Oracle International Corporation

Sandeep Agrawal

Nipun Agarwal

Sam Idicula

Venkatanathan Varadarajan

Anatoly Yakovlev

Hesam Fathi Moghadam

Ali Moharrer

Gaurav Chadha

Jingxiao Cai

Sanjay Jinturkar

Arun Raghavan

Nikan Chavoshi

Techniques are described herein for performing efficient matrix multiplication in architectures with scratchpad memories or associative caches using asymmetric allocation of space for the different matrices. The system receives a left matrix and a right matrix. In an embodiment, the system allocates, in a scratchpad memory, asymmetric memory space for tiles for each of the two matrices as well as a dot product matrix. The system proceeds with then performing dot product matrix multiplication involving the tiles of the left and the right matrices, storing resulting dot product values in corresponding allocated dot product matrix tiles. The system then proceeds to write the stored dot product values from the scratchpad memory into main memory.

Asymmetric allocation of SRAM and data layout for efficient matrix-matrix multiplication

Herein, a computer generates and evaluates many preprocessor configurations for a window preprocessor that transforms a training timeseries dataset for an ML model. With each preprocessor configuration, the window preprocessor is configured. The window preprocessor then converts the training timeseries dataset into a configuration-specific point-based dataset that is based on the preprocessor configuration. The ML model is trained based on the configuration-specific point-based dataset to calculate a score for the preprocessor configuration. Based on the scores of the many preprocessor configurations, an optimal preprocessor configuration is selected for finally configuring the window preprocessor, after which, the window preprocessor can optimally transform a new timeseries dataset such as in an offline or online production environment such as for real-time processing of a live streaming timeseries.

Automated machine learning pipeline for timeseries datasets utilizing point-based algorithms

Herein are techniques that train regressor(s) to predict how effective would a machine learning model (MLM) be if trained with new hyperparameters and/or dataset. In an embodiment, for each training dataset, a computer derives, from the dataset, values for dataset metafeatures. The computer performs, for each hyperparameters configuration (HC) of a MLM, including landmark HCs: configuring the MLM based on the HC, training the MLM based on the dataset, and obtaining an empirical quality score that indicates how effective was said training the MLM when configured with the HC. A performance tuple is generated that contains: the HC, the values for the dataset metafeatures, the empirical quality score and, for each landmark configuration, the empirical quality score of the landmark configuration and/or the landmark configuration itself. Based on the performance tuples, a regressor is trained to predict an estimated quality score based on a given dataset and a given HC.

Using metamodeling for fast and accurate hyperparameter optimization of machine learning and deep learning models

Techniques are described for generating and applying mini-machine learning variants of machine learning algorithms to save computational resources in tuning and selection of machine learning algorithms. In an embodiment, at least one of the hyper-parameter values for a reference variant is modified to a new hyper-parameter value thereby generating a new variant of machine learning algorithm from the reference variant of machine learning algorithm. A performance score is determined for the new variant of machine learning algorithm using a training dataset, the performance score representing the accuracy of the new machine learning model for the training dataset. By performing training of the new variant of machine learning algorithm with the training data set, a cost metric of the new variant of machine learning algorithm is measured by measuring usage the used computing resources for the training. Based on the cost metric of the new variant of machine learning algorithm and comparing the performance score for the new and reference variants, the system determines whether the modified reference machine algorithm is the mini-machine learning algorithm that is computationally less costly than the reference variant of machine learning algorithm but closely tracks the accuracy thereof.

Mini-machine learning

Herein, horizontally scalable techniques efficiently configure machine learning algorithms for optimal accuracy and without informed inputs. In an embodiment, for each particular hyperparameter, and for each epoch, a computer processes the particular hyperparameter. An epoch explores one hyperparameter based on hyperparameter tuples. A respective score is calculated from each tuple. The tuple contains a distinct combination of values, each of which is contained in a value range of a distinct hyperparameter. All values of a tuple that belong to the particular hyperparameter are distinct. All values of a tuple that belong to other hyperparameters are held constant. The value range of the particular hyperparameter is narrowed based on an intersection point of a first line based on the scores and a second line based on the scores. A machine learning algorithm is optimally configured from repeatedly narrowed value ranges of hyperparameters. The configured algorithm is invoked to obtain a result.

Gradient-based auto-tuning for machine learning and deep learning models

Techniques are provided for selection of machine learning algorithms based on performance predictions by using hyperparameter predictors. In an embodiment, for each mini-machine learning model (MML model), a respective hyperparameter predictor set that predicts a respective set of hyperparameter settings for a data set is trained. Each MML model represents a respective reference machine learning model (RML model). Data set samples are generated from the data set. Meta-feature sets are generated, each meta-feature set describing a respective data set sample. A respective target set of hyperparameter settings are generated for said each MML model using a hypertuning algorithm. The meta-feature sets and the respective target set of hyperparameter settings are used to train the respective hyperparameter predictor set. Each hyperparameter predictor set is used during training and inference to improve the accuracy of automatically selecting a RML model per data set.

