A method for adapting a trained neural network is provided. Input data is input to the trained neural network and a plurality of filters are applied to generate a plurality of channels of activation data. Differences between corresponding activation values in the plurality of channels of activation data are calculated and an order of the plurality of channels is determined based on the calculated differences. The neural network is adapted so that it will output channels of activation data in the determined order. The ordering of the channels of activation data is subsequently used to compress activation data values by taking advantage of a correlation between activation data values in adjacent channels.

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 system utilizing a high throughput coding mode for CABAC in HEVC is described. The system may include an electronic device configured to obtain a block of data to be encoded using an arithmetic based encoder; to generate a sequence of syntax elements using the obtained block; to compare an Absolute-3 value of the sequence or a parameter associated with the Absolute-3 value to a preset value; and to convert the Absolute-3 value to a codeword using a first code or a second code that is different than the first code, according to a result of the comparison.

A method for decoding a picture performed by a decoding apparatus according to the present disclosure includes receiving a bitstream including residual information, deriving a quantized transform coefficient for a current block based on the residual information included in the bitstream, deriving a transform coefficient from the quantized transform coefficient based on a dequantization process, deriving a residual sample for the current block by applying an inverse transform to the derived transform coefficient, and generating a reconstructed picture based on the residual sample for the current block.

Aspects of the disclosure provide methods and apparatuses for neural network model compression/decompression. In some examples, an apparatus for neural network model decompression includes receiving circuitry and processing circuitry. The processing circuitry decodes, from a bitstream corresponding to a representation of a neural network, at least a syntax element to be applied to multiple blocks in the neural network. Then, the processing circuitry reconstructs, from the bitstream, weight coefficients in the blocks based on the syntax element.

An idea used herein is to use the same function for the dependency of the context and the dependency of the symbolization parameter on previously coded/decoded transform coefficients. Using the same function—with varying function parameter—may even be used with respect to different transform block sizes and/or frequency portions of the transform blocks in case of the transform coefficients being spatially arranged in transform blocks. A further variant of this idea is to use the same function for the dependency of a symbolization parameter on previously coded/decoded transform coefficients for different sizes of the current transform coefficient's transform block, different information component types of the current transform coefficient's transform block and/or different frequency portions the current transform coefficient is located within the transform block.

An apparatus of neural network model decompression includes processing circuitry. The processing circuitry can be configured to receive, from a bitstream of a compressed neural network representation, one or more first syntax elements associated with a 3-dimensional coding unit (CU3D) partitioned from a 3-dimensional coding tree unit (CTU3D). The first CTU3D can be partitioned from a tensor in a neural network. The one or more first syntax elements can indicate that the CU3D is partitioned based on a 3D pyramid structure that includes multiple depths. Each depth corresponds to one or more nodes. Each node has a node value. Second syntax elements corresponding to the node values of the nodes in the 3D pyramid structure can be received from the bitstream in a breadth-first scan order for scanning the nodes in the 3D pyramid structure. Model parameters of the tensor can be reconstructed based on the received second syntax elements.

Methods, devices and systems for signaling the use of subpictures in coded video pictures are described. One example method of video processing includes performing a conversion between a video and a bitstream of a video, wherein the bitstream conforms to a format rule, and wherein the format rule specifies that a first syntax element, which indicates whether a picture of the video can be partitioned, is conditionally included in a picture parameter set (PPS) of the bitstream based on values of a second syntax element, which indicates whether identifiers of subpictures are signaled in the PPS, and a third syntax element, which indicates, in the PPS, a number of the subpictures.

The invention relates to a method for compressing an input data set, wherein the coefficients in the input data set are grouped in groups of coefficients, a number of bit planes, GCLI, needed for representing each group is determined, a quantization is applied, keeping a limited number of bit planes, a prediction mechanism is applied to the GCLIs for obtaining residues, and an entropy encoding of the residues is performed. The entropy-encoded residues, and the bit planes kept allow the decoder to reconstruct the quantized data, at a minimal cost in meta-data.

The invention describes an encoding scheme for encoding a sequence of integer values, particularly by first using encoding a value at a predetermined position and then consecutively calculating and encoding differences between neighboring values. A respective decoding scheme is also provided. These encoding and decoding schemes provide a more efficient encoding concept.