Hewlett Packard Enterprise Development LP

Sergey Serebryakov

Dejan S Milojicic

John Paul Strachan

Sai Rahul Chalamalasetti

Paolo Faraboschi

Catherine Graves

Martin Foltin

Tahir Cader

Harumi Kuno

Suparna Bhattacharya

Duncan Roweth

Alan Lynn Davis

Mehmet Kivanc Ozonat

Giacomo Pedretti

Ryan Dean Menhusen

John L Byrne

Torsten Wilde

Daniel William Dauwe

Aalap Tripathy

Annmary Justine Koomthanam

Vipin Kumar Kukkala

Nanjundaiah Deepak

Deepak Nanjundaiah

Testing for bias in a machine learning (ML) model in a manner that is independent of the code/weights deployment path is described. If bias is detected, an alert for bias is generated, and optionally, the ML model can be incrementally re-trained to mitigate the detected bias. Re-training the ML model to mitigate the bias may include enforcing a bias cost function to maintain a level of bias in the ML model below a threshold bias level. One or more statistical metrics representing the level of bias present in the ML model may be determined and compared against one or more threshold values. If one or more metrics exceed corresponding threshold value(s), the level of bias in the ML model may be deemed to exceed a threshold level of bias, and re-training of the ML model to mitigate the bias may be initiated.

Machine learning model bias detection and mitigation

An example device includes processing circuitry and a memory. The memory includes instructions that cause the device to perform various functions. The functions include receiving datastreams from a plurality of sensors of a high performance computing system, classifying each datastream of the each sensor to one of a plurality of datastream models, selecting an anomaly detection algorithm from a plurality of anomaly detection algorithms for each datastream, determining parameters of the each anomaly detection algorithm, determining an anomaly threshold for each datastream, and generating an indication that the sensor associated with the datastream is acting anomalously.

Classification based anomaly detection

Systems and methods are provide for automatically constructing data lineage representations for distributed data processing pipelines. These data lineage representations (which are constructed and stored in a central repository shared by the multiple data processing sites) can be used to among other things, clone the distributed data processing pipeline for quality assurance or debugging purposes. Examples of the presently disclosed technology are able to construct data lineage representations for distributed data processing pipelines by (1) generating a hash content value for universally identifying each data artifact of the distributed data processing pipeline across the multiple processing stages/processing sites of the distributed data processing pipeline; and (2) creating an data processing pipeline abstraction hierarchy for associating each data artifact to input and output events for given executions of given data processing stages (performed by the multiple data processing sites).

Method to track and clone data artifacts associated with distributed data processing pipelines

Systems and methods are provided for an accelerator system that includes a baseline (production) accelerator, optimizing accelerator, and control hardware accelerator, and an operation of alternatingly switching the production/optimizing accelerators between production and optimizing. With two production/optimizing accelerators, at any given point in time, one accelerator adapts while another accelerator processes data. Once the second accelerator starts doing a better job (e.g., has adapted to data drift), the accelerators change their modes, and the trainable accelerator becomes the 'optimized' one. The accelerators do this non-stop, thus maintaining redundancy, providing expected quality of service (QoS) and adapting to data/concept drift.

Self-adaptable accelerators having alternating production/optimizing modes

Embodiments of the disclosure provide a system, method, or computer readable medium for providing a differentiable content addressable memory (aCAM) that implements an analog input analog storage and analog output learning memory. The analog output of the differentiable CAM can provide input to a learning algorithm, which may compute the gradients in comparison to historic values and reduce data inaccuracies and power consumption.

Differentiable content addressable memory

In exemplary aspects, a golden data structure can be used to validate the stability of machine learning (ML) models and weights. The golden data structure includes golden input data and corresponding golden output data. The golden output data represents the known correct results that should be output by a ML model when it is run with the golden input data as inputs. The golden data structure can be stored in a secure memory and retrieved for validation separately or together with the deployment of the ML model for a requested ML operation. If the golden data structure is used to validate the model and/or weights concurrently with the performance of the requested operation, the golden input data is combined with the input data for the requested operation and run through the model. Relevant outputs are compared with the golden output data to validate the stability of the model and weights.

Resiliency for machine learning workloads

A crossbar array includes a number of memory elements. An analog-to-digital converter (ADC) is electronically coupled to the vector output register. A digital-to-analog converter (DAC) is electronically coupled to the vector input register. A processor is electronically coupled to the ADC and to the DAC. The processor may be configured to determine whether division of input vector data by output vector data from the crossbar array is within a threshold value, and if not within the threshold value, determine changed data values as between the output vector data and the input vector data, and write the changed data values to the memory elements of the crossbar array.

