Abstract: An embodiment may include a processor that identifies a plurality of weights from the propositional logical neural network. The embodiment may convert the plurality of weights into a sparse matrix. The embodiment may convert a training set into a plurality of bound vectors. The embodiment may update the sparse matrix using a graphical processing unit (GPU). The embodiment may compute a loss parameter and based on determining the loss function is below threshold, update the plurality of weights of the propositional neural network.

Abstract: Methods, systems, and computer program products for accelerating inference of transformer-based models are provided herein. A computer-implemented method includes obtaining a machine learning model comprising a plurality of transformer blocks, a task, and a natural language dataset; generating a compressed version of the machine learning model based on the task and the natural language dataset, wherein the generating comprises: obtaining at least one set of tokens, wherein each token in the set corresponds to one of the items in the natural language dataset, identifying and removing one or more redundant output activations of different ones of the plurality of transformer blocks for the at least one set of tokens, and adding one or more input activations corresponding to the one or more removed output activations into the machine learning model at subsequent ones of the plurality of the transformer blocks; and outputting the compressed version of the machine learning model to at least one user.

Abstract: Methods, systems, and computer program products for cold-start service placement over on-demand resources are provided herein. A computer-implemented method includes obtaining a performance requirement profile comprising performance requirements of a service that vary over time; determining a plurality of incarnations for the service, wherein each incarnation is associated with a level of performance provided by the incarnation for the service, resource requirements of the incarnation, and a type of computing node the incarnation is configured to execute on; identifying computing nodes having different types and different resource capacities; jointly scheduling (i) the computing nodes and (ii) one or more of the incarnations on the computing nodes over a time interval such that a cumulative level of performance of the incarnations scheduled at each timepoint in the time interval satisfies the performance requirement profile of the service.

Abstract: Techniques for utilizing model and hyperparameter optimization for multi-objective machine learning are disclosed. In one example, a method comprises the following steps. One of a plurality of hyperparameter optimization operations and a plurality of model parameter optimization operations are performed to generate a first solution set. The other of the plurality of hyperparameter optimization operations and the plurality of model parameter optimization operations are performed to generate a second solution set. At least a portion of the first solution set and at least a portion of the second solution set are combined to generate a third solution set.

Abstract: A computer-implemented machine learning model training method and resulting machine learning model. One embodiment of the method may comprise receiving at a computer memory training data; and training on a computer processor a machine learning model on the received training data using a plurality of batch sizes to produce a trained processor. The training may include calculating a plurality of activations during a forward pass of the training and discarding at least some of the calculated plurality of activations after the forward pass of the training.

Abstract: One embodiment provides a method, including: receiving at least one deep learning job for scheduling and running on a distributed system comprising a plurality of nodes; receiving a batch size range indicating a minimum batch size and a maximum batch size that can be utilized for running the at least one deep learning job; determining a plurality of runtime estimations for running the at least one deep learning job; creating a list of optimal combinations of (i) batch sizes and (ii) numbers of the plurality of nodes for running both (a) the at least one deep learning job and (b) current deep learning jobs; and scheduling the at least one deep-learning job at the distributed system, responsive to identifying, by utilizing the list, that the distributed system has necessary processing resources for running both (iii) the at least one deep learning job and (iv) the current deep learning jobs.

Abstract: Methods, systems, and computer program products for accelerating inference of transformer-based models are provided herein. A computer-implemented method includes obtaining a machine learning model comprising a plurality of transformer blocks, a task, and a natural language dataset; generating a compressed version of the machine learning model based on the task and the natural language dataset, wherein the generating comprises: obtaining at least one set of tokens, wherein each token in the set corresponds to one of the items in the natural language dataset, identifying and removing one or more redundant output activations of different ones of the plurality of transformer blocks for the at least one set of tokens, and adding one or more input activations corresponding to the one or more removed output activations into the machine learning model at subsequent ones of the plurality of the transformer blocks; and outputting the compressed version of the machine learning model to at least one user.

Abstract: Methods, systems, and computer program products for accelerating inference of neural network models via dynamic early exits are provided herein. A computer-implemented method includes determining a plurality of candidate exit points of a neural network model; obtaining a plurality of outputs of the neural network model for data samples in a target dataset, wherein the plurality of outputs comprises early outputs of the neural network model from the plurality of candidate exit points and regular outputs of the neural network model; and a set of one or more exit points from the plurality of candidate exits points that are dependent on the target dataset based at least in part on the plurality of outputs.

Abstract: Methods, systems, and computer program products for determining operating range of hyperparameters are provided herein. A computer-implemented method includes obtaining a machine learning model, a list of candidate values for a hyperparameter, and a dataset; performing one or more hyperparameter range trials based on the machine learning model, the list of candidate values for the hyperparameter, and the dataset; automatically determining an operating range of the hyperparameter based on the one or more hyperparameter range trials; and training the machine learning model to convergence based at least in part on the determined operating range.