Using hyperparameter predictors to improve accuracy of automatic machine learning model selection

According to an embodiment, a method includes generating a first dataset sample from a dataset, calculating a first validation score for the first dataset sample and a machine learning model, and determining whether a difference in validation score between the first validation score and a second validation score satisfies a first criteria. If the difference in validation score does not satisfy the first criteria, the method includes generating a second dataset sample from the dataset. If the difference in validation score does satisfy the first criteria, the method includes updating a convergence value and determining whether the updated convergence value satisfies a second criteria. If the updated convergence value satisfies the second criteria, the method includes returning the first dataset sample. If the updated convergence value does not satisfy the second criteria, the method includes generating the second dataset sample from the dataset.

Adaptive sampling for imbalance mitigation and dataset size reduction in machine learning

Techniques are provided for selection of machine learning algorithms based on performance predictions by trained algorithm-specific regressors. In an embodiment, a computer derives meta-feature values from an inference dataset by, for each meta-feature, deriving a respective meta-feature value from the inference dataset. For each trainable algorithm and each regression meta-model that is respectively associated with the algorithm, a respective score is calculated by invoking the meta-model based on at least one of: a respective subset of meta-feature values, and/or hyperparameter values of a respective subset of hyperparameters of the algorithm. The algorithm(s) are selected based on the respective scores. Based on the inference dataset, the selected algorithm(s) may be invoked to obtain a result. In an embodiment, the trained regressors are distinctly configured artificial neural networks. In an embodiment, the trained regressors are contained within algorithm-specific ensembles. Techniques are also provided for optimal training of regressors and/or ensembles.

Algorithm-specific neural network architectures for automatic machine learning model selection

Herein are techniques for exploring hyperparameters of a machine learning model (MLM) and to train a regressor to predict a time needed to train the MLM based on a hyperparameter configuration and a dataset. In an embodiment that is deployed in production inferencing mode, for each landmark configuration, each containing values for hyperparameters of a MLM, a computer configures the MLM based on the landmark configuration and measures time spent training the MLM on a dataset. An already trained regressor predicts time needed to train the MLM based on a proposed configuration of the MLM, dataset meta-feature values, and training durations and hyperparameter values of landmark configurations of the MLM. When instead in training mode, a regressor in training ingests a training corpus of MLM performance history to learn, by reinforcement, to predict a training time for the MLM for new datasets and/or new hyperparameter configurations.

Predicting machine learning or deep learning model training time

Assymetric allocation of SRAM and data layout for efficient matrix multiplication

Herein are techniques for automatic tuning of hyperparameters of machine learning algorithms. System throughput is maximized by horizontally scaling and asynchronously dispatching the configuration, training, and testing of an algorithm. In an embodiment, a computer stores a best cost achieved by executing a target model based on best values of the target algorithm's hyperparameters. The best values and their cost are updated by epochs that asynchronously execute. Each epoch has asynchronous costing tasks that explore a distinct hyperparameter. Each costing task has a sample of exploratory values that differs from the best values along the distinct hyperparameter. The asynchronous costing tasks of a same epoch have different values for the distinct hyperparameter, which accomplishes an exploration. In an embodiment, an excessive update of best values or best cost creates a major epoch for exploration in a subspace that is more or less unrelated to other epochs, thereby avoiding local optima.

Scalable and efficient distributed auto-tuning of machine learning and deep learning models

Techniques related to matrix multiplication at memory bandwidth are disclosed. Computing device(s) perform multiplication of a first matrix with a second matrix to generate a third matrix. A first register stores contiguous element values of the first matrix. Furthermore, a second register stores a first set of contiguous element values of the second matrix, and a third register stores a second set of contiguous element values of the second matrix. The first set and the second set correspond to a first row and a second row, respectively, of the second matrix. The first row and the second row are contiguous rows. A single instruction is executed to cause at least a partial computation of contiguous element values of the third matrix. The single instruction causes multiplication of element values stored in the first register with element values stored in the second and third registers and grouped accumulation of the products.

Matrix multiplication at memory bandwidth

Techniques related to memory management for sparse matrix multiplication are disclosed. Computing device(s) may perform a method for multiplying a row of a first sparse matrix with a second sparse matrix to generate a product matrix row. A compressed representation of the second sparse matrix is stored in main memory. The compressed representation comprises a values array that stores non-zero value(s). Tile(s) corresponding to row(s) of second sparse matrix are loaded into scratchpad memory. The tile(s) comprise set(s) of non-zero value(s) of the values array. A particular partition of an uncompressed representation of the product matrix row is generated in the scratchpad memory. The particular partition corresponds to a partition of the second sparse matrix comprising non-zero value(s) included in the tile(s). When a particular tile is determined to comprise non-zero value(s) that are required to generate the particular partition, the particular tile is loaded into the scratchpad memory.

Growing community of inventors

San Jose, CA, United States of America

Sandeep Agrawal