has notbeen necessary in our testing.To specify the global number of training examples shared with the distributedSGD worker nodes between model averaging, use thetrain samples per iteration parameter. If the specified value is -1,all nodes process all their local training data on each iteration.If replicate training data is enabled, which is the default setting, thiswill result in training N epochs (passes over the data) per iteration on N nodes;otherwise, one epoch will be trained per iteration. Specifying 0 always resultsin one epoch per iteration regardless of the number of compute nodes. Ingeneral, this parameter supports any positive number. For large datasets, werecommend specifying a fraction of the dataset.A value of -2, which is the default value, enables auto-tuning for this parameterbased on the computational performance of the processors and the networkof the system and attempts to find a good balance between computation andcommunication. This parameter can affect the convergence rate during training.For example, if the training data contains 10 million rows, and the number oftraining samples per iteration is specified as 100, 000 when running on fournodes, then each node will process 25, 000 examples per iteration, and it willtake 40 distributed iterations to process one epoch.If the value is too high, it might take too long between synchronization andmodel convergence may be slow. If the value is too low, network communicationoverhead will dominate the runtime and computational performance will suffer.

18 H2O’s Deep Learning Architecture5.3RegularizationH2O’s Deep Learning framework supports regularization techniques to preventoverfitting. 1 (L1: Lasso) and 2 (L2: Ridge) regularization enforce the samepenalties as they do with other models: modifying the loss function so as tominimize loss:L0 (W, B j) L(W, B j) λ1 R1 (W, B j) λ2 R2 (W, B j).For 1 regularization, R1 (W, B j) is the sum of all 1 norms for the weightsand biases in the network; 2 regularization via R2 (W, B j) represents thesum of squares of all the weights and biases in the network. The constants λ1and λ2 are generally specified as very small (for example 10 5 ).The second type of regularization available for Deep Learning is a moderninnovation called dropout (Hinton et al., 2012). Dropout constrains the onlineoptimization so that during forward propagation for a given training example,each neuron in the network suppresses its activation with probability P , whichis usually less than 0.2 for input neurons and up to 0.5 for hidden neurons.There are two effects: as with 2 regularization, the network weight values arescaled toward 0. Although they share the same global parameters, each trainingexample trains a different model. As a result, dropout allows an exponentiallylarge number of models to be averaged as an ensemble to help prevent overfittingand improve generalization.If the feature space is large and noisy, specifying an input dropout using theinput dropout ratio parameter can be especially useful. Note that input dropout can be specified independently of the dropout specification inthe hidden layers (which requires activation to be TanhWithDropout,MaxoutWithDropout, or RectifierWithDropout). Specify the amountof hidden dropout per hidden layer using the hidden dropout ratios parameter, which is set to 0.5 by default.5.4Advanced OptimizationH2O features manual and automatic advanced optimization modes. The manualmode features include momentum training and learning rate annealing and theautomatic mode features an adaptive learning rate.

H2O’s Deep Learning Architecture 195.4.1Momentum TrainingMomentum modifies back-propagation by allowing prior iterations to influencethe current version. In particular, a velocity vector, v, is defined to modify theupdates as follows: θ represents the parameters W, B µ represents the momentum coefficient α represents the learning ratevt 1 µvt α L(θt )θt 1 θt vt 1Using the momentum parameter can aid in avoiding local minima and anyassociated instability (Sutskever et al, 2014). Too much momentum can leadto instability, so we recommend incrementing the momentum slowly. The parameters that control momentum are momentum start, momentum ramp,and momentum stable.When using momentum updates, we recommend using the Nesterov accelerated gradient method, which uses the nesterov accelerated gradientparameter. This method modifies the updates as follows:vt 1 µvt α L(θt µvt )Wt 1 Wt vt 15.4.2Rate AnnealingDuring training, the chance of oscillation or “optimum skipping” creates theneed for a slower learning rate as the model approaches a minimum. As opposedto specifying a constant learning rate α, learning rate annealing graduallyreduces the learning rate αt to “freeze” into local minima in the optimizationlandscape (Zeiler, 2012).For H2O, the annealing rate (rate annealing) is the inverse of the numberof training samples required to divide the learning rate in half (e.g., 10 6 meansthat it takes 106 training samples to halve the learning rate).

20 H2O’s Deep Learning Architecture5.4.3Adaptive LearningThe implemented adaptive learning rate algorithm ADADELTA (Zeiler, 2012)automatically combines the benefits of learning rate annealing and momentumtraining to avoid slow convergence. To simplify hyper parameter search, specifyonly ρ and .In some cases, a manually controlled (non-adaptive) learning rate and momentum specifications can lead to better results but require a hyperparameter searchof up to seven parameters. If the model is built on a topology with many localminima or long plateaus, a constant learning rate may produce sub-optimalresults. However, the adaptive learning rate generally produces the best resultsduring our testing, so this option is the default.The first of two hyper parameters for adaptive learning is ρ (rho). It is similarto momentum and is related to the memory of prior weight updates. Typicalvalues are between 0.9 and 0.999. The second hyper parameter, (epsilon),is similar to learning rate annealing during initial training and allows furtherprogress during momentum at later stages. Typical values are between 10 10and 10 4 .5.5Loading DataLoading a dataset in R or Python for use with H2O is slightly different thanthe usual methodology. Instead of using data.frame or data.table inR, or pandas.DataFrame or numpy.array in Python, datasets must beconverted into H2OFrame objects (distributed data frames).5.5.1Data Standardization/NormalizationAlong with categorical encoding, H2O’s Deep Learning preprocesses the datato standardize it for compatibility with the activation functions (refer to to thesummary of each activation function’s target space in Activation Functions).Since the activation function generally does not map into the full spectrumof real numbers, R, we first standardize our data to be drawn from N (0, 1).Standardizing again after network propagation allows us to compute moreprecise errors in this standardized space, rather than in the raw feature space.For autoencoding, the data is normalized (instead of standardized) to thecompact interval of U( 0.5, 0.5) to allow bounded activation functions liketanh to better reconstruct the data.

H2O’s Deep Learning Architecture 215.5.2Convergence-based Early StoppingEarly stopping based on convergence of a user-specified metric is an especiallyhelpful feature for finding the optimal number of epochs. By default, it usesthe metrics on the validation dataset, if provided. Otherwise, training metricsare used. To stop model building if misclassification improves (is reduced) by lessthan one percent between individual scoring epochs, specifystopping rounds 1, stopping tolerance 0.01 andstopping metric "misclassification". To stop model building if the simple moving average (window length 5) ifthe AUC improves (increases) by less than 0.1 percent for 5 consecutivescoring epochs, use stopping rounds 5, stopping metric "AUC",and stopping tolerance 0.001. To stop model building if the logloss on the validation set does not improveat all for 3 consecutive scoring epochs, specify a validation frame,stopping rounds 3, stopping tolerance 0 andstopping metric "logloss". To continue model building even after metrics have converged, disablethis feature using stopping rounds 0. To compute the best number of epochs with cross-validation, simplyspecify stopping rounds 0 as in the examples above, in combinationwith nfolds 1, and the main model will pick the ideal number of epochsfrom the convergence behavior of the nfolds cross-validation models.5.5.3Time-based Early StoppingTo stop model training after a given amount of seconds, specify max runtime secs 0. This option is also available for grid searches and models with crossvalidat

Deep Learning tasks. Deep Learning architectures are models of hierarchical feature extraction, typically involving multiple levels of nonlinearity. Deep Learning models are able to learn useful representations of raw data and have exhibited high performance on comp