Abstract: In one aspect of the invention, there is a computer-implemented method including: detecting, by a processor set of a first sensor agent, sensor data from one or more sensors comprised in the first sensor agent; determining, by the processor set, an own series of estimates, based on the sensor data; transmitting, by the processor set, the own series of estimates; receiving, by the processor set, at least one additional series of estimates from additional sensor agents; restoring, by the processor set, in response to detecting that a second sensor agent of the additional sensor agents has become disconnected and then re-connected, the transmitting of the series of estimates to the second sensor agent and the receiving of the series of estimates from the second sensor agent; and outputting, by the processor set, based on the own series of estimates and the additional series of estimates, a series of consensus estimates.

Abstract: One or more group-specific aggregate losses, one or more group-agnostic aggregate losses, and a joint loss are computed. A regularizer loss is computed based on the one or more group-specific aggregate losses and the one or more group-agnostic aggregate losses. One or more group-specific models are trained based on the one or more group-specific aggregate losses. A feature extractor is updated based on the regularizer loss and a joint classifier is updated based on the joint loss.

Abstract: A computer implemented method for certifying robustness of image classification in a neural network is provided. The method includes initializing a neural network model. The neural network model includes a problem space and a decision boundary. A processor receives a data set of images, image labels, and a perturbation schedule. Images are drawn from the data set in the problem space. A distance from the decision boundary is determined for the images in the problem space. A re-weighting value is applied to the images. A modified perturbation magnitude is applied to the images. A total loss function for the images in the problem space is determined using the re-weighting value. A confidence level of the classification of the images in the data set is evaluated for certifiable robustness.

Abstract: Embodiments of the present invention provide computer-implemented methods, computer program products and computer systems. Embodiments of the present invention can identify a plurality of constraints on states of data and actions of data associated with a data model. Embodiments of the present invention can then identify constraints on safety policy parameters associated with a computing device. Embodiments of the present invention can then convert the identified constraints into a uniform domain syntax that considers coupled and decoupled constraints and introduce buffer data within the converted constraints, wherein the buffer data filters outlier constraints within the plurality of constraints. Embodiments of the present invention can then dynamically generate optimal safety policies associated with the computing device based on the remaining constraints.

Abstract: A pretrained machine learning model is obtained. A Bayesian meta-model is configured to cooperate with the pretrained machine learning model, the Bayesian meta-model being configured to quantify different kinds of uncertainties associated with the pretrained machine learning model, wherein the Bayesian meta-model comprises a plurality of linear layers attached to different intermediate features of the pretrained machine learning model with a final linear layer generating a Dirichlet distribution. Multiple intermediate features extracted from the pretrained machine learning model are received as inputs. A Dirichlet distribution is generated over a probability simplex as output, wherein the Dirichlet distribution is parameterized by the Bayesian meta-model and allows quantification of uncertainty of model prediction, and the Bayesian meta-model and the pretrained machine learning model are used in a downstream task.

Abstract: Prompt embedding samples are drawn from a prior distribution and are passed into a pretrained model to receive a corresponding token label prediction for a batch of text data. Prompt embedding samples are accepted from a distribution of a first iteration; the accepted samples satisfy a condition of a distance function between a ground truth label and the corresponding token label prediction being less than a first tolerance. Embeddings are resampled from the accepted prompt embedding samples with probability proportional to weights and the resampled embeddings are perturbed via a perturbation kernel to obtain a new sample. The perturbed resampled embeddings are propagated through the pretrained model, and those that satisfy a condition are projected, where the second tolerance is decayed by one step per iteration. The projected resampled embeddings are concatenated with an embedding of a given input and inferencing is performed.

Abstract: A computer-implemented method and system for generating a predictive model include a computation engine learning a predictive model having missing or crippled data. A processor applies a formulated problem of missing or crippled data based learning to the predictive model. The computation engine reduces one or more tasks associated with the predictive model to a quadratically constrained quadratic problem (QCQP). The computation engine characterizes one or more solutions associated with the QCQP, where each of the one or more solutions is a loss function value.

Abstract: Obtain, as input, in electronic form, structured information including tasks for a plurality of occupations in a plurality of industries over a length of time; compute, from the structured information, a time series of normalized occupation task shares over the length of time; train a computerized machine learning model, on the time series, to predict future task shares for the plurality of occupations in the plurality of industries; and, with the trained computerized machine learning model, predict the future task shares.

Abstract: Obtain, using at least one hardware processor, data characterizing a physical system governed by a physical conservation law. Apply, using the at least one hardware processor, contrastive learning to the data to automatically capture system invariants of the physical system. Employ, using the at least one hardware processor, a neural projection layer to guarantee that a corresponding dynamic machine learning model preserves the captured system invariants. Optionally, predict performance of the physical system using the corresponding dynamic machine learning model.

Abstract: A computer implemented method determines skill shares for skills from job advertisements. A skill share for a skill identifies a number of times a skill has appeared in job advertisements during a given period of time. The computer implemented method creates a time series of skill demand using the skill shares. The computer implemented method extracts embeddings from job advertisements for an occupation using natural language processing. The computer implemented method clusters the skills using the embeddings to form skill clusters. The computer implemented method defines a training dataset using as a time series of skill demand for all skills within a cluster containing a selected skill to be predicted. The computer implemented method trains a time series prediction model using the training dataset.

