Abstract: In various implementations, the techniques may include accessing a machine learning model, input and output ground-truth data. The techniques may include determining model deviation values between output data and the ground-truth data. The techniques may include determining a stochastic dependence value between the input data and the model deviation values using a loss function. The techniques may include training the model to reduce the stochastic dependence value below a predefined threshold value. The techniques may include determining a probability value indicating an existence of further deterministic relations to extract. If the stochastic dependence value is above the predefined threshold value, the techniques may continue training the model. If the stochastic dependence value is at or below the predefined threshold value or the probability value is greater than, less than, or equal to a predefined probability value the techniques can include storing the one or more weights the trained model.

Abstract: Described herein are techniques for analyzing historic time-series data to determine right-size commissioned resources from a software provider such as a hyperscaler. The historic time-series data may be analyzed to determine whether a commissioned resource should be upsized or downsized based on historical usage of the resource. Advantages to right-sizing a commissioned resource include improved performance and reduced spending.

Abstract: Embodiments of the present disclosure include techniques for predictive memory maintenance. In one embodiment, locations of correctable errors in a memory are observed. A machine learning (ML) system may be trained with patterns of correctable errors that result in uncorrectable errors. A trained ML monitors correctable errors to predict when memory requires maintenance. In another embodiment, error rates from multiple memories are monitored to predict memory channel and other upstream device failures.

Abstract: Described herein are techniques for intelligently pruning a machine learning or generative AI model. The model may first be split up into subunits. Each subunit may be analyzed to calculate a suitable measure such as a stochastic independence score or mutual information score. The subunits may in turn be ranked by their associated score and the lowest ranked subunit or subunits may be pruned from the model. The pruned model is then retrained, and accuracy of the pruned model is evaluated. A determination is then made whether to prune more or to return the pruned model.

Abstract: Described herein are techniques for stacking machine learning models to better capture deterministic relations in a dataset. In some instances, a first machine learning model may not be capable of capturing all of the deterministic relations in a dataset due to the limitations of the model. Supplemental models may be trained so that the corrections generated by the supplemental models, when combined with the first machine learning model, perform better at capturing the deterministic models in the dataset. Techniques are described for training supplemental models to capture deterministic relations associated with ordinal data and nominal data and continuous data.

Abstract: Described herein are techniques for determining whether a trained machine learning model has captured all of the deterministic relations in a dataset. In some examples, the techniques may be applied to the training dataset along with the validation or test dataset. First, the input variables from the dataset are fed into the trained machine learning model to generate predicted outputs. Second, the correctness of the predicted outputs is compared against the output variables from the dataset, also known as the ground truth. The correctness is represented by residuals. Third, the residuals and the input variables are correlated. If correlation exists, then the trained machine learning model has not captured all of the deterministic relations in the dataset.

Abstract: Described herein are techniques for automatically detecting root cause failures in a computing environment. A data center may experience a critical system failure that renders the data center inoperable. An administrator may utilize a root cause detector to analyze alerts generated from the data center to automatically detect the root cause of the critical system failure. Once detected, the administrator may investigate the device with the root cause failure to repair the data center. In some examples, the root cause detector may compare alerts received from the data center with root cause failures in a failure repository that it is familiar with to determine whether the present sequence of alerts is similar to alert patterns it has seen before. The root cause detector may be implemented with a classifier model, a large language model, a rule-based heuristics identifying spurious alert patterns, or combinations of these techniques.

Abstract: Systems and Methods for Automatically Detecting Root Cause Failures Described herein are techniques for automatically detecting root cause failures in a computing environment. A data center may experience a critical system failure that renders the data center inoperable. An administrator may utilize a root cause detector to analyze alerts generated from the data center to automatically detect the root cause of the critical system failure. Once detected, the administrator may investigate the device with the root cause failure to repair the data center. In some examples, the root cause detector may compare alerts received from the data center with root cause failures in a failure repository that it is familiar with to determine whether the present sequence of alerts is similar to alert patterns it has seen before. The root cause detector may be implemented with a classifier model, a large language model, a rule-based heuristics identifying spurious alert patterns, or combinations of these techniques.