Acceleration of model/weight programming in memristor crossbar arrays

Multi-metric artificial intelligence (AI)/machine learning (ML) models for detection of anomalous behavior of a machine/system are disclosed. The multi-metric AI/ML models are configured to detect anomalous behavior of systems having multiple sensors that measure correlated sensor metrics such as coolant distribution units (CDUs). The multi-metric AI/ML models perform the anomalous system behavior detection in a manner that enables both a reduction in the amount of sensor instrumentation needed to monitor the system's operational behavior as well as a corresponding reduction in the complexity of the firmware that controls the sensor instrumentation. As such, AI-enabled systems and corresponding methods for anomalous behavior detection disclosed herein offer a technical solution to the technical problem of increased failure rates of existing multi-sensor systems, which is caused by the presence of redundant sensor instrumentation that necessitates complex firmware for controlling the sensor instrumentation.

Anomalous behavior detection by an artificial intelligence-enabled system with multiple correlated sensors

One embodiment provides a system and method for predicting network power usage associated with workloads. During operation, the system configures a simulator to simulate operations of a plurality of network components, which comprises embedding one or more event counters in each simulated network component. A respective event counter is configured to count a number of network-power-related events. The system collects, based on values of the event counters, network-power-related performance data associated with one or more sample workloads applied to the simulator; and trains a machine-learning model with the collected network-power-related performance data and characteristics of the sample workloads as training data 1, thereby facilitating prediction of network-power-related performance associated with a to-be-evaluated workload.

System and method for predicting power usage of network components

Systems and methods are provided for implementing hardware optimization for a hardware accelerator. The hardware accelerator emulates a neural network. Training of the neural network integrates a regularized pruning technique to systematically reduce a number of weights. A crossbar array included in hardware accelerator can be programmed to calculate node values of the pruned neural network to selectively reduce the number of weight column lines in the crossbar array. During deployment, the hardware accelerator can be programmed to power off periphery circuit elements that correspond to a pruned weight column line to optimize the hardware accelerator for power. Alternatively, before deployment, the hardware accelerator can be optimized for area by including a finite number of weight column line. Then, regularized pruning of the neural network selectively reduces the number of weights for consistency with the finite number of weight columns lines in the hardware accelerator.

Systems and methods for neural network training and deployment for hardware accelerators

Disclosed techniques provide for dynamically changing precision of a multi-stage compute process. For example, changing neural network (NN) parameters on a per-layer basis depending on properties of incoming data streams and per-layer performance of an NN among other considerations. NNs include multiple layers that may each be calculated with a different degree of accuracy and therefore, compute resource overhead (e.g., memory, processor resources, etc.). NNs are usually trained with 32-bit or 16-bit floating-point numbers. Once trained, an NN may be deployed in production. One approach to reduce compute overhead is to reduce parameter precision of NNs to 16 or 8 for deployment. The conversion to an acceptable lower precision is usually determined manually before deployment and precision levels are fixed while deployed. Disclosed techniques and implementations address automatic rather than manual determination or precision levels for different stages and dynamically adjusting precision for each stage at run-time.

Adjustable precision for multi-stage compute processes

A computer system includes multiple memory array components that include respective analog memory arrays which are sequenced to implement a multi-layer process. An error array data structure is obtained for at least a first memory array component, and from which a determination is made as to whether individual nodes (or cells) of the error array data structure are significant. A determination can be made as to any remedial operations that can be performed to mitigate errors of significance.

Determining significance levels of error values in processes that include multiple layers

Systems and methods are provided for optimizing parameters of a system across an entire stack, including algorithms layer, toolchain layer, execution or runtime layer, and hardware layer. Results from the layer-specific optimization functions of each domain can be consolidated using one or more consolidation optimization functions to consolidate the layer-specific optimization results, capturing the relationship between the different layers of the stack. Continuous monitoring of the programming model during execution may be implemented and can enable the programming model to self-adjust based on real-time performance metrics. In this way, programmers and system administrators are relieved of the need for domain knowledge and are offered a systematic way for continuous optimization (rather than an ad hoc approach).

Self-adjustable end-to-end stack programming

A dot-product engine (DPE) implemented on an integrated circuit as a crossbar array (CA) includes memory elements comprising a memristor and a transistor in series. A crossbar with N rows, M columns may have N&#xd7;M memory elements. A vector input for N voltage inputs to the CA and a vector output for M voltage outputs from the CA. An analog-to-digital converter (ADC) and/or a digital-to-analog converter (DAC) may be coupled to each input/output register. Values representing a first matrix may be stored in the CA. Voltages/currents representing a second matrix may be applied to the crossbar. Ohm's Law and Kirchoff's Law may be used to determine values representing the dot-product as read from the crossbar. A portion of the crossbar may perform Error-correcting Codes (ECC) concurrently with calculating the dot-product results. ECC codes may be used to only indicate detection of errors, or for both detection and correction of results.

Growing community of inventors

Palo Alto, CA, United States of America

Sergey Serebryakov