Exploration of Effective Attention Strategies for Neural Automatic Post-editing with Transformer

Automatic post-editing (APE) is the study of correcting translation errors in the output of an unknown machine translation (MT) system and has been considered as a method of improving translation quality without any modification to conventional MT systems. Recently, several variants of Transformer that take both the MT output and its corresponding source sentence as inputs have been proposed for APE; and models introducing an additional attention layer into the encoder to jointly encode the MT output with its source sentence recorded a high-rank in the WMT19 APE shared task. We examine the effectiveness of such joint-encoding strategy in a controlled environment and compare four types of decoder multi-source attention strategies that have been introduced into previous APE models. The experimental results indicate that the joint-encoding strategy is effective and that taking the final encoded representation of the source sentence is the more proper strategy than taking such representation within the same encoder stack. Furthermore, among the multi-source attention strategies combined with the joint-encoding, the strategy that applies attention to the concatenated input representation and the strategy that adds up the individual attention to each input improve the quality of APE results over the strategy using the joint-encoding only.

Two-color super-resolution localization microscopy via joint encoding of emitter location and color

Multi-color super-resolution localization microscopy (SRLM) provides great opportunities for studying the structural and functional details of biological samples. However, current multi-color SRLM methods either suffer from medium to high crosstalk, or require a dedicated optical system and a complicated image analysis procedure. To address these problems, here we propose a completely different method to realize multi-color SRLM. This method is built upon a customized RGBW camera with a repeated pattern of filtered (Red, Green, Blue and Near-infrared) and unfiltered (White) pixels. With a new insight that RGBW camera is advantageous for color recognition instead of color reproduction, we developed a joint encoding scheme of emitter location and color. By combing this RGBW camera with the joint encoding scheme and a simple optical set-up, we demonstrated two-color SRLM with ∼20 nm resolution and < 2% crosstalk (which is comparable to the best reported values). This study significantly reduces the complexity of two-color SRLM (and potentially multi-color SRLM), and thus offers good opportunities for general biomedical research laboratories to use multi-color SRLM, which is currently mastered only by well-trained researchers.

SECURITY TECHNIQUE BASED ON WATERMARKING AND CHAOTIC ENCRYPTION FOR DIGITAL IMAGE

With the widespread usage of the Internet, security takes precedence above anything else when transmitting data. This research proposed a hybrid encoding approach with watermark embedding that provides excellent security. Our novel strategy is based on cryptography, which allows it to recognize higher secrecy and efficiency. In this research, a hybrid method is used to support visual watermarking and cryptography to embed vital data. Our effort aims to improve the security of hidden facts embedded in the cover picture. The embedding and encryption of the watermark image into the cover image document were performed in two steps. To begin, two keys and an XOR bit operation were used to generate a large number of distinct keys for encryption. Second, a modified approach of the least significant bit (LSB) technique was adopted to hide a high-resolution watermark picture in the cover picture. The suggested second stage involved encrypting the cover image using the asymmetric key cryptography method (RSA), which provides additional secrecy during picture transmission. The original picture, the watermark photo, can be recovered and decoded using the permitted techniques. As a result, encoding and watermarking may be combined, giving the term "joint encoding and watermarking" legitimacy. Peak Signal to Noise Ratio (PSNR) and relationship aspects are greater with this method. For the implementation of hiding and encoding for the watermark and encoding the digital cover picture, MATLAB-GUI software was utilized. Experimental results demonstrated a good performance with a good correlation for all encrypted images and very high PSNR of the Stego images.

Shape-invariant perceptual encoding of dynamic facial expressions across species

Dynamic facial expressions are crucial for communication in primates. Due to the difficulty to control shape and dynamics of facial expressions across species, it is unknown how species-specific facial expressions are perceptually encoded and interact with the representation of facial shape. While popular neural-network theories predict a joint encoding of facial shape and dynamics, the neuromuscular control of faces evolved more slowly than facial shape, suggesting a separate encoding. To investigate this hypothesis, we developed photo-realistic human and monkey heads that were animated with motion-capture data from monkeys and human. Exact control of expression dynamics was accomplished by a Bayesian machine-learning technique. Consistent with our hypothesis, we found that human observers learned cross-species expressions very quickly, where face dynamics was represented independently of facial shape. This result supports the co-evolution of the visual processing and motor-control of facial expressions, while it challenges popular neural-network theories of dynamic expression-recognition.

Neural Entity Summarization with Joint Encoding and Weak Supervision

In a large-scale knowledge graph (KG), an entity is often described by a large number of triple-structured facts. Many applications require abridged versions of entity descriptions, called entity summaries. Existing solutions to entity summarization are mainly unsupervised. In this paper, we present a supervised approach NEST that is based on our novel neural model to jointly encode graph structure and text in KGs and generate high-quality diversified summaries. Since it is costly to obtain manually labeled summaries for training, our supervision is weak as we train with programmatically labeled data which may contain noise but is free of manual work. Evaluation results show that our approach significantly outperforms the state of the art on two public benchmarks.