Disclosed are a method of encoding and decoding an audio signal and an encoder and a decoder performing the method. The method of encoding an audio signal includes identifying an input signal, and generating a bitstring of each encoding layer by applying, to the input signal, an encoding model including a plurality of successive encoding layers that encodes the input signal, in which a current encoding layer among the encoding layers is trained to generate a bitstring of the current encoding layer by encoding an encoded signal which is a signal encoded in a previous encoding layer and quantizing an encoded signal which is a signal encoded in the current encoding layer.

A Unit Element (UE) has a positive UE and a negative UE, each having a digital X input and a digital W input with a sign bit, the sign bit is exclusive ORed with a chop clock to generate a chopped sign bit. The positive UE is enabled when the chopped sign bit is positive and the negative UE is enabled when the chopped sign bit is negative. Each positive and negative UE comprises groups of NAND gates generating an output and complementary output which are coupled to a differential charge transfer bus comprising a positive charge transfer line and a negative charge transfer line. The NAND gate outputs and complementary outputs are coupled through binary weighted charge transfer capacitors the positive charge transfer line and negative charge transfer line.

Provided is a processor implemented method that includes performing training or an inference operation with a neural network by obtaining a parameter for the neural network in a floating-point format, applying a fractional length of a fixed-point format to the parameter in the floating-point format, performing an operation with an integer arithmetic logic unit (ALU) to determine whether to round off a fixed point based on a most significant bit among bit values to be discarded after a quantization process, and performing an operation of quantizing the parameter in the floating-point format to a parameter in the fixed-point format, based on a result of the operation with the ALU.

A method and device for model compression of a neural network. The method comprises: recording input and output parameters of each layer of network in a network structure; dividing the network structure into several small networks according to the input and output parameters; setting a pruning flag bit of a first convolutional layer in each small network to be zero to obtain a pruned small network; training each pruned small network to obtain a network weight and a weight mask; recording a pruned channel index number of each convolutional layer of a pruned small network with the weight mask of zero; and carrying out decomposition calculation on each pruned small network according to the pruned channel index number. According to the method, a calculation amount and the size of a model is reduced, and during network deployment, the model can be loaded with one click, thus reducing usage difficulty.

When it is assumed that a large-scale Deep Neural Network for autonomous driving applied compression, there are problems of a decrease in recognition accuracy of a post-compression Neural Network (NN) model and an increase in a compression design period, due to a large number of harmful or unnecessary training images (invalid training images). A training image selection unit B100 calculates an influence value on an inference, and generates an indexed training image set 1004-1 necessary for an NN compression design, by using the influence value. A neural network compression unit P200 notified of the result via a memory P300 compresses the NN.

A method of generating a quantized neural network comprises (i) receiving a pre-trained neural network model and (ii) modifying the pre-trained neural network model to calculate one or more statistics on an output of one or more layers of the pre-trained neural network model based on a current image and set up an output data format for one or more following layers of the pre-trained neural network model for one or more of the current image and a subsequent image dynamically based on the one or more statistics.

A data processing method related to the field of artificial intelligence includes adding an architecture parameter to each feature interaction item in a first model, to obtain a second model, where the first model is a factorization machine (FM)-based model, and the architecture parameter represents importance of a corresponding feature interaction item; performing optimization on architecture parameters in the second model to obtain the optimized architecture parameters; and obtaining, based on the optimized architecture parameters and the first model or the second model, a third model through feature interaction item deletion.

A system for non-reference video-quality prediction includes a video-processing block to receive an input bitstream and to generate a first vector, and a neural network to provide a predicted-quality vector after being trained using training data. The training data includes the first vector and a second vector, and elements of the first vector include high-level features extracted from a high-level syntax processing of the input bitstream.

Artificial intelligence is an increasingly important sector of the computer industry. However, artificial intelligence is extremely computationally intensive field such that it can be expensive, time consuming, and energy consuming. Fortunately, many of the calculations required for artificial intelligence can be performed in parallel such that specialized processors can great increase computational performance. Specifically, artificial intelligence generally requires large numbers of matrix operations to implement neural networks such that specialized matrix processor circuits can improve performance. To perform all these matrix operations, the matrix processor circuits must be quickly and efficiently supplied with data to process or else the matrix processor circuits end up idle or spending large amounts of time loading in different weight matrix data. Thus, this document discloses apparatus and methods for efficiently operating external interfaces on neural network processors that efficiently interleave and overlap external memory operations along with computation operations such that both the external interface and matrix processor circuits are used efficiently.

A method for implementing an artificial neural network in a computing system that comprises performing a compute operation using an input activation and a weight to generate an output activation, and modifying the output activation using a noise value to increase activation sparsity.