Abstract: Deep learning in situ retraining uses deep learning nodes to provide a human perception state on a user device. A plurality of images including facial data is obtained for human perception state analysis. A server device trains a set of weights on a set of layers for deep learning that implements the analysis, where the training is performed with a first set of training data. A subset of weights is deployed on deep learning nodes on a user device, where the deploying enables at least part of the human perception state analysis. An additional set of weights is retrained on the user device, where the additional set of weights is trained using a second set of training data. A human perception state based on the subset of the set of weights, the additional set of weights, and input images obtained by the user device is provided on the user device.

Abstract: Methods and systems disclosed herein relate generally to systems and methods for generating visual relationship graphs that identify relationships between objects depicted in an image. A vision-language application uses transformer encoders to generate a graph structure, in which the graph structure represents a dependency between a first region and a second region of an image. The dependency indicates that a contextual representation of the first region was derived, at least in part, by processing the second region. The contextual representation identifies a predicted identity of an image object depicted in the first region. The predicted identity is determined at least in part by identifying a relationship between the first region and other data objects associated with various modalities.

Abstract: A system and method are described herein for training and evaluating machine learning models trained using large complex datasets in an efficient manner. The systems enable methods for training a first model and generating predictions using the first trained model. The predictions and a portion of the data may be used to train a second model and generate predictions as an output of the second model. The first model may be used on a third dataset to generate predictions, which are appended to the third dataset. The appended third dataset is then input into the second model for generation of additional outputs that are compared against the third dataset for accuracy.

Abstract: The present invention in a first embodiment is a method of constructing a geometry-induced sparse local connected network architecture, the method comprising: choosing a geometry, choosing a direction of data flow in the geometry, choosing a node set as a finite subset of the geometry, choosing local edges between each node and nodes in preceding layers with respect to the geometry and direction of data flow, and choosing sparse nonlocal edges between each node and nodes in preceding layers with respect to the geometry and direction of data flow.

Abstract: A method to generate synthetic data instances. The method includes generating a synthetic data instance for an input variable value of an input variable supplied to the generative model, classifying the synthetic data instance to generate a classification result, determining a loss function value of a loss function, the loss function evaluating the classification result and determining the gradient of the loss function with respect to the input variable. Depending on the absolute value of the gradient, the method includes generating a plurality of modified input variable values, determining, for each modified input variable value, the gradient of the loss function, combining the gradients of the loss function to generate an estimated gradient, and modifying the input variable value in a direction determined by the estimated gradient to generate a further input variable value. The generative model generates a further synthetic data instance for the further input variable value.

Abstract: A real-time temporal convolution network (RT-TCN) algorithm reuses the output of prior convolution operations in all layers of the network to minimize the computational requirements and memory footprint of a TCN during real-time evaluation. Further, a TCN trained via the fixed-window view, where the TCN is trained using fixed time splices of the input time series, can be executed in real-time continually using RT-TCN.

Abstract: Techniques are disclosed relating to dynamically determining a data load process usable with a decision service for responding to requests to the decision service. The data load process may be selected from either a full data load process that loads data needed for all the rules in the decision service, or a selective data load process that loads data needed for a subset of the rules in the decision service according to embodiments. The decision service may determine the data load process to implement based on an evaluation of a decision blueprint in response to receiving the request, where the decision blueprint includes a set of rules interconnected by flow paths. Evaluation of properties of the rules and flow paths based on input data in the request may be used to determine the data load process to implement. Overall, this results in faster execution times in various circumstances.

Abstract: A method of interpreting tabular data includes receiving, at a deep tabular data learning network (TabNet) executing on data processing hardware, a set of features. For each of multiple sequential processing steps, the method also includes: selecting, using a sparse mask of the TabNet, a subset of relevant features of the set of features; processing using a feature transformer of the TabNet, the subset of relevant features to generate a decision step output and information for a next processing step in the multiple sequential processing steps; and providing the information to the next processing step. The method also includes determining a final decision output by aggregating the decision step outputs generated for the multiple sequential processing steps.

Abstract: In an example embodiment, a single machine learned model that allows for ranking of entities across all of the different combinations of node types and edge types is provided. The solution calibrates the scores from Edge-FPR models to a single scale. Additionally, the solution may utilize a per-edge type multiplicative factor dictated by the true importance of an edge type, which is learned through a counterfactual experimentation process. The solution may additionally optimize on a single, common downstream metric, specifically downstream interactions that can be compared against each other across all combinations of node types and edge types.