Abstract: Root causes of network anomalies can be identified as follows. A subset of network entities that have experienced network anomalies during a time period are determined based on historical network data. A set of root cause candidates are selected among the plurality of network entities by iterating through the network topology, each root cause candidate being directly upstream of two or more network entities in the subset of network entities that have experienced network anomalies according to the network topology. Network entities that are root causes of the network anomalies are identified by removing root cause candidates that have a common upstream network entity that is also a root cause candidate from the set of root cause candidates leaving a set of remaining root cause candidates that are the root causes.

Abstract: To predict hardware safety margins, historic records of hardware metrics indicating amounts of allocated and used resources for one or more software applications are obtained. Feedback metrics indicating performance issues for the software are determined based on the metrics. Then a histogram is generated plotting a frequency of the feedback metric using bins based on a difference between the allocated resources and the used resources. A threshold value is determined for the difference by iteratively determining, starting with a rightmost bin, whether data points in that bin indicate poor performance of the software based on the difference between the allocated resources and the used resources. The threshold value indicates a safety margin for operating the one software applications without performing poorly. Resources for the one or more software applications are then re-allocated according to the safety margin.

Abstract: Embodiments of the present disclosure include techniques for predictive memory maintenance. In one embodiment, error locations in a RAM are specified by columns and rows. Error locations are detected and stored in a storage system. One or more plots of the error locations may be presented to a user. In some embodiments, the error locations are time stamped. Rules may be defined to automatically detect patterns of error locations statically or over time. Alerts may be generated automatically to perform maintenance of a computer system with failing memory.

Abstract: Embodiments of the present disclosure include techniques for predictive memory maintenance. In one embodiment, locations of correctable errors in a memory are observed. A machine learning (ML) system may be trained with patterns of correctable errors that result in uncorrectable errors. A trained ML monitors correctable errors to predict when memory requires maintenance. In another embodiment, error rates from multiple memories are monitored to predict memory channel and other upstream device failures.

Abstract: To predict network maintenance, historic records of hardware metrics are obtained for a plurality of network interfaces. An average of the metrics over a specified time span is determined for a plurality of time spans. Feedback metrics are determined for the network interfaces for each of the time spans. A histogram is generated that plots a frequency of the feedback metric for specified ranges of the hardware metric. A threshold value for the hardware metric is determined by iteratively determining whether a hardware metric bin of the histogram meets a specified non-zero value for the feedback metric starting from a highest hardware metric bin of the histogram. Then new records of hardware metrics are obtained and one or more network interfaces are determined to be needing maintenance based on an average of the hardware metrics in the new records meeting or exceeding the determined threshold value for the hardware metric.

Abstract: Root causes of network anomalies can be identified as follows. A subset of network entities that have experienced network anomalies during a time period are determined based on historical network data. A set of root cause candidates are selected among the plurality of network entities by iterating through the network topology, each root cause candidate being directly upstream of two or more network entities in the subset of network entities that have experienced network anomalies according to the network topology. Network entities that are root causes of the network anomalies are identified by removing root cause candidates that have a common upstream network entity that is also a root cause candidate from the set of root cause candidates leaving a set of remaining root cause candidates that are the root causes.

Abstract: Embodiments of the present disclosure include techniques for predictive memory maintenance. In one embodiment, locations of correctable errors in a memory are observed. A machine learning (ML) system may be trained with patterns of correctable errors that result in uncorrectable errors. A trained ML monitors correctable errors to predict when memory requires maintenance. In another embodiment, error rates from multiple memories are monitored to predict memory channel and other upstream device failures.

Abstract: Embodiments of the present disclosure include techniques for predictive memory maintenance. In one embodiment, locations of correctable errors in a memory are observed. A machine learning (ML) system may be trained with patterns of correctable errors that result in uncorrectable errors. A trained ML monitors correctable errors to predict when memory requires maintenance. In another embodiment, error rates from multiple memories are monitored to predict memory channel and other upstream device failures.

Abstract: Embodiments of the present disclosure include techniques for predictive memory maintenance. In one embodiment, error locations in a RAM are specified by columns and rows. Error locations are detected and stored in a storage system. One or more plots of the error locations may be presented to a user. In some embodiments, the error locations are time stamped. Rules may be defined to automatically detect patterns of error locations statically or over time. Alerts may be generated automatically to perform maintenance of a computer system with failing memory.