Real-world point clouds usually have inconsistent orientations and often suffer from data missing issues. To solve this problem, we design a neural network, CF-Net, to address challenges in rotation invariant completion. In our network, we modify and integrate complementary operators to extract features that are robust against rotation and incompleteness. Our CF-Net can achieve competitive results both geometrically and semantically as demonstrated in this paper.

The Zuker algorithm is a widely-adopted method for rapid prediction of RNA secondary structures. In this work, we present a new hardware architecture based on the algorithm and implement its accelerator using TSMC 40-nm technology. The proposed architecture is capable of handling more types of dangling bases when compared to the previous hardware work. Besides, with better data re-use, the idling time and memory usage of our hardware modules can be effectively reduced, and thus a higher processing speed can be achieved. Compared to the well-known software approach, RNAstructure, our 16-processing-element design has an $11\times$ speedup for RNA sequences around 3000 bases.

MOTIVATION: MinION, a third-generation sequencer from Oxford Nanopore Technologies, is a portable device that can provide long-nucleotide read data in real-time. It primarily aims to deduce the makeup of nucleotide sequences from the ionic current signals generated when passing DNA/RNA fragments through nanopores charged with a voltage difference. To determine nucleotides from measured signals, a translation process known as basecalling is required. However, compared to NGS basecallers, the calling accuracy of MinION still needs to be improved. RESULTS: In this work, a simple but powerful neural network architecture called multi-scale recurrent caller (MSRCall) is proposed. MSRCall comprises a multi-scale structure, recurrent layers, a fusion block and a connectionist temporal classification decoder. To better identify both short-and long-range dependencies, the recurrent layer is redesigned to capture various time-scale features with a multi-scale structure. The results show that MSRCall outperforms other basecallers in terms of both read and consensus accuracies. AVAILABILITY AND IMPLEMENTATION: MSRCall is available at: https://github.com/d05943006/MSRCall. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Traditional High Dynamic Range (HDR) techniques require inputs of multiple Low Dynamic Range (LDR) images with different exposure time taken in serial shots or from a camera array. These methods are either suffered from ghosting effects or very expensive when setting up the system. In this paper, we design a hand-held pinhole-mask-based light field camera and use it to capture all LDR data in one shot. We also propose a light-field HDR algorithm to process LDR data collected. Finally, a hardware accelerator for the algorithm is implemented to complete calculations within 0.3 seconds.

This paper addresses three issues in integrating part-based representations into convolutional neural networks (CNNs) for object recognition. First, most part-based models rely on a few pre-specified object parts. However, the optimal object parts for recognition often vary from category to category. Second, acquiring training data with part-level annotation is labor-intensive. Third, modeling spatial relationships between parts in CNNs often involves an exhaustive search of part templates over multiple network streams. We tackle the three issues by introducing a new network layer, called co-occurrence layer. It can extend a convolutional layer to encode the co-occurrence between the visual parts detected by the numerous neurons, instead of a few pre-specified parts. To this end, the feature maps serve as both filters and images, and mutual correlation filtering is conducted between them. The co-occurrence layer is end-to-end trainable. The resultant co-occurrence features are rotation-and translation-invariant, and are robust to object deformation. By applying this new layer to the VGG-16 and ResNet-152, we achieve the recognition rates of 83.6% and 85.8% on the Caltech-UCSD bird benchmark, respectively. The source code is available at https://github.com/yafangshih/Deep-COOC.