Abstract: A method including receiving a user input for input to a language processing machine learning model (MLM). The method also includes generating modified inputs that are based on, and semantically related to, the user input. The method also includes executing the MLM to generate model outputs of the MLM. The MLM takes as input instances of each of the modified inputs. The method also includes sampling the model outputs using a statistical sampling strategy to generate sampled model outputs. The method also includes clustering the sampled model outputs into clusters. Each cluster of the clusters represents a distinct semantic meaning of the sampled model outputs. The method also includes generating a confidence metric for the user input. The confidence metric includes a predictive entropy of the clusters. The method also includes routing the user input based on whether the confidence metric satisfies or fails to satisfy a threshold value.

Abstract: Augmented Denoising Diffusion Implicit Models (“DDIMs”) using a latent trajectory optimization process can be used for image generation and manipulation using text input and one or more source images to create an output image. Noise bias and textual bias inherent in the model representing the image and text input is corrected by correcting trajectories previously determined by the model at each step of a diffusion inversion process by iterating multiple starts the trajectories to find determine augmented trajectories that minimizes loss at each step. The trajectories can be used to determine an augmented noise vector, enabling use of an augmented DDIM and resulting in more accurate, stable, and responsive text-based image manipulation.

Abstract: Aspects of the present disclosure provide techniques for machine learning and rules integration. Embodiments include receiving input values corresponding to a subset of a set of input variables associated with an automated determination. Embodiments include generating a directed acyclic graph (DAG) representing a set of constraints corresponding to the set of input variables. The set of constraints relate to one or more machine learning models and one or more rules. Embodiments include receiving one or more outputs from the one or more machine learning models based on one or more of the input values. Embodiments include determining outcomes for the one or more rules based on at least one of the input values. Embodiments include populating the DAG based on the input values, the one or more outputs, and the outcomes. Embodiments include making the automated determination based on logic represented by the DAG.

Abstract: Certain aspects of the present disclosure provide techniques for processing natural language utterances in a knowledge graph. An example method generally includes receiving a long-tail query comprising a natural language utterance from a user of an application. Operands and operators are extracted from the natural language utterance using a natural language model. Operands may be mapped to nodes in a knowledge graph, the nodes representing values calculated from data input into the application, and operators may be mapped to operations to be performed on data extracted from the knowledge graph. The functions associated with the operators are executed using data extracted from the nodes in the knowledge graph associated with the operands to generate a query result. The query result is returned as a response to the received long-tail query.

Abstract: A method includes executing a Optical Character Recognition (OCR) preprocessor on training images to obtain OCR preprocessor output, executing an OCR engine on the OCR preprocessor output to obtain OCR engine output, and executing an approximator on the OCR preprocessor output to obtain approximator output. The method further includes iteratively adjusting the approximator to simulate the OCR engine using the OCR engine output and the approximator output, and generating OCR preprocessor losses using the approximator output and target labels. The method further includes iteratively adjusting the OCR preprocessor using the OCR preprocessor losses to obtain a customized OCR preprocessor.

Abstract: Systems and methods for forecasting cashflows across one or more accounts of a user disclosed. One example method may include retrieving a data set for each of a plurality of accounts from a database, constructing a graph including a plurality of nodes linked together by a multitude of edges, wherein each node identifies a time series value corresponding to one of the accounts, and each edge indicates a time series value of a corresponding set of transactions occurring between a corresponding pair of accounts, determining a plurality of constraints, determining a specified loss function based on the plurality of constraints, back-propagating a derivative of the specified loss function into a deep neural network (DNN) to determine a set of neural network parameters, forecasting, using the DNN, a time sequence for one or more of the nodes and one or more of the edges, and providing the forecasted time sequences to the user.

Abstract: In one embodiment a finite rank deep kernel learning method includes: receiving a training dataset; forming a plurality of training data subsets from the training dataset; for each respective training data subset of the plurality of training data subsets: calculating a subset-specific loss based on a loss function and the respective training data subset; and optimizing a model based on the subset-specific loss; determining a set of embeddings based on the optimized model; determining, based on the set of embeddings, a plurality of dot kernels; combining the plurality of dot kernels to form a composite kernel for a Gaussian process; receiving live data from an application; and predicting a plurality of values and a plurality of uncertainties associated with the plurality of values simultaneously using the composite kernel.

Abstract: Systems and methods are disclosed for model based document image enhancement. Instead of requiring paired dirty and clean images for training a model to clean document images (which may cause privacy concerns), two models are trained on the unpaired images such that only the dirty images are accessed or only the clean images are accessed at one time. One model is a first implicit model to translate the dirty images from a source space to a latent space, and the other model is a second implicit model to translate the images from the latent space to clean images in a target space. The second implicit model is trained based on translating electronic document images in the target space to the latent space. In some implementations, the implicit models are diffusion models, such as denoising diffusion implicit models based on solving ordinary differential equations.

Abstract: A method for generating a synthetic dataset involves generating discretized synthetic data based on driving a model of a cumulative distribution function (CDF) with random numbers. The CDF is based on a source dataset. The method further includes generating the synthetic dataset from the discretized synthetic data by selecting, for inclusion into the synthetic dataset, values from a multitude of entries of the source dataset, based on the discretized synthetic data, and providing the synthetic dataset to a downstream application that is configured to operate on the source dataset.

