Ricoh Company, Ltd.

Harman International Industries, Inc.

Aashish Kumar

Manuel Martinello

Shachar Mendelowitz

Rajesh Biswal

Alexander Kreines

Nir Morgulis

Srinivas Kruthiventi Subrahmanyeswara Sai

Jigar Mistry

Sambuddha Saha

Srinivas Kruthiveti Subrahmanyeswara Sai

George Jose

The current disclosure relates to systems and methods for wakeword or keyword detection in Virtual Personal Assistants (VPAs). In particular, systems and methods are provided for wakeword detection using deep neural networks including a parametric pooling layer, wherein the parametric pooling layer includes trainable parameters, enabling the layer to learn to distinguish between informative feature vectors and non-informative/noisy feature vectors extracted from a variable length acoustic signal. In one example, a parametric pooling layer may aggregate a variable length feature map, comprising a plurality of feature vectors extracted from an acoustic signal, into an embedding vector of pre-determined length, by weighting each of the plurality of feature vectors based on one or more learned parameters in a parametric pooling layer, and aggregating the plurality of weighted feature vectors into the embedding vector.

System and method for robust wakeword detection in presence of noise in new unseen environments without additional data

Techniques are disclosed for detecting adversarial attacks. A machine learning (ML) system processes the input into and output of a ML model using an adversarial detection module that does not include a direct external interface. The adversarial detection module includes a detection model that generates a score indicative of whether the input is adversarial using, e.g., a neural fingerprinting technique or a comparison of features extracted by a surrogate ML model to an expected feature distribution for the output of the ML model. In turn, the adversarial score is compared to a predefined threshold for raising an adversarial flag. Appropriate remedial measures, such as notifying a user, may be taken when the adversarial score satisfies the threshold and raises the adversarial flag.

Defending machine learning systems from adversarial attacks

In one aspect, a color plenoptic imaging system captures a plenoptic image of an object. The plenoptic image is made up of a plurality of superpixels, and each superpixel includes a center subpixel. The collection of center subpixels from the plurality of superpixels forms a set of captured center view data for the object. The sensor array includes at least two arrays of different color sensors that capture subpixels of different colors. The microlens array and the sensor array are positioned such that, within the set of captured center view data, for each of the different colors, adjacent center subpixels of that color are separated by not more than three intervening center subpixels of a different color.

Reducing color artifacts in plenoptic imaging systems

In one aspect, a plenoptic imaging system includes imaging optics, a microlens array and a sensor array. A method for color correction of views captured by the plenoptic imaging system includes the following. A plenoptic image of an object captured by the plenoptic imaging system is accessed. The plenoptic image includes many superpixels. Each superpixel includes many subpixels corresponding to different views of the object. The subpixels are of at least two different colors. An initial color view of the object is generated from the plenoptic image. The initial color view includes color pixels. Pixel-wise color correction factors are applied to the color pixels in the initial color view, thereby generating a corrected color view. The pixel-wise color correction factors are determined based on an initial color view generated for a calibration object of known color, as captured by the plenoptic imaging system.

Color correction of preview images for plenoptic imaging systems

A color preview image is generated, preferably in real-time, for a plenoptic imaging system. The raw plenoptic image has a structure containing superpixels and also contains different color channels. For a principal color channel, a center view is generated by determining the pixel value for a center view of the superpixel, i.e. at the centroid of the superpixel. For each of the other color channels, the ratio of that color channel to the principal color channel is calculated for each superpixel. A center view for each non-principal color channel is determined by multiplying the color ratios times the pixel values for the principal color center view. These center views for the principal and non-principal color channels are combined into a color preview image. The calculations preferably can be performed in real-time at video rates.

Growing community of inventors

Bengaluru, India

Aashish Kumar