Search

Prompting has shown impressive success in enabling large pretrained language models (LMs) to perform diverse NLP tasks, especially when only few downstream data are available. Automatically finding the optimal prompt for each task, however, is challenging. Most existing work resorts to tuning soft prompt (e.g., embeddings) which falls short of interpretability, reusability across LMs, and applicability when gradients are not accessible. Discrete prompt, on the other hand, is difficult to optimize, and is often created by "enumeration (e.g., paraphrasing)-then-selection" heuristics that do not explore the prompt space systematically. This thesis will mainly introduce our proposed RLPrompt, an efficient discrete prompt optimization approach with reinforcement learning (RL). RLPrompt formulates a parameter-efficient policy network that generates the desired discrete prompt after training with reward. To overcome the complexity and stochasticity of reward signals by the large LM environment, we incorporate effective reward stabilization that substantially enhances the training efficiency. RLPrompt is flexibly applicable to different types of LMs, such as masked (e.g., BERT) and left-to-right models (e.g., GPTs), for both classification and generation tasks. Experiments on few-shot classification and unsupervised text style transfer show superior performance over a wide range of existing finetuning or prompting methods. Interestingly, the resulting optimized prompts are often ungrammatical gibberish text; and surprisingly, those gibberish prompts are transferrable between different LMs to retain significant performance, indicating LM prompting may not follow human language patterns.

Modeling object is one of the core problems in computer vision. A good object model can be applied to multiple visual recognition tasks. In this thesis, we use compositional relations to model the objects and parts for articulated objects, which is very challenging due to the large variability by subtypes, viewpoints and poses. Intuitively, objects are composed of parts, and each part is then composed of subparts. By recursively applying the compositional relations, we can obtain a hierarchical structure of the object, which is called compositional model. Compared to the popular black-box deep learning system, our model enjoys the advantages of explainability, and adaptivity to more complex scenarios. There are three challenges for the compositional model: 1) How to learn the parts and subparts, 2) How to learn the compositional structure between parts and subparts, 3) How to allow efficient learning and inference. In this thesis, we mainly focus on three relevant projects using the concept of compositions and parts.In the first project, we study the problem of semantic part segmentation for animals. Compositional model is used to represent boundaries of objects and parts. Given part-level boundary annotations, a mixture of models is learned to deal with various viewpoints and poses. We also incorporate edge, appearance, and semantic part cues into the compositional model. Besides, a linear complexity algorithm is developed for efficient inference. The evaluation using horse and cow images from PASCAL VOC demonstrates the effectiveness of our method.In the second project, we provide a novel unsupervised method to learn object semantic parts from internal states of CNNs. Our hypothesis is that semantic parts are represented by populations of neurons rather than by single filters. We propose a simple clustering technique to extract part representations, which we call visual concepts. We qualitatively show that visual concepts are semantically coherent in that they represent most semantic parts of the object, and visually coherent in that their corresponding image patches appear very similar. Next, we treat the visual concepts as part detectors and quantitatively evaluate their performance for detecting semantic parts. The experiments on the PASCAL3D+ dataset and our newly annotated VehicleSemanticPart dataset support that CNNs have internal representations of object parts encoded by populations of neurons.In the third project, we use the visual concepts for detecting the semantic parts on partially occluded objects. We consider a scenario where the model is trained using non-occluded images but tested on occluded images. The motivation is that there are infinite number of occlusion patterns in real world, so the models should be inherently robust and adaptive to occlusions instead of learning the occlusion patterns in the training data. Our approach uses a simple voting scheme, based on log-likelihood ratio tests and spatial constraints, to combine the evidence of local cues, which are derived from visual concepts. Experiments show that our algorithm outperforms several competitors in semantic part detection, especially in the presence of occlusions.

In recent years, synthetic aperture radar (SAR) has played an increasingly important role in the military and civil fields. Since the SAR image reflects the scattering characteristics of the target， it is of great significance to achieve multi-angle fusion of the target. However, there is a problem of angular loss in real SAR images. Through the electromagnetic simulation method, SAR images of 0–360° can be obtained, but the similarity to real images is low. Here, the authors combine electromagnetic simulation with conditional generative adversarial networks (cGANs). The image obtained by the electromagnetic simulation is taken as the input of the cGANs, and then the generator generates photorealistic SAR images containing the label information. Thereby, authors’ method complement the missing angles in the real SAR image dataset. Finally, they qualitatively and quantitatively evaluated the synthetic images generated through their model to verify the quality of the dataset.