Abstract: In one aspect of the invention, there is a computer-implemented method including: detecting, by a processor set of a first sensor agent, sensor data from one or more sensors comprised in the first sensor agent; determining, by the processor set, an own series of estimates, based on the sensor data; transmitting, by the processor set, the own series of estimates; receiving, by the processor set, at least one additional series of estimates from additional sensor agents; restoring, by the processor set, in response to detecting that a second sensor agent of the additional sensor agents has become disconnected and then re-connected, the transmitting of the series of estimates to the second sensor agent and the receiving of the series of estimates from the second sensor agent; and outputting, by the processor set, based on the own series of estimates and the additional series of estimates, a series of consensus estimates.

Abstract: One or more computer processors create a fully convolution network (FCN) comprising a plurality of 1×1 convolutions. The one or more computer processors append linear mapping layer (LM) to created FCN. The one or more computer processors capture a plurality of features utilizing multi-scale dilated convolutional kernels from the linear mapped FCN (LM-FCN). The one or more computer processors apply an average pool layer to the captured plurality of features along a temporal axis of a dilated convolutional kernel within the LM-FCN. The one or more computer processors predict a classification for subsequent time-series data utilizing the pooled plurality of features.

Abstract: Providing a trained reinforcement learning (RL) model by formulating a decision process problem for the RL model, defining at least one of a logarithmic loss function for the RL model and defining an initiation point for the RL model according to an optimized spectral norm of the RL model, training the system according to the logarithmic loss function or from the initiation point, and providing the trained RL model.

Abstract: Using a first sensor device in a network of sensor devices, sensor data is measured. A second sensor device comprising a trusted sensor device is selected from the network of sensor devices. A gain matrix is updated. Using the gain matrix, the sensor data, and a second parameter estimate received from the second sensor device, a first parameter estimate is updated. The first parameter estimate comprises an estimate of a parameter of a model representing the first sensor device. Using the gain matrix, an estimation error covariance matrix and a cross-variance matrix are updated. Using the updated first parameter estimate, second sensor data measured by the first sensor device is adjusted.

Abstract: In aspects of the disclosure, a method comprises training, by a computing system, a dynamics model of a cooperative multi-agent reinforcement learning (c-MARL) environment. The method further comprises processing, by the computing system, a perturbation optimizer to generate a state perturbation of the c-MARL environment, based on the dynamics model. The method further comprises selecting one or more agents of the c-MARL system as having enhanced vulnerability. The method further comprises attacking, by the computing system, the c-MARL system based on the state perturbation and the selected one or more agents.

Abstract: A computer implemented method for certifying robustness of image classification in a neural network is provided. The method includes initializing a neural network model. The neural network model includes a problem space and a decision boundary. A processor receives a data set of images, image labels, and a perturbation schedule. Images are drawn from the data set in the problem space. A distance from the decision boundary is determined for the images in the problem space. A re-weighting value is applied to the images. A modified perturbation magnitude is applied to the images. A total loss function for the images in the problem space is determined using the re-weighting value. A confidence level of the classification of the images in the data set is evaluated for certifiable robustness.

Abstract: An embodiment for identifying skill adjacencies and skill gaps to generate reskilling recommendations. The embodiment may receive input from a user including candidate details and a job description. The embodiment may automatically extract a first set of skill keywords from the candidate description and a second set of skill keywords from the job description. The embodiment may automatically input the first and second set of skill keywords into a first type of word embedding model and a second type of word embedding model to automatically generate word embeddings. The embodiment may automatically compare the generated word embeddings and calculate cosine similarity scores for the first and second set of skill keywords.

Abstract: One or more group-specific aggregate losses, one or more group-agnostic aggregate losses, and a joint loss are computed. A regularizer loss is computed based on the one or more group-specific aggregate losses and the one or more group-agnostic aggregate losses. One or more group-specific models are trained based on the one or more group-specific aggregate losses. A feature extractor is updated based on the regularizer loss and a joint classifier is updated based on the joint loss.

Abstract: Embodiments of the present invention provide computer-implemented methods, computer program products and computer systems. Embodiments of the present invention can identify a plurality of constraints on states of data and actions of data associated with a data model. Embodiments of the present invention can then identify constraints on safety policy parameters associated with a computing device. Embodiments of the present invention can then convert the identified constraints into a uniform domain syntax that considers coupled and decoupled constraints and introduce buffer data within the converted constraints, wherein the buffer data filters outlier constraints within the plurality of constraints. Embodiments of the present invention can then dynamically generate optimal safety policies associated with the computing device based on the remaining constraints.

Abstract: A method for creating a question answering system includes receiving user stories, wherein each of the user stories is structured as a plurality of first phrasal entities within a template; applying a Natural Language Processing to discover first data relationships between the first phrasal entities and first context relationships between the first phrasal entities; constructing a knowledge graph that captures second data relationships and second contextual relationships of a plurality of second phrasal entities; enriching the KG by linking the first phrasal entities to the second phrasal entities to form enriched phrasal entities in the KG; receiving a selection of ones of the enriched phrasal entities for completing a story template; identifying a technical requirement based on the selection of the enriched phrasal entities; and training a model matching at least one of the user stories to the technical requirement.