Emc IP Holding Company LLC

Dell Products Inc.

Tiago Salviano Calmon

Eduardo Vera Sousa

Vinicius Michel Gottin

Vinícius Michel Gottin

Hugo De Oliveira Barbalho

Jonas Furtado Dias

Paulo Abelha Ferreira

Pablo Nascimento Da Silva

Adriana Bechara Prado

Roberto Nery Stelling Neto

Rômulo Teixeira De Abreu Pinho

Romulo D Pinho

Flavia Coimbra Delicato

Paulo F Pires

Igor Leão Dos Santos

Alex Laier Bordignon

Daniel Sadoc Menasché

Ana Cristina Bernardo De Oliveira

Diego Salomone Bruno

Percy E Rivera Salas

Owen Martin

John T Cardente

Edward José Pacheco Condori

Victor Bursztyn

Marcelo Pitanga

Ana C Oliveira

Amit Bhaya

Oumar Diene

Tiago Cariolano De Souza Xavier

Jonathan Ferreira Passoni

Method for model updating in a federated learning environment, including distributing a current model to client nodes; receiving a first set of gradient sign vectors, wherein each gradient sign vector of the first set of gradient sign vectors is received from one client node; generating a first updated model based on the first set of gradient sign vectors; distributing the first updated model to the plurality of client nodes; storing a first shape parameter and a second shape parameter; receiving, in response to distributing the first updated model, a second set of gradient sign vectors, wherein each gradient sign vector of the second set of gradient sign vectors is received from one client node; generating a second updated model based on the second set of gradient sign vectors, the first shape parameter, and the second shape parameter; and distributing the second updated model to the plurality of client nodes.

Fast converging gradient compressor for federated learning

Methods and apparatus are provided for real-time anomaly detection over sets of time-series data. One method comprises: obtaining a state-space representation of a plurality of states and transitions between said states based on sets of historical time-series data; obtaining an anomaly detection model trained using a supervised learning technique, wherein the anomaly detection model associates sequences of states in the state-space representation with annotated anomalies in the sets of historical time-series data and assigns a probability to said sequences of states; and, for incoming real-time time-series data, determining a likelihood of a current state belonging to a plurality of possible states in the state-space representation; and determining a probability of incurring said annotated anomalies based on a plurality of likely current state sequences that satisfy a predefined likelihood criteria. Anomalous behavior is optionally distinguished from previously unknown behavior based on a predefined likelihood threshold.

Methods and apparatus for real-time anomaly detection over sets of time-series data

A framework for rapidly prototyping federated learning algorithms. Specifically, the disclosed framework proposes a method and system for evaluating different hypotheses for configuring learning model, which may be optimized through federated learning. Through the disclosed framework, these hypotheses may be tested for scalability, hardware and network resource performance, as well as for new learning state compression and/or aggregation technique effectiveness. Further, these hypotheses may be tested through federated learning simulations, which avoid costs associated with deploying these hypotheses to be tested across production systems.

Framework for rapidly prototyping federated learning algorithms

Techniques are provided for allocating resources for one or more workloads. One method comprises obtaining a current performance of a workload; determining an adjustment to a current allocation of a resource allocated to the workload by evaluating a representation of a relationship between: (i) the current allocation of the resource allocated to the workload, (ii) a performance metric, and (iii) the current performance of the workload; and initiating an application of the determined adjustment to the current allocation of the resource for the workload. The performance metric may comprise a nominal value of a predefined service metric and the current performance of the workload may comprise a current value of a variable that tracks a given predefined service metric of the workload. An amount (or percentage) of the adjustment permitted for each iteration may be controlled. A sum of allocated resources can be constrained to an amount of available resources.

Resource adaptation using nonlinear relationship between system performance metric and resource usage

Methods for compression switching that includes distributing a model to client nodes, which use the model to generate a gradient vector (GV) based on a client node data set. The method includes receiving a model update that includes a gradient sign vector (GSV) based on the gradient vector; generating an updated model using the GSV; and distributing the updated model to the client nodes. The client node uses the updated model to generate a second GV based on a second client node data set. The method also includes a determination that a compression switch condition exists; based on the determination, transmitting an instruction to the client node to perform a compression switch; receiving, in response to the instruction, another model update including a subset GSV based on the second gradient vector; generating a second updated model using the subset GSV; and distributing the second updated model to the client nodes.

Compression switching for federated learning

Techniques described herein relate to a method for managing data of data nodes. The method includes obtaining, by a data node manager, a soft labeling request; in response to obtaining the soft labeling request: sending, by the data node manager, requests for processed data to data nodes associated with the data node manager; obtaining, by the data node manager, processed data from the data nodes; merging, by the data node manager, the processed data to obtain processed data; performing, by the data node manager, clustering on the processed data to obtain soft label metadata; associating, by the data node manager, the soft label metadata with live data associated with the data nodes; and performing, by the data node manager, labeling actions using the live data and the soft label metadata.

Method and system for generating and assigning soft labels for data node data

An information handling system for managing detection of objects includes a storage and a processor. The storage is for storing an encoder; a critical class classifier; a general classifier; and a decoder. The processor obtains data that may include one or more of the objects; encodes the data using the encoder to obtain encoded data; obtains a critical class classification for the encoded data using the critical class classifier; obtains a general classification for the encoded data using the general classifier; conditions the encoded data to obtain conditioned encoded data; decodes the conditioned encoded data using the decoder to obtain reconstructed data; makes a determination that the reconstructed data and the critical class classification indicate that the data is an unknown classification; classifies the data as being an unknown classification based on the determination; and performs an action set based on the unknown classification of the data.

