International Business Machines Corporation

Guangnan Ye

Wei Zhang

Yada Zhu

Xiaodong Cui

Ching-Yung Lin

Jui-Hsin Lai

Ruichi Yu

Kumar Bhaskaran

Chun-Fu Chen

Keith C Houck

Toyotaro Suzumura

Hessian matrix-free sample-based techniques for model explanations that are faithful to the model are provided. In one aspect, a method for explaining a machine learning model {circumflex over (θ)} (e.g., for natural language processing) is provided. The method includes: training the machine learning model {circumflex over (θ)} with training data D; obtaining a decision of the machine learning model {circumflex over (θ)}; and explaining the decision of the machine learning model {circumflex over (θ)} using training examples from the training data D.

Faithful and efficient sample-based model explanations

A machine learning model can be trained to predict one or more financial indicators using earnings call transcripts augmented with counterfactual information. Using faithful gradient-based method, prediction results with respect to a particular counterfactual information can be explained. Based on the explanation, the counterfactual information determined to have most impact on prediction results can be selected for updating the machine learning model.

Counterfactual neural network learning for contextual enhanced earnings call analysis

A method for improving computing efficiency of a computing device for temporal directed cycle detection in a transaction graph includes preparing the transaction graph based on a plurality of transactions, the transaction graph including nodes indicating transaction origination points and transaction destination points, and edges indicating interactions between the nodes. Irrelevant nodes in the transaction graph are identified and pruned to provide a pruned, preprocessed transaction graph which can be partitioning into sections, where each section includes selected nodes that are linked to other linked nodes therein. Each of the sections having non-cyclic nodes can be trimmed prior to performing cycle detection on the resulting pruned transaction graph. Postprocessing pruning can be performed to further reduce the number of detected cycles that may be of interest to a particular application, such as in anti-money laundering.

Temporal directed cycle detection and pruning in transaction graphs

Method and apparatus for optimizing a convolutional neural network (CNN). A respective measure of importance is calculated for each of a plurality of elements within a CNN. A first one of the measures of importance is calculated by back propagating a second one of the measures of importance through the CNN. One or more of the plurality of elements is pruned from the CNN, based on the calculated measures of importance.

Pruning redundant neurons and kernels of deep convolutional neural networks

Growing community of inventors

Scarsdale, NY, United States of America

Guangnan Ye