Abstract: A system, computer program product, and method are presented for performing multi-objective automated machine learning, and, more specifically, to identifying a plurality of machine learning pipelines as Pareto-optimal solutions to optimize a plurality of objectives. The method includes receiving input data directed toward one or more subjects of interest and determining a plurality of objectives to be optimized. The method also includes ingesting at least a portion of the input data through one or more machine learning (ML) models. The method further includes aggregating the plurality of objectives into one or more aggregated single objectives. The method also includes determining a plurality of Pareto-optimal solutions, thereby defining a plurality of ML pipelines that optimize the one or more aggregated single objectives. The method further includes selecting one ML pipeline from the plurality of ML pipelines.

Abstract: One or more computer processors decompose a weight matrix associated with a neural network utilizing a permutation dependent decomposition. The one or more computer processors regenerate a recovered matrix utilizing the decomposed weight matrix. The one or more computer processors reduce an error between the decomposed weight matrix and regenerated recovered matrix.

Abstract: Methods, systems, and computer program products for determining optimal compute resources for distributed batch based optimization applications are provided herein. A method includes obtaining a size of an input dataset, a size of a model, and a set of batch sizes corresponding to a job to be processed using a distributed computing system; computing, based at least in part on the set of batch sizes, one or more node counts corresponding to a number of nodes that can be used for processing said job; estimating, for each given one of the node counts, an execution time to process the job based on an average computation time for a batch of said input dataset and an average communication time for said batch of said input dataset; and selecting, based at least in part on said estimating, at least one of said node counts for processing the job.

Abstract: One embodiment provides a computer implemented method, including: receiving an application including a plurality of services to be hosted on a remote services environment; obtaining, a resource profile identifying a usage of a resource by a given service of the plurality of services over a period of time; splitting, based upon the resource profile corresponding to a given service, the given service into a plurality of service slices; selecting, for each of the plurality of service slices an incarnation fulfilling a resource demand requirement and a service-performance-offering, wherein an incarnation has a total demand value based upon a resource capacity of a node at the remote services environment; and assigning, for each of the plurality of service slices the incarnation selected to at least one node within the remote services environment based upon the resource capacity of the at least one node.

Abstract: A computer-implemented machine learning model training method and resulting machine learning model. One embodiment of the method may comprise receiving at a computer memory training data; and training on a computer processor a machine learning model on the received training data using a plurality of batch sizes to produce a trained processor. The training may include calculating a plurality of activations during a forward pass of the training and discarding at least some of the calculated plurality of activations after the forward pass of the training.

Abstract: One embodiment provides a method, including: receiving at least one deep learning job for scheduling and running on a distributed system comprising a plurality of nodes; receiving a batch size range indicating a minimum batch size and a maximum batch size that can be utilized for running the at least one deep learning job; determining a plurality of runtime estimations for running the at least one deep learning job; creating a list of optimal combinations of (i) batch sizes and (ii) numbers of the plurality of nodes for running both (a) the at least one deep learning job and (b) current deep learning jobs; and scheduling the at least one deep-learning job at the distributed system, responsive to identifying, by utilizing the list, that the distributed system has necessary processing resources for running both (iii) the at least one deep learning job and (iv) the current deep learning jobs.

Abstract: Methods, systems, and computer program products for determining operating range of hyperparameters are provided herein. A computer-implemented method includes obtaining a machine learning model, a list of candidate values for a hyperparameter, and a dataset; performing one or more hyperparameter range trials based on the machine learning model, the list of candidate values for the hyperparameter, and the dataset; automatically determining an operating range of the hyperparameter based on the one or more hyperparameter range trials; and training the machine learning model to convergence based at least in part on the determined operating range.

Abstract: In an approach to determining an optimal training approach for a large deep learning model based on model characteristics and system characteristics. The one or more computer processors identify one or more model characteristics associated with a deep learning model. The one or more computer processors identify one or more system configurations associated with a system training the deep learning model. The one or more computer processors determine a training approach for the deep learning model utilizing a trained large model predictor fed with the one or more identified model characteristics and the one or more identified system configurations. The one or more computer processors train the deep learning model utilizing the determined training approach.

Abstract: Methods and systems for managing energy sources and energy consumers in an integrated system are provided. Certain subject matter presented herein relates to automatically scheduling jobs and their sub-tasks to maximize profit by comparing power source configurations and determining the best job schedule for the power source configurations. This system broadly involves two sub-problems: determining the best power source configuration and determining the best job schedule for the given power source configuration.

Abstract: Methods, systems, and computer program products for determining optimal compute resources for distributed batch based optimization applications are provided herein. A method includes obtaining a size of an input dataset, a size of a model, and a set of batch sizes corresponding to a job to be processed using a distributed computing system; computing, based at least in part on the set of batch sizes, one or more node counts corresponding to a number of nodes that can be used for processing said job; estimating, for each given one of the node counts, an execution time to process the job based on an average computation time for a batch of said input dataset and an average communication time for said batch of said input dataset; and selecting, based at least in part on said estimating, at least one of said node counts for processing the job.

Abstract: Methods, systems, and computer program products for dataset dependent low rank decomposition of neural networks are provided herein. A computer-implemented method includes obtaining a target dataset and a trained model of a neural network; providing at least a portion of the target dataset to the trained model; determining relevance of each of one or more of filters of the neural network and channels of the neural network to the target dataset based on the provided portion, wherein the one or more of the filters and the channels correspond to at least one layer of the neural network; and compressing the trained model of the neural network based at least in part on the determined relevancies.