This letter presents an efficient technique to reduce the computational complexity involved in training binary convolutional neural networks (BCNN). The BCNN training shall be conducted focusing on the optimization of the sign of each weight element rather than the exact value itself in convention; in which, the sign of an element is not likely to be flipped anymore after it has been updated to have such a large magnitude to be clipped out. The proposed technique does not update such elements that have been clipped out and eliminates the computations involved in their optimization accordingly. The complexity reduction by the proposed technique is as high as 25.52% in training the BCNN model for the CIFAR-10 classification task, while the accuracy is maintained without severe degradation.

Mobile operators need to migrate from 2G to 3G networks in a cost-effective manner. Cognitive radio systems are currently being investigated as a promising solution to achieve spectrum efficiency by allowing coexistence of unlicensed (secondary) networks and licensed (primary) networks. However, conventional mechanisms to operate these systems incur additional complexity and fail to maximize network performance. In this paper, we propose a pilot sensing and frequency selection method with low complexity for OFDMA-based cognitive radio systems. Subject to the interference constraints imposed by the primary network, capacity maximization problems involving both up-link and down-link connections are considered for overall network performance improvement. The throughput and outage probability of the proposed method are evaluated by simulations. Our proposed method shows outstanding performance if the channel varies frequently in the primary network and the frequency reuse factor of the primary network is high.

This letter presents a novel technique to achieve a fast inference of the binarized convolutional neural networks (BCNN). The proposed technique modifies the structure of the constituent blocks of the BCNN model so that the input elements for the max-pooling operation are binary. In this structure, if any of the input elements is +1, the result of the pooling can be produced immediately; the proposed technique eliminates such computations that are involved to obtain the remaining input elements, so as to reduce the inference time effectively. The proposed technique reduces the inference time by up to 34.11%, while maintaining the classification accuracy.

Barrel distortion is a critical problem that can hinder the successful application of wide-angle cameras. This letter presents an implementation method for fast correction of the barrel distortion. In the proposed method, the required scaling factor is obtained by interpolating a mapping polynomial with a non-uniform spline instead of calculating it directly, which reduces the number of computations required for the distortion correction. This reduction in the number of computations leads to faster correction while maintaining quality: when compared to the conventional method, the reduction ratio of the correction time is about 89%, and the correction quality is 35.3 dB in terms of the average peak signal-to-noise ratio.

An efficient pruning method is proposed for the infinity-norm sphere decoding based on Schnorr-Euchner enumeration in multiple-input multiple-output spatial multiplexing systems. The proposed method is based on the characteristics of the infinity norm, and utilizes the information of the layer at which the infinity-norm value is selected in order to decide unnecessary sub-trees that can be pruned without affecting error-rate performance. Compared to conventional pruning, the proposed pruning decreases the average number of tree-visits by up to 37.16% in 44 16-QAM systems and 33.75% in 66 64-QAM systems.

The generalized spatial modulation (GSM) is a new transmission technique that can realize high-performance multiple-input multiple-output (MIMO) communication systems with a low RF complexity. This paper presents an efficient sphere decoding method used to perform the symbol detection for the generalized spatial modulation (GSM) multiple-input multiple-output (MIMO) systems. In the proposed method, the cost metric is modified so that it does not include the cancellation of the nonexistent interference. The modified cost metric can be computed by formulating a detection tree that has a regular structure representing the transmit antenna combinations as well as the symbol vectors, both of which are detected efficiently by finding the shortest path on the basis of an efficient tree search algorithm. As the tree search algorithm is performed for the regular detection tree to compute the modified but mathematically-equivalent cost metric, the efficiency of the sphere decoding is improved while the bit-error rate performance is not degraded. The simulation results show that the proposed method reduces the complexity significantly when compared with the previous method: for the 6×6 64QAM GSM-MIMO system with two active antennas, the average reduction rate of the complexity is as high as 45.8% in the count of the numerical operations.

This letter presents a technique to reduce the complexity of the soft-output multiple-input multiple-output symbol detection based on Dijkstra's algorithm. By observing that the greedy behavior of Dijkstra's algorithm can entail unnecessary tree-visits for the symbol detection, this letter proposes a technique to evict non-promising candidates early from the search space. The early eviction technique utilizes layer information to determine if a candidate is promising, which is simple but effective. When the SNR is 30dB for 6×6 64-QAM systems, the average number of tree-visits in the proposed method is reduced by 72.1% in comparison to that in the conventional Dijkstra's algorithm-based symbol detection without the early eviction.

Square-related functions such as square, inverse square, square-root and inverse square-root operations are widely used in digital signal processing and digital communication algorithms, and their efficient realizations are commonly required to reduce the hardware complexity. In the implementation point of view, approximate realizations are often desired if they do not degrade performance significantly. In this paper, we propose new linear approximations for the square-related functions. The traditional linear approximations need multipliers to calculate slope offsets and tables to store initial offset values and slope values, whereas the proposed approximations exploit the inherent properties of square-related functions to linearly interpolate with only simple operations, such as shift, concatenation and addition, which are usually supported in modern VLSI systems. Regardless of the bit-width of the number system, more importantly, the maximum relative errors of the proposed approximations are bounded to 6.25% and 3.13% for square and square-root functions, respectively. For inverse square and inverse square-root functions, the maximum relative errors are bounded to 12.5% and 6.25% if the input operands are represented in 20 bits, respectively.