A scalable hierarchical coreference method that employs a homomorphic compression scheme that supports addition and partial subtraction to more efficiently represent the data and the evolving intermediate results of probabilistic inference. The method may encode the features underlying conditional random field models of coreference resolution so that cosine similarities can be efficiently computed. The method may be applied to compressing features and intermediate inference results for conditional random fields. The method may allow compressed representations to be added and subtracted in a way that preserves the cosine similarities.

Certain aspects of the present disclosure provide a method of encoding data. The method generally includes receiving data comprising a fractional number comprising an exponential component and a fractional component, the exponential component being represented by an exponential bit sequence, the fractional component being represented by a fractional bit sequence. The method further includes determining if the fractional component is within a threshold of 0 or 1. The method further includes setting the fractional component to 0 when the fractional component is within the threshold of 0 or 1. The method further includes downscaling the fractional bit sequence based on a difference between the exponential component and a second threshold. The method further includes encoding the data. The method further includes transmitting the encoded data.

The disclosure is directed at a method of data compression. The method includes creating a set of single composite data structures and then calculating a set of bit probabilities based on the set of single data structures. The bit probabilities are then used to create a set of intermediate buffers which are then sorted and traversed for data compression.

A method, computer program product, and system includes a processor obtaining data including values and generating a value conversion dictionary by applying a parse tree based compression algorithm to the data, where the value conversion dictionary includes dictionary entries that represent the values. The processor obtains a distribution of the values and estimates a likelihood for each based on the distribution. The processor generates a code word to represent each value, a size of each code word is inversely proportional to the likelihood of the word. The processor assigns a rank to each code word, the rank for each represents the likelihood of the value represented by the code word; and based on the rank associated with each code word, the processor reorders each dictionary entry in the value conversion dictionary to associate each dictionary entry with an equivalent rank, the reordered value conversion dictionary comprises an architected dictionary.

Computer-implemented methods, systems, and devices to perform lossless compression of floating point format time-series data are disclosed. A first data value may be obtained in floating point format representative of an initial time-series parameter. For example, an output checkpoint of a computer simulation of a real-world event such as weather prediction or nuclear reaction simulation. A first predicted value may be determined representing the parameter at a first checkpoint time. A second data value may be obtained from the simulation. A prediction error may be calculated. Another predicted value may be generated for a next point in time and may be adjusted by the previously determined prediction error (e.g., to increase accuracy of the subsequent prediction). When a third data value is obtained, the adjusted prediction value may be used to generate a difference (e.g., XOR) for storing in a compressed data store to represent the third data value.

Embedded codec (EBC) circuitry for frequency-dependent coding of transform coefficients, groups a plurality of transform coefficients for an input image block into a plurality of groups of transform coefficients. The plurality of transform coefficients are grouped based on a frequency distribution of the plurality of transform coefficients for the input image block. The EBC circuitry selects a different entropy coding parameter from a set of entropy coding parameters for each group of the plurality of groups, based on the frequency distribution. Thereafter, the EBC circuitry applies an entropy coding scheme from a set of entropy coding schemes to each group of transform coefficients, in accordance with the selected entropy coding parameter.

Methods of encoding and decoding data in which some data symbols are entropy coded and some data symbols are bypass coded. The encoder separates the coded symbols into an entropy coded stream and a bypass coded stream. The streams are packaged in a data unit that has a payload structured to contain one of the streams in forward order and the other stream in reverse order, with the reverse order stream aligned with the end of the data unit. In this manner, at the decoder, the decoder may begin decoding the forward order stream from its beginning and may also begin decoding the reverse order stream from its beginning at the end of the data unit by extracting symbols in reverse order. The data unit does not need to signal the length of the streams. The decoder determines the length of the data unit from explicit or implicit signaling.

In some embodiments, a method determines a plurality of classes of bins that are used to determine a context model for entropy coding of a current block in a video. The method calculates a first value for a first class of bins in the plurality of classes of bins and calculates a second value for a second class of bins in the plurality of classes of bins. The first value for the first class of bins is weighted by a first weight to generate a weighted first value and the second value for the second class of bins is weighted by a second weight to generate a weighted second value. The method then selects a context model based on the first weighted value and the second weighted value.

A method (C) for converting a data signal (U), comprising (i) providing an input symbol stream (B) representative of the data signal (U), (ii) demultiplexing (DMX) the input symbol stream (B) to consecutively decompose the input symbol stream (B) into a number m of decomposed partial symbol streams (B_1, . . . , B_m), (iii) applying on each of the decomposed partial symbol streams (B_1, . . . , B_m) an assigned distribution matching process (DM_1, . . . , DM_m), thereby generating and outputting for each decomposed partial symbol stream (B_1, . . . , B_m) a respective pre-sequence (bn_1, . . . , bn_m) or n_j symbols as an intermediate output symbol sequence, and (iv) supplying the pre-sequences (bn_1, . . . , bn_m) to at least one symbol mapping process (BM) to generate and output a signal representative for a final output symbol sequence (S) as a converted data signal. Each of the distribution matching processes (DM_1, . . . , DM_m) and the symbol mapping process (BM) are based on a respective assigned alphabet (ADM_1, . . . , ADM_m; ABM) of symbols and the cardinality of each of the alphabets (ADM_1, . . . , ADM_m) of the distribution matching processes (DM_1, . . . , DM_m) is lower than the cardinality of the alphabet (ABM) of the symbol mapping process (BM).

An example method of compressing a data set includes determining whether individual values from a data set correspond to a first category or a second category of values. Based on one of the values corresponding to the first category, the value is added to a compressed data set. Based on one of the values corresponding to the second category, the value is excluded from the compressed data set, and a statistical distribution of values of the second category is updated based on the value. During a first phase, the determining is performed for a plurality of values from a first portion of the data set based on comparison of the values to criteria. During a second phase, the determining is performed for a plurality of values from a second portion of the data set based on the statistical distribution.