Search

Quantization is a prominent approach to compress model sizes of deep neural networks (DNNs), which clusters high-precision weights into a smaller set of quantization levels and represents high-precision weights by low-precision indexes. To achieve the same accuracy, nonuniform quantized DNNs (NUQ-DNNs) with unequal quantization intervals need lower index precision than uniform quantized DNNs (UQ-DNNs) with equal intervals, achieving smaller model sizes. Hence, deploying NUQ-DNNs on accelerators costs less on- and off-chip memory accesses than UQ-DNNs, which are more valuable for edge devices. However, accelerating NUQ-DNNs is nontrivial, since weight indexes cannot be directly used for computations. Previous NUQ-DNN accelerators adopt standard convolutions by decoding weight indexes into actual-weights multiplied with activations, causing abundant look-up overhead and redundant computations. In this work, we propose a weight-repetition-aware activation aggregating (WPAA) convolution approach to accelerate inference of variable-precision NUQ- and UQ-DNNs. By merging convolutions of multiple kernels, WPAA requires no look-up operation and removes redundant computations. Based on WPAA, we design a generic quantized DNN accelerator (GQNA). Furthermore, we propose a layer-adaptive kernel-reordering merging scheme to off-line adjust merging order of kernels for minimizing energy consumption of GQNA. Implemented under TSMC 28-nm technology, GQNA achieves 31.9 and 32.6 TOPS/W energy efficiency for 1-b UQ- and NUQ-VGG-16, respectively. [ABSTRACT FROM AUTHOR]

In this paper, we propose an improvement of the normalized min-sum (MS) decoding algorithm and novel MS decoder architectures with reduced word length using nonuniform quantization schemes for low-density parity-check (LDPC) codes. The proposed normalized MS algorithm introduces a more exact adjustment with two optimized correction factors for check-node-updating computations, while the conventional normalized MS algorithm applies only one correction factor. The proposed algorithm provides a significant performance gain without any additional computation or hardware complexity. The finite word-length analysis in implementing an LDPC decoder is a very important factor since it directly impacts the size of memory to store the intrinsic and extrinsic messages and the overall hardware area in the partially parallel LDPC decoder. The proposed nonuniform quantization scheme can reduce the finite word length while achieving similar performances compared to a conventional quantization scheme. From simulation results, it is shown that the proposed 4-bit nonuniform quantization scheme achieves an acceptable decoding performance, unlike the conventional 4-bit uniform quantization scheme. Finally, the proposed MS decoder architectures by the nonuniform quantization scheme provide significant reductions of 20% and up to 8% for the memory area and combinational logic area, respectively, compared to the conventional 5-bit ones. [ABSTRACT FROM AUTHOR]

A direction-of-arrival (DOA) estimation method using an array antenna has been developed based on a database retrieval technique. This method uses a database consisting of a set of correlation matrices of the array output vectors for various combinations of the quantized angles and signal powers. When a correlation matrix is estimated from an observed output vector, several correlation matrices close to the estimated one are searched out from the database, and the DOA is estimated based on the retrieved data. This method gives an accurate estimation, but the use of uniform quantization step size requires a large amount of storage space. In this paper, the relation between the quantization step size and the estimation accuracy is analyzed, and a nonuniform quantization scheme is developed to reduce the database size without sacrificing the estimation accuracy. A clustering technique is also introduced to alleviate the performance degradation caused by the retrieval of data which have similar correlation matrices but have much different angles. We show by simulations that the nonuniform quantization reduces the database size and the clustering improves the estimation accuracy, and that the proposed method is applicable to the array of three elements at the present. [Copyright &y& Elsevier]