A method is disclosed to reduce computation in a self-attention deep-learning model. A feature-map regularization term is added to a loss function while training the self-attention model. At least one low-magnitude feature is removed from at least one feature map of the self-attention model during inference. Weights of the self-attention model are quantized after the self-attention model has been trained. Adding the feature-map regularization term reduces activation values of feature maps, and removing the at least one low-magnitude feature from at least one feature map may be performed by setting the low-magnitude feature to be equal to zero based on the low-magnitude feature having a value that is less than a predetermined threshold. Feature maps of the self-attention model quantized and compressed.

A method of combining compressed uplink signals according to one aspect of the present disclosure is a method of combining IQ data of uplink signals compressed without decompression by a block floating point compression method, a block scaling compression method, or a -law compression method.

Methods, computer program products, and/or systems are provided that perform the following operations: in a data replication environment, analyzing a database workload to generate a knowledge base of information related to compression; dividing a transfer data stream into different segments based, at least in part, on the knowledge base; obtaining candidate compression types for the transfer data stream based, at least in part, on the knowledge base; assigning respective compression types of the candidate compression types to the different segments; generating compressed segments based, at least in part, on the respective compression types assigned to the different segments; and providing the compressed segments to a replication target.

One example method includes file specific compression selection. Compression metrics are generated for a chunk of a file using a reference compressor. Compression metrics for other compressors are determined from the metrics of the reference compressor. A compressor is then selected to compress the file.

A system and method for highly efficient encoding of data that includes extended functionality for asymmetric encoding/decoding and network policy enforcement. In the case of asymmetric encoding/decoding the original data is encoded by an encoder according to a codebook and sent to a decoder, but the output of the decoder depends on data manipulation rules applied at the decoding stage to transform the decoded data into a different data set from the original data. In the case of network policy enforcement, a behavior appendix into the codebook, such that the encoder and/or decoder at each node of the network comply with network behavioral rules, limits, and policies during encoding and decoding.

Apparatus and methods for training a neural network accelerator using quantized precision data formats are disclosed, and in particular for storing activation values from a neural network in a compressed format having lossy or non-uniform mantissas for use during forward and backward propagation training of the neural network. In certain examples of the disclosed technology, a computing system includes processors, memory, and a compressor in communication with the memory. The computing system is configured to perform forward propagation for a layer of a neural network to produced first activation values in a first block floating-point format. In some examples, activation values generated by forward propagation are converted by the compressor to a second block floating-point format having a non-uniform and/or lossy mantissa. The compressed activation values are stored in the memory, where they can be retrieved for use during back propagation.

Compressing files is disclosed. An input file to be compressed is first aligned. Aligning the file includes splitting the file into sequences that can be aligned. When splitting the file into sequences or when performing subsequent recursive splitting, the splitting is based on a longest sequence match. The result is a compression matrix, where each row of the matrix corresponds to part of the file. A consensus sequence is determined from the compression matrix. Using the consensus sequence, pointer pairs are generated. Each pointer pair identifies a subsequence of the consensus matrix. The compressed file includes the pointer pairs and the consensus sequence.

A data output method, a data acquisition method, a device, and an electronic apparatus are provided, and a specific technical solution is: reading a first data sub-block, and splicing the first data sub-block into a continuous data stream, wherein the first data sub-block is a data sub-block in transferred data in a neural network; compressing the continuous data stream to acquire a second data sub-block; determining, according to a length of the first data sub-block and a length of the second data sub-block, whether there is a gain in compression of the continuous data stream; outputting the second data sub-block if there is the gain in the compression of the continuous data stream.

A data output method, a data acquisition method, a device, and an electronic apparatus are provided, and a specific technical solution is: reading a first data sub-block, and splicing the first data sub-block into a continuous data stream, wherein the first data sub-block is a data sub-block in transferred data in a neural network; compressing the continuous data stream to acquire a second data sub-block; determining, according to a length of the first data sub-block and a length of the second data sub-block, whether there is a gain in compression of the continuous data stream; outputting the second data sub-block if there is the gain in the compression of the continuous data stream.

An electronic device that includes a decompression engine that includes N decoders and a decompressor decompresses compressed input data that includes N streams of data. Upon receiving a command to decompress compressed input data, the decompression engine causes each of the N decoders to decode a respective one of the N streams from the compressed input data separately and substantially in parallel with others of the N decoders. Each decoder outputs a stream of decoded data of a respective type for generating commands associated with a compression standard for decompressing the compressed input data. The decompressor next generates, from the streams of decoded data output by the N decoders, commands for decompressing the data using the compression standard to recreate the original data. The decompressor next executes the commands to recreate the original data and stores the original data in a memory or provides the original data to another entity.