Zhongfeng Zhang

We propose a method for minimizing global buffer access within a deep learning accelerator for convolution operations by maximizing the data reuse through a local register file, thereby substituting the local register file access for the power-hungry global buffer access. To fully exploit the merits of data reuse, this study proposes a rearrangement of the computational sequence in a deep learning accelerator. Once input data are read from the global buffer, repeatedly reading the same data is performed only through the local register file, saving significant power consumption. Furthermore, different from prior works that equip local register files in each computation unit, the proposed method enables sharing a local register file along the column of the 2D computation array, saving resources and controlling overhead. The proposed accelerator is implemented on an off-the-shelf field-programmable gate array to verify the functionality and resource utilization. Then, the performance improvement of the proposed method is demonstrated relative to popular deep learning accelerators. Our evaluation indicates that the proposed deep learning accelerator reduces the number of global-buffer accesses to nearly 86.8%, consequently saving up to 72.3% of the power consumption for the input data memory access with a minor increase in resource usage compared to a conventional deep learning accelerator.

In a Wi-Fi indoor positioning system (IPS), the performance of the IPS depends on the channel state information (CSI), which is often limited due to the multipath fading effect, especially in indoor environments involving multiple non-line-of-sight propagation paths. In this paper, we propose a novel IPS utilizing trajectory CSI observed from predetermined trajectories instead of the CSI collected at each stationary location; thus, the proposed method enables all the CSI along each route to be continuously encountered in the observation. Further, by using a generative adversarial network (GAN), which helps enlarge the training dataset, the cost of trajectory CSI collection can be significantly reduced. To fully exploit the trajectory CSI’s spatial and temporal information, the proposed IPS employs a deep learning network of a one-dimensional convolutional neural network–long short-term memory (1DCNN-LSTM). The proposed IPS was hardware-implemented, where digital signal processors and a universal software radio peripheral were used as a modem and radio frequency transceiver, respectively, for both access point and mobile device of Wi-Fi. We verified that the proposed IPS based on the trajectory CSI far outperforms the state-of-the-art IPS based on the CSI collected from stationary locations through extensive experimental tests and computer simulations.

Recently, filter bank multicarrier with offset quadrature amplitude modulation (FBMC/OQAM) has received increasing attention from researchers, owing to its merits and superior spectral efficiency. High peak-to-average power ratio (PAPR) occurs in approximately all multicarrier systems, including FBMC/OQAM, and may cause bit-error-rate (BER) degradation if not appropriately handled. Conventional PAPR reduction methods for orthogonal frequency division multiplexing (OFDM), such as partial transmit sequence (PTS), selective mapping (SLM), and discrete Fourier transform (DFT) spreading, are ineffective in FBMC/OQAM because of the different structure of the symbols. This study proposes a novel method combining DFT spreading and PTS methods to reduce the PAPR of FBMC/OQAM systems with reasonable computational complexity. Numerical results obtained from various computer simulations show that the proposed method achieves a noticeable enhancement in the PAPR performance of the FBMC/OQAM signal compared to other existing methods without affecting the BER performance. Further, the computational complexity analysis and BER performance of the proposed method are presented in comparison to typical existing methods. From our computer simulations, the proposed method reduces the PAPR by approximately 32.8% compared to that of the conventional methods, and the BER performance is improved by 25% with a high-power amplifier effect.

Backward compatibility is one of the key issues for radio equipment that supports IEEE 802.11, which is a typical communication protocol for wireless local area networks (WLANs). For achieving successful packet decoding with backward compatibility, frame format detection is the core precondition. In this study, we present a novel, deep-learning-based frame format detection method for IEEE 802.11 WLANs. Considering that the detection performance of conventional methods is mainly degraded because of poor performance in symbol synchronization and/or channel estimation in environments with a low signal-to-noise ratio, we propose a novel detection method based on a deep learning network to replace conventional detection procedures. The proposed deep-learning network method achieves robust detection directly from the received (Rx) data. Through extensive computer simulations performed in multipath fading channel environments (modeled by Project IEEE 802.11 Task Group ac), we confirmed that the proposed method exhibits significantly higher frame format detection performance than that of the conventional method.