Sas Institute Inc.

North Carolina State University

James Allen Cox

Samuel Paul Leeman-Munk

Zheng Zhao

Saratendu Sethi

Jordan Riley Benson

Ravinder Devarajan

David James Caira

Christopher Graham Healey

Kalpesh Padia

Lawrence Edmund Lewis

Shaoliang Nie

Russell Albright

David B Styles

Mustafa Kabul

Jared Peterson

Arila Barnes

Samantha DuPont

Russel Albright

Barry DeVille

Nancy Anne Rausch

Jun Liu

Richard Welland Crowell

Barry De Ville

David Rawlings Duling

Bradford Wayne Mott

Warren Sarle

Oliver Dain

Ning Jin

Reza Soleimani

Gowtham Dinakaran

Wookhee Min

James Curtis Lester, Ii

In one example, a system can execute a first machine-learning model to determine an overall classification for a textual dataset. The system can also determine classification scores indicating the level of influence that each token in the textual dataset had on the overall classification. The system can select a first subset of the tokens based on their classification scores. The system can also execute a second machine-learning model to determine probabilities that the textual dataset falls into various categories. The system can determine category scores indicating the level of influence that each token had on a most-likely category determination. The system can select a second subset of the tokens based on their category scores. The system can then generate a first visualization depicting the first subset of tokens color-coded to indicate their classification scores and a second visualization depicting the second subset of tokens color-coded to indicate their category scores.

Graphical user interface for visualizing contributing factors to a machine-learning model's output

Computerized pipelines can transform input data into data structures compatible with models in some examples. In one such example, a system can obtain a first table that includes first data referencing a set of subjects. The system can then execute a sequence of processing operations on the first data in a particular order defined by a data-processing pipeline to modify an analysis table to include features associated with the set of subjects. Executing each respective processing operation in the sequence to generate the modified analysis table may involve: deriving a respective set of features from the first data by executing a respective feature-extraction operation on the first data; and adding the respective set of features to the analysis table. The system may then execute a predictive model on the modified analysis table for generating a predicted value based on the modified analysis table.

Computerized pipelines for transforming input data into data structures compatible with models

A system trains a classification model. Text windows are defined from tokens based on a window size. A network model including a transformer network is trained with the text windows to define classification information. A first accuracy value is computed. (A) The window size is reduced using a predefined reduction factor value. (B) Second text windows are defined based on the reduced window size. (C) Retrain the network model with the second text windows to define classification information. (D) A second accuracy value is computed. (E) An accuracy reduction value is computed from the second accuracy value relative to the first accuracy value. When the computed accuracy reduction value is &#x2265;an accuracy reduction tolerance value, repeat (A)-(E) until the accuracy reduction value is &#x3c;the accuracy reduction tolerance value. Otherwise, increase the window size, define final text windows based on the increased window size, and retrain the network model.

Machine learning classification system

A computing system receives a collection comprising multiple sets of ordered terms, including a first set. The system generates a dataset indicating an association between each pair of terms within a same set of the collection by generating co-occurrence score(s) for the first set. The system generates computed probabilities based on the co-occurrence score(s) for the first set. The computed probabilities indicate a likelihood that one term in a given pair of terms of the collection appears in a given set of the collection given that another term in the given pair of terms of the collection occurs. The system smoothes the computed probabilities by adding one or more random observations. The system generates one or more association indications for the first set based on the smoothed computed probabilities. The system outputs an indication of the dataset. Additionally, or alternatively, based on association measure(s), the system generates a virtual term.

Word embeddings and virtual terms

Recurrent neural networks (RNNs) can be visualized. For example, a processor can receive vectors indicating values of nodes in a gate of a RNN. The values can result from processing data at the gate during a sequence of time steps. The processor can group the nodes into clusters by applying a clustering method to the values of the nodes. The processor can generate a first graphical element visually indicating how the respective values of the nodes in a cluster changed during the sequence of time steps. The processor can also determine a reference value based on multiple values for multiple nodes in the cluster, and generate a second graphical element visually representing how the respective values of the nodes in the cluster each relate to the reference value. The processor can cause a display to output a graphical user interface having the first graphical element and the second graphical element.

Interactive visualizations for a recurrent neural network

Methods, processes and computer-program products are disclosed for use in a parallelized computing system in which representations of large sparse matrices are efficiently encoded and communicated between grid-computing devices. A sparse matrix can be encoded and stored as a collection of character strings wherein each character string is a Base64 encoded string representing the non-zero elements of a single row of the sparse matrix. On a per-row basis, non-zero elements can be identified by column indices and error correction metadata can be included. The resultant row data can be converted to IEEE 754 8-byte representations and then encoded into Base64 characters for storage as strings. These character strings of even very large-dimensional sparse matrices can be efficiently stored in databases or communicated to grid-computing devices.

Sparse matrix storage in a database

Convolutional neural networks can be visualized. For example, a graphical user interface (GUI) can include a matrix of symbols indicating feature-map values that represent a likelihood of a particular feature being present or absent in an input to a convolutional neural network. The GUI can also include a node-link diagram representing a feed forward neural network that forms part of the convolutional neural network. The node-link diagram can include a first row of symbols representing an input layer to the feed forward neural network, a second row of symbols representing a hidden layer of the feed forward neural network, and a third row of symbols representing an output layer of the feed forward neural network. Lines between the rows of symbols can represent connections between nodes in the input layer, the hidden layer, and the output layer of the feed forward neural network.

