Abstract: Some embodiments are directed to a neural network training device for training a neural network. At least one layer of the neural network layers is a projection layer. The projection layer projects a layer input vector (x) of the projection layer to a layer output vector (y). The output vector (y) sums to the summing parameter (k).

Abstract: A mobile communication terminal device may include one or more image sensors, configured to generate image sensor data representing an environment of the mobile communication terminal device; one or more processors, configured to receive the image sensor data from the one or more image sensors; implement at least one artificial neural network to receive the image sensor data as an artificial neural network input and output an artificial neural network output representing a detected environment parameter of the environment of the mobile communication terminal; determine a navigation instruction based on the artificial neural network output; and send a signal representing the navigation instruction to a robot terminal via a communication interface.

Abstract: An example apparatus for imaging in low-light environments includes a raw sensor data receiver to receive raw sensor data from an imaging sensor. The apparatus also includes a convolutional neural network trained to generate an illuminated image based on the received raw sensor data. The convolutional neural network is trained based on images captured by a sensor similar to the imaging sensor.

Abstract: Techniques are provided for speech denoising using a denoising neural network (NN) trained with deep feature losses obtained from an audio classifier NN. A methodology implementing the techniques according to an embodiment includes applying the speech denoising NN, to be trained, to a noisy sample of a training speech signal to generate a processed training speech signal. The method further includes applying a trained audio classifier NN to the processed training speech signal to generate a first set of activation features, and applying the trained audio classifier NN to a clean sample of the training speech signal to generate a second set of activation features. The method further includes calculating a loss value based on the first and second sets of activation features, and performing a back-propagation training update of the denoising NN, based on the loss value. The method includes iterating this process to further train the denoising NN.

Abstract: To address the needs of applications that work with large-scale unstructured point clouds and other noisy data (e.g. image and video data), tangent convolution of 3D data represents 3D data as tangent planes. Tangent convolution estimates tangent planes for each 3D data point in one or more channels of 3D data. Tangent convolution further computes the tangent image signals for the estimated tangent planes. Tangent convolution precomputes the tangent planes and tangent image signals to enable convolution to be performed with greater efficiency and better performance than can be achieved with other 3D data representations.

Abstract: Techniques are provided for estimation of optical flow between images using 4-dimensional cost volume processing. A methodology implementing the techniques according to an embodiment includes extracting a first set of feature vectors from a first image and extracting a second set of feature vectors from a second image. Each feature vector of the first set is associated with a pixel of the first image and each feature vector of the second set is associated with a pixel of the second image. The method further includes constructing a 4-dimensional (4D) cost volume to store a distance metric between each feature vector of the first set of feature vectors and a selected subset of feature vectors of the second set of feature vectors. The method further includes performing a flow-semi-global matching (Flow-SGM) on the 4D cost volume to estimate an optical flow vector for pixels of the first image.

Abstract: Techniques are provided for approximating image processing functions using convolutional neural networks (CNNs). A methodology implementing the techniques according to an embodiment includes performing, by a CNN, a sequence of non-linear operations on an input image to generate an output image. The generated output image approximates the application of a targeted image processing operator to the input image. The CNN is trained on pairs of training input and output images, wherein the training output images are generated by application of the targeted image processing operator to the training input images. The CNN training process generates bias parameters and convolutional kernel parameters to be employed by the CNN for processing of intermediate image layers associated with processing stages between the input image and the output image, each of the processing stages associated with one of the sequence of non-linear operations. The parameters are associated with the targeted image processing operator.

Abstract: To address the needs of applications that work with large-scale unstructured point clouds and other noisy data (e.g. image and video data), tangent convolution of 3D data represents 3D data as tangent planes. Tangent convolution estimates tangent planes for each 3D data point in one or more channels of 3D data. Tangent convolution further computes the tangent image signals for the estimated tangent planes. Tangent convolution precomputes the tangent planes and tangent image signals to enable convolution to be performed with greater efficiency and better performance than can be achieved with other 3D data representations.

Abstract: Techniques are provided for speech denoising using a denoising neural network (NN) trained with deep feature losses obtained from an audio classifier NN. A methodology implementing the techniques according to an embodiment includes applying the speech denoising NN, to be trained, to a noisy sample of a training speech signal to generate a processed training speech signal. The method further includes applying a trained audio classifier NN to the processed training speech signal to generate a first set of activation features, and applying the trained audio classifier NN to a clean sample of the training speech signal to generate a second set of activation features. The method further includes calculating a loss value based on the first and second sets of activation features, and performing a back-propagation training update of the denoising NN, based on the loss value. The method includes iterating this process to further train the denoising NN.

Abstract: Systems, apparatuses and methods may provide for technology that generates, by a first neural network, an initial set of model weights based on input data and iteratively generates, by a second neural network, an updated set of model weights based on residual data associated with the initial set of model weights and the input data. Additionally, the technology may output a geometric model of the input data based on the updated set of model weights. In one example, the first neural network and the second neural network reduce the dependence of the geometric model on the number of data points in the input data.

Abstract: An example apparatus for imaging in low-light environments includes a raw sensor data receiver to receive raw sensor data from an imaging sensor. The apparatus also includes a convolutional neural network trained to generate an illuminated image based on the received raw sensor data. The convolutional neural network is trained based on images captured by a sensor similar to the imaging sensor.

Abstract: Techniques are provided for approximating image processing functions using convolutional neural networks (CNNs). A methodology implementing the techniques according to an embodiment includes performing, by a CNN, a sequence of non-linear operations on an input image to generate an output image. The generated output image approximates the application of a targeted image processing operator to the input image. The CNN is trained on pairs of training input and output images, wherein the training output images are generated by application of the targeted image processing operator to the training input images. The CNN training process generates bias parameters and convolutional kernel parameters to be employed by the CNN for processing of intermediate image layers associated with processing stages between the input image and the output image, each of the processing stages associated with one of the sequence of non-linear operations. The parameters are associated with the targeted image processing operator.

Abstract: Techniques are provided for estimation of optical flow between images using 4-dimensional cost volume processing. A methodology implementing the techniques according to an embodiment includes extracting a first set of feature vectors from a first image and extracting a second set of feature vectors from a second image. Each feature vector of the first set is associated with a pixel of the first image and each feature vector of the second set is associated with a pixel of the second image. The method further includes constructing a 4-dimensional (4D) cost volume to store a distance metric between each feature vector of the first set of feature vectors and a selected subset of feature vectors of the second set of feature vectors. The method further includes performing a flow-semi-global matching (Flow-SGM) on the 4D cost volume to estimate an optical flow vector for pixels of the first image.

Abstract: In an embodiment of the present invention, an interactive system employs sets of interior design guidelines. In an embodiment, the user begins by specifying the shape of a room and the set of furniture that must be arranged within it. The user then interactively moves furniture pieces. In response, the system suggests a set of furniture layouts that follow the interior design guidelines. The user can interactively select a suggestion and move any piece of furniture to modify the layout.