A method and an apparatus for adapting feature data in a convolutional neural network. The method includes selecting a plurality of consecutive layers; determining an expected number of subdata blocks and a layout position, width and height of each subdata block in an output feature data of a last layer; determining, for each current layer, a layout position, width, and height of each subdata block of an input feature data for the current layer according to the layout position, width, and height of each subdata block of the output feature data for the current layer; determining an actual position of each subdata block of the input feature data for a first layer in the input feature data for the first layer; and obtaining the expected number of subdata blocks of the input feature data for the first layer according to the actual position, width and height of each subdata block of the input feature data for the first layer.

Systems and methods for programming field programmable gate array (FPGA) devices are provided. A trained model for a deep learning process is obtained and converted to design abstraction (DA) code defining logic block circuits for programming an FPGA device. Each of these logic block circuits represents one of a plurality of modules that executes a processing step between different layers of the deep learning process.

An electronic apparatus for compressing a recurrent neural network and a method thereof are provided. The electronic apparatus and the method thereof include a sparsification technique for the recurrent neural network, obtaining first to third multiplicative variables to learn the recurrent neural network, and performing sparsification for the recurrent neural network to compress the recurrent neural network.

A system, method and computer product for training a neural network system. The method comprises inputting an audio signal to the system to generate plural outputs f(X, Θ). The audio signal includes one or more of vocal content and/or musical instrument content, and each output f(X, Θ) corresponds to a respective one of the different content types. The method also comprises comparing individual outputs f(X, Θ) of the neural network system to corresponding target signals. For each compared output f(X, Θ), at least one parameter of the system is adjusted to reduce a result of the comparing performed for the output f(X, Θ), to train the system to estimate the different content types. In one example embodiment, the system comprises a U-Net architecture. After training, the system can estimate various different types of vocal and/or instrument components of an audio signal, depending on which type of component(s) the system is trained to estimate.

Aspects of the present disclosure enable humanly-specified relationships to contribute to a mapping that enables compression of the output structure of a machine-learned model. An exponential model such as a maximum entropy model can leverage a machine-learned embedding and the mapping to produce a classification output. In such fashion, the feature discovery capabilities of machine-learned models (e.g., deep networks) can be synergistically combined with relationships developed based on human understanding of the structural nature of the problem to be solved, thereby enabling compression of model output structures without significant loss of accuracy. These compressed models provide improved applicability to “on device” or other resource-constrained scenarios.

An electronic apparatus is provided. The electronic apparatus includes sample data and memory storing a first matrix included in an artificial intelligence model trained based on sample data, and a processor configured to prunes each of a plurality of first elements included in the first matrix based on a first threshold, and acquire a first pruning index matrix that indicates whether each of the plurality of first elements has been pruned with binary data, factorize the first matrix to a second matrix of which size was determined based on the number of rows and the rank, and a third matrix of which size was determined based on the rank and the number of columns of the first matrix, prunes each of a plurality of second elements included in the second matrix based on a second threshold, and acquire a second pruning index matrix that indicates whether each of the plurality of second elements has been pruned with binary data, prunes each of a plurality of third elements included in the third matrix based on a third threshold, and acquire a third pruning index matrix that indicates whether each of the plurality of third elements has been pruned with binary data, acquire a final index matrix based on the second pruning index matrix and the third pruning index matrix, and update at least one of the second pruning index matrix or the third pruning index matrix by comparing the final index matrix with the first pruning index matrix.

Presented herein are systems and methods that provide for automated analysis of medical images to determine a predicted disease status (e.g., prostate cancer status) and/or a value corresponding to predicted risk of the disease status for a subject. The approaches described herein leverage artificial intelligence (AI) to analyze intensities of voxels in a functional image, such as a PET image, and determine a risk and/or likelihood that a subject's disease, e.g., cancer, is aggressive. The approaches described herein can provide predictions of whether a subject that presents a localized disease has and/or will develop aggressive disease, such as metastatic cancer. These predictions are generated in a fully automated fashion and can be used alone, or in combination with other cancer diagnostic metrics (e.g., to corroborate predictions and assessments or highlight potential errors). As such, they represent a valuable tool in support of improved cancer diagnosis and treatment.

Techniques are presented for accelerating in-memory matrix multiplication operations for a convolution neural network (CNN) inference in which the weights of a filter are stored in the memory of a storage class memory device, such as a ReRAM or phase change memory based device. To improve performance for inference operations when filters exhibit sparsity, a zero column index and a zero row index are introduced to account for columns and rows having all zero weight values. These indices can be saved in a register on the memory device and when performing a column/row oriented matrix multiplication, if the zero row/column index indicates that the column/row contains all zero weights, the access of the corresponding bit/word line is skipped as the result will be zero regardless of the input.

Various embodiments are provided for automating decision making for a neural architecture search by one or more processors in a computing system. One or more specifications may be automatically selected for a dataset, tasks, and one or more constraints for a neural architecture search. The neural architecture search may be performed based on the one or more specifications. A deep learning model may be suggested, predicted, and/or configured for the dataset, the tasks, and the one or more constraints based on the neural architecture search.

Methods, techniques and systems for providing self-pruning neural networks are disclosed. A neural network including a plurality of layers may be trained using a batch sampled from a dataset. In addition to simulated neurons, individual ones of the plurality of layers include respective auxiliary parameters that may identify relative contributions of respective layers to the accuracy of the trained model. The respective layers of the neural network may be trained using a training batch to determine a penalty according to a regularization penalty for the neural network, the penalty determined according to a number of layers in the neural network. Prior to completion of the training batch and in accordance with the regularization penalty, one or more neurons of the neural network may be identified and deleted using the respective auxiliary parameters, thus providing a self-pruning mechanism to control growth and resource demands for the neural network.