Abstract: A method of blank detection involves receiving a document from a user, where the document includes derived text; applying a trained blank detection model to the document to make a first prediction, where the first prediction indicates whether at least one field in the document is blank; comparing the first prediction with a second prediction, where the second prediction is made by an extraction model; and extracting the at least one field using the extraction model.

Abstract: One embodiment facilitates generating synthetic data objects using a semi-supervised GAN. During operation, a generator module synthesizes a data object derived from a noise vector and an attribute label. The system passes, to an unsupervised discriminator module, the data object and a set of training objects which are obtained from a training data set. The unsupervised discriminator module calculates: a value indicating a probability that the data object is real; and a latent feature representation of the data object. The system passes the latent feature representation and the attribute label to a supervised discriminator module. The supervised discriminator module calculates a value indicating a probability that the attribute label given the data object is real. The system performs the aforementioned steps iteratively until the generator module produces data objects with a given attribute label which the unsupervised and supervised discriminator modules can no longer identify as fake.

Abstract: Aspects of the present disclosure provide techniques for intent matching. Embodiments include receiving input of text by a user via a user interface. Embodiments include determining weights for portions of the text based on a plurality of keywords. Embodiment include generating an embedding of the text. Embodiments include determining an intent of the text by weighting, based on the weights, word mover's distances from the embedding of the text to a known embedding of known text associated with the intent in order to determine a similarity measure between the text and the known text. Embodiments include providing content to the user via the user interface based on the intent.

Abstract: Aspects of the present disclosure provide techniques for confidence score calibration for automatic transaction categorization. Embodiments include providing one or more first inputs to a prediction model based on a transaction of a user. Embodiments include receiving a prediction of an account with a confidence score from the prediction model based on the one or more first inputs. Embodiments include providing one or more second inputs to a calibration model based on the confidence score, a detail type associated with the account, and a number of accounts of the user. Embodiments include receiving a calibrated confidence score from the calibration model based on the one or more second inputs. Embodiments include determining whether to automatically categorize the transaction into the account based on the calibrated confidence score.

Abstract: Aspects of the present disclosure provide techniques for machine learning and rules integration. Embodiments include receiving input values corresponding to a subset of a set of input variables associated with an automated determination. Embodiments include generating a directed acyclic graph (DAG) representing a set of constraints corresponding to the set of input variables. The set of constraints relate to one or more machine learning models and one or more rules. Embodiments include receiving one or more outputs from the one or more machine learning models based on one or more of the input values. Embodiments include determining outcomes for the one or more rules based on at least one of the input values. Embodiments include populating the DAG based on the input values, the one or more outputs, and the outcomes. Embodiments include making the automated determination based on logic represented by the DAG.

Abstract: In one embodiment a finite rank deep kernel learning method includes: receiving a training dataset; forming a plurality of training data subsets from the training dataset; for each respective training data subset of the plurality of training data subsets: calculating a subset-specific loss based on a loss function and the respective training data subset; and optimizing a model based on the subset-specific loss; determining a set of embeddings based on the optimized model; determining, based on the set of embeddings, a plurality of dot kernels; combining the plurality of dot kernels to form a composite kernel for a Gaussian process; receiving live data from an application; and predicting a plurality of values and a plurality of uncertainties associated with the plurality of values simultaneously using the composite kernel.

Abstract: Certain aspects of the present disclosure provide techniques for extracting data from a document. An example method generally includes identifying a bounding polygon of the region from an electronic image of the document and extracting data from within the bounding polygon of the region. The method further includes generating revised extracted data based on the extracted data, and combining the revised extracted data with other data extracted from the electronic image of the document to generate input data for a data processing application.

Abstract: Certain aspects of the present disclosure provide techniques for executing a function in a software application through a conversational user interface based on a knowledge graph associated with the function. An example method generally includes receiving a request to execute a function in a software application through a conversational user interface. A graph definition of the function is retrieved from a knowledge engine. Input is iteratively requested through the conversational user interface for each parameter of the parameters identified in the graph definition of the function based on a traversal of the graph definition of the function. Based on a completeness graph associated with the function, it is determined that the requested inputs corresponding to the parameters identified in the graph definition of the function have been provided through the conversational user interface. The function is executed using the requested inputs as parameters for executing the function.

Abstract: Embodiments described herein provide a system for generating semantically accurate synthetic images. During operation, the system generates a first synthetic image using a first artificial intelligence (AI) model and presents the first synthetic image in a user interface. The user interface allows a user to identify image units of the first synthetic image that are semantically irregular. The system then obtains semantic information for the semantically irregular image units from the user via the user interface and generates a second synthetic image using a second AI model based on the semantic information. The second synthetic image can be an improved image compared to the first synthetic image.

Abstract: A method including retrieving first data from first and second databases. The databases include different, incompatible formats and described different aspect of subjects. The data in the respective databases is referenceable using a common key type. The method also includes converting, into combined data, the first data and the second data into a canonical form configured for use as input to a machine learning model. The method also includes generating, using the common key type, pre-processed data by correlating, for ones of the subjects in the combined data, the first aspect of the subjects to the second aspect of the subjects. The machine learning model takes, as an input, the first aspect and the second aspect for each of the subjects in the pre-processed data, and generates, as an output, a prediction for a selected subject in the subjects. The method also includes presenting the output.

Abstract: Systems and methods are provided to determine consensus values for duplicate fields in a document or form.