Voxel-based neural networks have been widely applied to the processing of large-scale outdoor point cloud data, which first convert points into voxels and then extract features using several sparse convolution and normal convolution layers. The hardware implementation of these networks suffers from complex rulebook generation, irregular memory access, and low hardware utilization. Meanwhile, these networks still have much data sparsity. In this paper, we propose an energy and area-efficient query-based sparsity-aware accelerator for voxel-based point cloud neural networks, namely QSAP. Specifically, a dedicated unit is used to improve the efficiency of rulebook generation. A query-based input feature-writing method is proposed to enhance parallel reading potential. An efficient weight-mapping method is introduced to store unpruned weights in on-chip buffers. A novel input-feature-reading method with consecutive queries is proposed to enable out-of-order execution of convolution operations, thereby improving hardware utilization. The hardware utilization is further enhanced by a proposed pop strategy that minimizes the total number of empty FIFOs. The MAC related to zero value is also skipped in this process. As a result, QSAP achieves superior performance on 22 nm technology with the throughput, energy efficiency, area efficiency, frame rate, and frame energy of 1074 GOPS, 7.79 TOPS/W, 590 GOPS/mm$\mathbf {^{2}}$, 35.2 FPS, and 3.91 mJ/Frame, respectively, better than state-of-the-art works.

The traditional Mongolian script poses significant challenges for text rendering due to its vertical orientation, complex glyph variations, and context-dependent shapes, compounded by limited linguistic resources. This paper introduces the first method specifically designed to address the out-of-vocabulary (OOV) word problem in traditional Mongolian script conversion. We present a novel dataset and propose MonTransformer, the first Transformer-based model tailored for converting Unicode phonetic sequences into glyph sequences. MonTransformer utilizes dynamic optimization and contrastive learning to effectively manage OOV cases and improve sequence generation accuracy, transforming Unicode inputs into visually accurate glyph representations. Our approach achieves state-of-the-art performance, maintaining the script’s visual fidelity and advancing efforts in digital preservation for complex scripts.

The game of Tibetan Go faces the scarcity of expert knowledge and research literature. Therefore, we study the zero learning model of Tibetan Go under limited computing power resources and propose a novel scale-invariant U-Net style two-headed output lightweight network TibetanGoTinyNet. The lightweight convolutional neural networks and capsule structure are applied to the encoder and decoder of TibetanGoTinyNet to reduce computational burden and achieve better feature extraction results. Several autonomous self-attention mechanisms are integrated into TibetanGoTinyNet to capture the Tibetan Go board’s spatial and global information and select important channels. The training data are generated entirely from self-play games. TibetanGoTinyNet achieves 62%–78% winning rate against other four U-Net style models including Res-UNet, Res-UNet Attention, Ghost-UNet, and Ghost Capsule-UNet. It also achieves 75% winning rate in the ablation experiments on the attention mechanism with embedded positional information. The model saves about 33% of the training time with 45%–50% winning rate for different Monte-Carlo tree search (MCTS) simulation counts when migrated from 9 × 9 to 11 × 11 boards. Code for our model is available at https://github.com/paulzyy/TibetanGoTinyNet .

The rules of the two phases of Tibetan Jiu chess, layout and battle, are very different, and using the same UCT search algorithm globally will result in a large overhead of time and storage space in the search process, so a phased game algorithm for Tibetan Jiu chess is proposed, with different strategies designed for the layout and battle phases, respectively. First, the layout phase uses a combination of Gaussian distribution and fast online estimation to improve the UCT algorithm, thus generating the optimal action selection scheme. Second, in order to take full advantage of reinforcement learning and deep learning, a neural network model with residual network structure is used in the battle phase to guide the search of Monte Carlo trees, and the default strategy is improved by "pruning" in the expansion step to improve the quality of the expanded nodes. The dataset is generated by self-play and used to train the neural network model to obtain the optimal model. It is verified through experiments that the phased gaming algorithm proposed in this study effectively reduces the process of blindly exploring the board state during the layout and battle phases of the UCT search algorithm, and improves the quality of the layout and the self-learning efficiency of the neural network model.