Visualizing convolutional neural networks

Interactive visualizations of a convolutional neural network are provided. For example, a graphical user interface (GUI) can include a matrix having symbols indicating feature-map values that represent likelihoods of particular features being present or absent at various locations in an input to a convolutional neural network. Each column in the matrix can have feature-map values generated by convolving the input to the convolutional neural network with a respective filter for identifying a particular feature in the input. The GUI can detect, via an input device, an interaction indicating that that the columns in the matrix are to be combined into a particular number of groups. Based on the interaction, the columns can be clustered into the particular number of groups using a clustering method. The matrix in the GUI can then be updated to visually represent each respective group of columns as a single column of symbols within the matrix.

Interactive visualizations of a convolutional neural network

Deep neural networks can be visualized. For example, first values for a first layer of nodes in a neural network, second values for a second layer of nodes in the neural network, and/or third values for connections between the first layer of nodes and the second layer of nodes can be received. A quilt graph can be output that includes (i) a first set of symbols having visual characteristics representative of the first values and representing the first layer of nodes along a first axis; (ii) a second set of symbols having visual characteristics representative of the second values and representing the second layer of nodes along a second axis; and/or (iii) a matrix of blocks between the first axis and the second axis having visual characteristics representative of the third values and representing the connections between the first layer of nodes and the second layer of nodes.

Visualizing deep neural networks

Training data for training a neural network usable for electronic sentiment analysis can be automatically constructed. For example, an electronic communication usable for training the neural network and including multiple characters can be received. A sentiment dictionary including multiple expressions mapped to multiple sentiment values representing different sentiments can be received. Each expression in the sentiment dictionary can be mapped to a corresponding sentiment value. An overall sentiment for the electronic communication can be determined using the sentiment dictionary. Training data usable for training the neural network can be automatically constructed based on the overall sentiment of the electronic communication. The neural network can be trained using the training data. A second electronic communication including an unknown sentiment can be received. At least one sentiment associated with the second electronic communication can be determined using the neural network.

Automatically constructing training sets for electronic sentiment analysis

Electronic communications can be normalized using a neural network. For example, a noncanonical communication that includes multiple terms can be received. The noncanonical communication can be preprocessed by (I) generating a vector including multiple characters from a term of the multiple terms; and (II) repeating a substring of the term in the vector such that a last character of the substring is positioned in a last position in the vector. The vector can be transmitted to a neural network configured to receive the vector and generate multiple probabilities based on the vector. A normalized version of the noncanonical communication can be determined using one or more of the multiple probabilities generated by the neural network. Whether the normalized version of the noncanonical communication should be outputted can also be determined using at least one of the multiple probabilities generated by the neural network.

Normalizing electronic communications using a vector having a repeating substring as input for a neural network

Systems and methods for performing analyses on data sets to display canonical rules sets with dimensional targets are disclosed. A cross-corpus rule set for a given Topic can be generated based on the entire corpus of data. A first dimensional rule set can be generated based on a first context (e.g., based on the same Topic but using a first sub-domain of the corpus of data). A second dimensional rule set can be generated based on a second context (e.g., based on the same Topic but using a second sub-domain of the corpus of data). Key dimensional differentiators (e.g., for each dimension, or context, of the Topic) can be determined based on a comparison of the general rule set, the first dimensional rule set, and the second dimensional rule set. A canonical rule set visualization can be displayed. The visualization can highlight the dimensional selectors (e.g., those tokens, or nodes, that differ between the first dimensional rule set and the second dimensional rule set).

Generating and displaying canonical rule sets with dimensional targets

Electronic communications can be normalized using neural networks. For example, an electronic representation of a noncanonical communication can be received. A normalized version of the noncanonical communication can be determined using a normalizer including a neural network. The neural network can receive a single vector at an input layer of the neural network and transform an output of a hidden layer of the neural network into multiple values that sum to a total value of one. Each value of the multiple values can be a number between zero and one and represent a probability of a particular character being in a particular position in the normalized version of the noncanonical communication. The neural network can determine the normalized version of the noncanonical communication based on the multiple values. Whether the normalized version should be output can be determined based on a result from a flagger including another neural network.

Normalizing electronic communications using a neural-network normalizer and a neural-network flagger

This disclosure describes a method, system and computer-program product for parallelized feature selection. The method, system and computer-program product may be used to access a first set of features, wherein the first set of features includes multiple features, wherein the features are characterized by a variance measure, and wherein accessing the first set of features includes using a computing system to access the features, determine components of a covariance matrix, the components of the covariance matrix indicating a covariance with respect to pairs of features in the first set, and select multiple features from the first set, wherein selecting is based on the determined components of the covariance matrix and an amount of the variance measure attributable to the selected multiple features, and wherein selecting the multiple features includes executing a greedy search performed using parallelized computation.

Growing community of inventors

Cary, NC, United States of America

James Allen Cox