A convolutional neural network pruning method based on feature map sparsification, which relates to how to compress the convolutional neural network to reduce the number of parameters and amount of computation so as to facilitate actual deployment, is provided. In the training process, by adding regularization to the feature map L1 or L2 after the activation layer in the loss function, the corresponding feature map channels have different sparsity. Under a certain pruned ratio, the convolution kernels corresponding to the channels are pruned according to the sparsity of the feature map channels. After fine-tune pruning, the network obtains new accuracy, and the pruned ratio is adjusted according to the change of accuracy before and after pruning. After multiple iterations, the near-optimal pruned ratio is found, and pruning is realized to the maximum extent under the condition that the accuracy does not decrease.

A system, method and program product for implementing a sparse sampling strategy for acquiring MRI data. A method includes: collecting and labeling a training dataset of MRI scans for a predetermined diagnostic; selecting a sampling shape and associated parameter values; sampling each MRI scan in the training data set using the sampling shape and associated parameter values to generate a set of sparse samples; training a neural network using the sparse samples and assigning an accuracy to a resulting trained neural network; and adjusting the associated parameter values, and repeating the sampling and training until optimized parameter values are established.

An embodiment provided herein preprocesses the input samples to the classification neural network, e.g., by adding Gaussian noise to word/sentence representations to make the function of the neural network satisfy Lipschitz property such that a small change in the input does not cause much change to the output if the input sample is in-distribution. Method to induce properties in the feature representation of neural network such that for out-of-distribution examples the feature representation magnitude is either close to zero or the feature representation is orthogonal to all class representations. Method to generate examples that are structurally similar to in-domain and semantically out-of domain for use in out-of-domain classification training. Method to prune feature representation dimension to mitigate long tail error of unused dimension in out-of-domain classification. Using these techniques, the accuracy of both in-domain and out-of-distribution identification can be improved.

An embodiment proposed herein uses sparsification techniques to train the neural network with a high feature dimension that may yield desirable in-domain detection accuracy but may prune away dimensions in the output that are less important. Specifically, a sparsification vector is generated based on Gaussian distribution (or other probabilistic distribution) and is used to multiply with the higher dimension output to reduce the number of feature dimensions. The pruned output may be then used for the neural network to learn the sparsification vector. In this way, out-of-distribution detection accuracy can be improved.

This document describes “Multi-domain Neighborhood Embedding and Weighting” (MNEW) for use in processing point cloud data, including sparsely populated data obtained from a lidar, a camera, a radar, or combination thereof. MNEW is a process based on a dilation architecture that captures pointwise and global features of the point cloud data involving multi-scale local semantics adopted from a hierarchical encoder-decoder structure. Neighborhood information is embedded in both static geometric and dynamic feature domains. A geometric distance, feature similarity, and local sparsity can be computed and transformed into adaptive weighting factors that are reapplied to the point cloud data. This enables an automotive system to obtain outstanding performance with sparse and dense point cloud data. Processing point cloud data via the MNEW techniques promotes greater adoption of sensor-based autonomous driving and perception-based systems.

Systems, apparatuses and methods may provide for technology that decodes data via an instruction that indicates a number of rulebooks to be processed, an input feature size, an output feature size, and a plurality of feature map base addresses, rearranges spatially distributed voxel output feature maps in the decoded data based on weight planes, and performs a channel-wise multiply-accumulate (MAC) operation on the rearranged spatially distributed voxel output feature maps to obtain an output, wherein the channel-wise MAC operation is performed as partial accumulations by a plurality of processing elements.

Techniques to improve performance of matrix multiply operations are described in which a compute kernel can specify one or more element-wise operations to perform on output of the compute kernel before the output is transferred to higher levels of a processor memory hierarchy.

Image data is run through a neural network, and the neural network produces a vector representation of the image data. Random sparse sampling masks are created. The vector representation of the image data is masked with each of the random sparse sampling masks, the masking generating corresponding sparsely sampled vectors. The sparsely sampled vectors are transmitted to nodes of a consensus network, wherein a sparsely sampled vector of the sparsely sampled vectors is transmitted to a node of the consensus network. Votes from the nodes of the consensus network are received. Whether a consensus is achieved in the votes is determined. Responsive to determining that the consensus is achieved, at least one of identification and verification of the image data may be provided.

Described is a system for object recognition. The system generates a training image set of object images from multiple image classes. Using a training image set and annotated semantic attributes, a model is trained that maps visual features from known images to the annotated semantic attributes using joint sparse representations with respect to dictionaries of visual features and semantic attributes. The trained model is used for mapping visual features of an unseen input image to its semantic attributes. The unseen input image is classified as belonging to an image class, and a device is controlled based on the classification of the unseen input image.

Techniques to improve performance of matrix multiply operations are described in which a compute kernel can specify one or more element-wise operations to perform on output of the compute kernel before the output is transferred to higher levels of a processor memory hierarchy.