Nec Corporation

Asapp, Inc.

Felix Wu

Eric Cosatto

Alexandru Niculescu-Mizil

Kyu Jeong Han

Kwangyoun Kim

Prashant Sridhar

Kilian Quirin Weinberger

Jing Pan

Yoav Artzi

For any application that processes speech, improving the quality of the feature vectors may improve the quality of the speech application. The quality of feature vectors may be improved by modifying a neural network architecture for computing feature vectors to allocate computational resources where they are more effective for learning and computing the feature vectors. Contextual feature vectors may be computed from feature vectors by using a parameterized downsampling operation that decreases a vector sequence rate, processing the downsampled vectors with a neural network, and using a parameterized upsampling operation that increases a vector sequence rate. For example, parameterized downsampling may decrease a vector sequence rate by a factor of two, a neural may require fewer computational resources since it operates with a lower vector sequence rate, and parameterized upsampling may then increase the vector sequence rate by a factor of two.

Contextual feature vectors for processing speech

The amount of future context used in a speech processing application allows for tradeoffs between performance and the delay in providing results to users. Existing speech processing applications may be trained with a specified future context size and perform poorly when used in production with a different future context size. A speech processing application trained using a stochastic future context allows a trained neural network to be used in production with different amounts of future context. During an update step in training, a future-context size may be sampled from a probability distribution, used to mask a neural network, and compute an output of the masked neural network. The output may then be used to compute a loss value and update parameters of the neural network. The trained neural network may then be used in production with different amounts of future context to provide greater flexibility for production speech processing applications.

Stochastic future context for speech processing

A method is provided for model training to detect defective products. The method includes sampling training images of a product to (i) extract image portions therefrom made of a center patch and its context and (ii) black-out the center patch. The method further includes performing unsupervised back-propagation training of a Contextual Auto-Encoder (CAE) model using (i) the image portions with the blacked-out center patch as an input and, (ii) the center patch as a target output and, (iii) an image-based loss function, to obtain a trained CAE model. The method also includes sampling positive and negative center-patch-sized portions from the training images. The method additionally includes normalizing, using the trained CAE model, the positive and negative center-patch-sized portions. The method further includes performing supervised training of a classifier model using the normalized positive and negative center-patch-sized portions to obtain a trained supervised classifier model for detecting the defective products.

Anomaly detection with predictive normalization

Systems and methods for detecting and correcting defective products include capturing at least one image of a product with at least one image sensor to generate an original image of the product. An encoder encodes portions of an image extracted from the original image to generate feature space vectors. A decoder decodes the feature space vectors to reconstruct the portions of the image into reconstructed portions by predicting defect-free structural features in each of the portions according to hidden layers trained to predict defect-free products. Each of the reconstructed portions are merged into a reconstructed image of a defect-free representation of the product. The reconstructed image is communicated to a contrastor to detect anomalies indicating defects in the product.

Reconstructor and contrastor for anomaly detection

Systems and methods for diagnosing a patient condition include a medical imaging device for generating an anatomical image. A reconstructor reconstructs the anatomical image by reconstructing portions of the anatomical image to be a healthy representation of the portions and merging the portions into the anatomical image to generate a reconstructed image. A contrastor contrasts the anatomical image with the reconstructed image to generate an anomaly map indicating locations of difference between the anatomical image and the reconstructed image. An anomaly tagging device tags the locations of difference as anomalies corresponding to anatomical abnormalities in the anatomical image, and a display displays the anatomical image with tags corresponding to the anatomical abnormalities.

Growing community of inventors

Ithaca, NY, United States of America

Felix Wu