System and method for object detection

Techniques are provided for assigning workloads in a multi-node processing environment using resource allocation feedback from each node. One method comprises obtaining feedback from distributed nodes that process workloads, wherein the feedback for a given node indicates (i) an allocation of resources, and (ii) a number of executing workloads. In response to receiving a given workload to be processed, candidate nodes are identified to execute the given workload; and the given workload is assigned to a given candidate node based on an amount of available resources on each candidate node and/or a stability of resource adjustments made for each candidate node. The stability of the resource adjustments made for each candidate node can be evaluated based on a maximum resource adjustment made for a given candidate node relative to a maximum resource adjustment made for each of the candidate nodes.

Assigning workloads in a multi-node processing environment using feedback from each node

A global prediction manager for generating predictions using data from data zones includes storage for storing a model repository comprising a global model set and a prediction manager. The prediction manager obtains a local model set from a data zone of the data zones indicating that the global model set is unacceptable; makes a determination that the local model set is acceptable; in response to the determination: distributes the local model set to at least one second data zone of the data zones; obtains compressed telemetry data, that was compressed using the local model set, from the data zone and the at least one second data zone; and generates a global prediction regarding a future operating condition of the data zones using: the compressed local telemetry data and the local model set.

Locality-aware compressor-decompressor for keeping prediction models up-to-date in resource constrained networks

Techniques are provided for adaptive resource allocation for multiple workloads. One method comprises obtaining a dynamic system model based on a relation between an amount of a resource for multiple iterative workloads and a predefined service metric; obtaining an instantaneous value of the predefined service metric; applying to a given controller associated with a given iterative workload of the multiple iterative workloads: (i) the dynamic system model, (ii) an interference effect of one or more additional iterative workloads on the given iterative workload, and (iii) a difference between the instantaneous value and a target value for the predefined service metric. The given controller applies an adjustment to the amount of the resource for the given iterative workload based at least in part on the difference. The resource allocation for the multiple iterative workloads can be performed in a sequence substantially in parallel with an execution of the iterative workloads.

Adaptation of resource allocation for multiple workloads using interference effect of resource allocation of additional workloads on performance

Techniques are provided for allocation of shared computing resources using source code feature extraction and machine learning techniques. An exemplary method comprises obtaining source code for execution in a shared computing environment; extracting a plurality of discriminative features from the source code; obtaining a trained machine learning model; and generating a prediction of an allocation of one or more resources of the shared computing environment needed to satisfy one or more service level agreement requirements for the source code. The generated prediction is optionally adjusted using a statistical analysis of an error curve, based on one or more error boundaries obtained by the trained machine learning model. The trained machine learning model can be trained using a set of discriminative features extracted from training source code and corresponding measurements of metrics of the service level agreement requirements obtained by executing the training source code on a plurality of the resources of the shared computing environment.

Allocation of shared computing resources using source code feature extraction and machine learning

A method and system for assessing sequentiality of a data stream is disclosed. Specifically, the method and system disclosed herein may entail receiving an incoming request to access a page in a cache memory, wherein the page is identified by a page address of an address space in a main memory; identifying, in a memory, a bin corresponding to an address range including the page address of the page of the incoming request, wherein the bin includes k address ranges of the address space of the main memory; determining whether to update an occupation count of the bin in the memory; locating the bin in a heuristics table to obtain an estimated total number of expected proximal accesses based on an updated occupation count of the bin; and determining, based on the estimated total number of expected proximal accesses, sequentiality of the data stream to device in order to generate a policy for the cache memory.

Online heuristic sequentiality detection over input/output streams for cache systems in large address spaces

Techniques are provided for predicting a time to complete a data protection operation. One method comprises obtaining metadata for (i) a given data protection appliance, and/or (ii) a cluster of similar data protection appliances comprising the given data protection appliance; evaluating first level features using the obtained metadata; evaluating a second level feature using some of the evaluated first level features; and processing one or more of the first level features, and the second level feature, using a model that provides a predicted time to complete a data protection operation with respect to data of a protected device associated with the given data protection appliance. The predicted time may comprise a tolerance based on a robustness factor. The predicted time may be based on a number of protected devices that are concurrently undergoing a data protection operation with the protected device for one or more time intervals.

Artificial intelligence techniques for prediction of data protection operation duration

Reinforcement learning agents for resource allocation for iterative workloads, such as training Deep Neural Networks, are configured. One method comprises obtaining a specification of an iterative workload comprising multiple states and a set of available actions for each state, and a domain model of the iterative workload relating allocated resources with service metrics; adjusting weights of a reinforcement learning agent by performing iteration steps for each simulated iteration of the iterative workload and using variables from the simulated iteration to refine the reinforcement learning agent; and determining a dynamic resource allocation policy for the iterative workload. The exemplary iteration steps comprise: (a) selecting an action for a current state, obtaining a reward for the selected action and selecting a next state based on the current state and/or the selected action; (b) updating a function that evaluates a quality of a plurality of state-action combinations; and (c) repeating steps (a) and (b) with a new allocation of resources.

Building neural networks for resource allocation for iterative workloads using reinforcement learning

Techniques are provided for allocating shared computing resources using source code feature extraction and cluster-based training of machine learning models. An exemplary method comprises: obtaining a source code corpus with source code segments for execution in a shared computing environment; extracting discriminative features from the source code segments in the source code corpus; obtaining a trained machine learning model, wherein the trained machine learning model is trained using samples of source code segments from clusters derived from clustering the source code corpus based on (i) a term frequency metric, and/or (ii) observed values of execution metrics; and generating, using the trained machine learning model, a prediction of an allocation of resources of the shared computing environment needed to satisfy service level agreement requirements for source code to be executed in the shared computing environment. The discriminative features may be extracted from the source code corpus using natural language processing techniques and/or pattern-based techniques.

Growing community of inventors

Rio de Janeiro, Brazil

Tiago Salviano Calmon