Search

Efficient implementation of deep neural networks (DNNs) on CPU-based systems is critical owing to the proliferation of applications in embedded and Internet of Things systems. Nowdays, most CPUs are equipped with single instruction multiple data (SIMD) instructions, which are used to implement an efficient general matrix multiply (GEMM) library for accelerating DNN inference. Quantized neural networks are actively investigated to simplify DNN computation and memory requirements; however, the current CPU libraries do not efficiently support arithmetic operations below eight bits. Hence, we developed TernGEMM, a GEMM library composed of SIMD instructions for DNNs with ternary weights and sub-8-bit activations. TernGEMM is implemented using simple logical operations that replace the long-latency multiply–add operation. Instead of fixing the accumulation bit precision as 32-bit, TernGEMM accumulates the partial sums in a bit-incremental manner to exploit parallelism in 8-bit and 16-bit SIMD instructions. Furthermore, we propose different tile sizes for TernGEMM to better support the diverse dimensions of DNNs. Compared with a state-of–the-art reduced precision DNN GEMM library, i.e., GEMMLowp, TernGEMM achieve × 1.785 to × 4.147 speedup for ResNet50, MobileNet-V2, and EfficientNet-B0, as evaluated on both Intel and ARM CPUs. [ABSTRACT FROM AUTHOR]

Recent theory of compressed sensing (CS) tells us that sparse signals can be reconstructed from a small number of random samples. In reconstruction of sparse signals, greedy algorithms, such as the orthogonal matching pursuit (OMP), have been shown to be computationally efficient. In this paper, the performance of OMP is shown to be dependent on how well information of the underlying signals is preserved in the residual vector. Further, to improve the information preservation, we present a modification of OMP, called oblique projection matching pursuit (ObMP), which updates the residual in a oblique projection manor. Using the restricted isometric property (RIP), we build a solid yet very intuitive grasp of the more accurate phenomenon of ObMP. We also show from empirical experiments that the ObMP achieves improved reconstruction performance over the conventional OMP algorithm in terms of support detection ratio and mean squared error (MSE). [ABSTRACT FROM AUTHOR]

Frequency estimation of a tonal signal in passive sonar systems is crucial to the identification of the marine object. In the conventional techniques, a basis mismatch error caused by the discretization of the frequency domain is unavoidable, resulting in a severe degradation of the object detection quality. To overcome the basis mismatch error, we propose a tonal frequency estimation technique in the continuous frequency domain. Towards this end, we formulate the frequency estimation problem as an atomic norm minimization problem. From the numerical experiments, we show that the proposed technique is effective in identifying the tonal frequency components of marine objects. [ABSTRACT FROM AUTHOR]