The sparrow search algorithm (SSA) has received widespread attention as an emerging group intelligence algorithm. In this study, a QSAR (quantitative structure-activity relationship) prediction model for anti-ERa (Estrogen receptor alpha, an important target for the treatment of breast cancer) compounds and their bioactivity data was constructed based on a sparrow search algorithm optimized BP neural network combined with the selected relevant parameter indicators. Then the created model was used to predict the biological activity of the compounds. The results show that the created model not only has a certain self-learning function, but also improves the convergence speed and the accuracy of the prediction results compared with the BP neural network model optimized by genetic algorithm (which is more complex in coding and slower in solving, but has good solution accuracy). Further, it was demonstrated that the use of SSA-BP could improve the prediction of QSAR for anti-ERa compounds.

To tackle the problem of aquatic environment pollution, a vision-based autonomous underwater garbage cleaning robot has been developed in our laboratory. We propose a garbage detection method based on a modified YOLOv4, allowing high-speed and high-precision object detection. Specifically, the YOLOv4 algorithm is chosen as a basic neural network framework to perform object detection. With the purpose of further improvement on the detection accuracy, YOLOv4 is transformed into a four-scale detection method. To improve the detection speed, model pruning is applied to the new model. By virtue of the improved detection methods, the robot can collect garbage autonomously. The detection speed is up to 66.67 frames/s with a mean average precision (mAP) of 95.099%, and experimental results demonstrate that both the detection speed and the accuracy of the improved YOLOv4 are excellent.

Machine learning algorithms are widely applied in the fields such as Machine Vision, Natural Language Processing, Image Processing and Automatic Speech Recognition. Deep learning based on mass data has achieved success in many application, for example, Alpha Go. However, learning from small data remains a key challenge in machine learning. This paper introduces the basic ideas of the small data learning theory, present the main machine learning algorithms and analyze their major characteristics. The paper also summarizes the current research trends.

In this paper, the problems are discussed for shape design and position arrangement for Water Strider Robot's supporting legs are discussed. A supporting leg is approached as Euler-Bernoulli elastic curved beam and a method for designing its optimal shape is proposed by analysing elastic deformation and stress-strain. The objective of the proposed optimal method is to attain the maximum lift force in leg operation. The effectiveness and validity of design results are verified through simulations and lab experiments. A method for properly locating supporting legs on the robot body is proposed by analysing the influence of leg location to lift force and the relationship of supporting legs' with robot's roll-resistant capability. A layout scheme for the Water Dancer II-a prototype with ten supporting legs is presented with its operation successful designed.

A frame of modular design problems and requirements for underactuated mechanisms is discussed as related to robotic fingers. The proposed modular mechanism is connected sequentially by series units of underactuated mechanisms, which have the feature of passive self-adaptive in grasp operation and uniformizable in design procedure. The design considerations for modular underactuated mechanism are outlined. Optimality criteria are analyzed with the aim to formulate a general design algorithm. An example of a four-phalanx modular robotic finger is studied as an improvement of new version LARM Hand with the aim to show the practical feasibility for the proposed modular concepts and design methods.

A novel neural networks control structure for manipulators with flexible last link was proposed The manipulator with flexible last link was regarded as two parts: the flexible latter part (FLP) that includes the last joint and the last link, and the rigid former part (RFP), i.e. the rest. The kinematical and dynamic equations of the two parts were derived apart. The proposed control approach combines organically a conventional model-based controller with a neural networks controller. The neural networks approximate the dynamic anti-model of FLP and decompose the desired trajectory of end point into two parts, desired trajectory of axis of the last joint and desired outer corner of the last joint. Those would be realized by conventional controllers. By combining, the approach has not only solved the problem of nonlinearity and calculation inefficiency of the dynamic model of a flexible manipulator but also lessen greatly the size of the neural network and thus improve the convergence speed The proposed approach achieves fine control effect on the emulation with a three links planar manipulator with a flexible last link.