As a result of the geometric constraints of the stairs, pedestrian movement features on stairways are different from level ambulation. Therefore, it is necessary for us to consider the stairs' three-dimensional geometry when we try to explore how the stair configuration affects pedestrian movement efficiency. Based on the pedestrian fundamental features from a previous pedestrian single-file movement experiment on stairs, we investigated pedestrian flow under various stair configurations with an improved agent-based model. Our simulated indicates that both stair inclinations and tread depth of stair step are sensitive to the pedestrian dynamic on stairs. Generally speaking, pedestrian flow decreases with the increasing slope of stairs. When the stairs slope increased 5, 10, 15 and 20 degree, the evacuation efficiency dropped 5.8%, 12.8%, 19.3% and 28.4%. Besides, the effect of the tread depth on pedestrian total evacuation time is not changed monotonously. Scenarios with 0.3m tread depth size present the optimal evacuation efficiency. Other scenarios differ from 3% to 27.9% in terms of the total evacuation time.

As an important power system protection device, the condition of metal oxide surge arresters (MOSAs) in service could be monitored, and online monitoring might be considered in a few critical installations. In this paper, a new resistive components measurement method termed as the remote nonsimultaneous resistive current measurement method (RNS) for MOSAs online condition monitoring is proposed. The RNS remotely and nonsimultaneously measures all types of resistive current parameters of MOSAs by employing the remote nonsimultaneous phase difference measurement method and noninteger harmonic analysis algorithm. The simulation and application results indicate that the proposed method accurately measures various parameters of MOSAs under complicated field circumstances.

Blockchain has received a lot of attention recently for its characteristics of decentralization, immutability, traceability, etc., making it a promising technology for the development of various applications, especially the management of various digital information. However, most current blockchain systems exhibit problems such as high computational overhead and centralization of power. Reliance on cryptocurrency in many public blockchain-based applications is another factor that has hindered the application of blockchain technology in areas other than the financial sectors. This paper proposes a new blockchain consensus mechanism based on the contributions of participants. The proposed consensus mechanism, which is called proof-of-contribution (PoC), quantifies user behaviors and actions in a blockchain-based application as contribution values calculated through an algorithm. The node that has the highest contribution value in each round of consensus gets the right to generate the next new block. PoC preserves the properties of decentralization and resistance to hard fork and does not rely on cryptocurrency, making it attractive over cryptocurrency-based consensus mechanisms like proof-of-work (PoW) for a wide variety of applications that do not have to involve cryptocurrency. Contribution values can be abstracted from applications and used in the underlying blockchain consensus process to improve the security and trustworthiness of the applications. Intellectual property (IP) protection is one such application to apply the blockchain technology and the PoC consensus mechanism. However, existing blockchain-based IP protection systems are mostly developed based on public blockchain platforms such as the Ethereum while some others are developed as consortium blockchains, thus exhibiting inherent defects such as high system requirements, long consensus delay, and low user participation. In this paper, we will use IP protection as the application scenario to illustrate the development of our PoC consensus mechanism and will compare PoC to some existing consensus mechanisms. Experimental results will show that the proposed PoC consensus mechanism preserves most of the main security characteristics of blockchain and is superior to existing consensus mechanisms, making it more secure and efficient to use blockchain technology for digital information management.

Research on advertisement has mainly focused on how to accurately predict the click-through rate (CTR). Much less is known about fraud detection and malicious behavior defense. Previous studies usually use statistics, design threshold and manually make strategies, which cannot find potential fraud behavior effectively and suffer from new attacks. In this paper, we make the first step to understand the type of malicious activities on large-scale online advertising platforms. By analyzing each feature comprehensively, we propose a novel coding approach to transform nominal attributes into numeric while maintaining the most effective information of the original data for fraud detection. Next, we code important features such as IP and cookie in our dataset and train machine learning methods to detect fraud traffic automatically. Experimental results on real datasets demonstrate that the proposed fraud detection method performs well considering both the accuracy and efficiency. Finally, we conclude how to design a defense system by considering which methods could be used for the anti-spam gaming in the future.

Physical Distribution is a bottleneck in E-commerce and reverse logistics has become a bottleneck restricting the development of e-commerce. This paper constructs the distribution system in E-commerce logistics based on gridding management, via the comparison analysis between grid and non-grid distribution system in E-commerce logistics through computer simulation, verifies the grid used in e-commerce distribution system is correct and sets up a service quality evaluation system. Comprehensive simulations and comparisons show that the gridding model is practically optimal and theoretically promising.

In this paper, we present randomized, distributed algorithms that allow n >- 1 nodes to solve agreement problems in a single-hop, time-slotted wireless sensor network (WSN). In particular, we consider the problem of consensus - where all nodes must come to agreement on any single value - and the problem of leader election - where all nodes must agree on a single node to act as a leader. Both consensus and leader election are well-studied challenges whose solutions provide critical building blocks for protocols in distributed systems. Our results extend this study to the domain of WSNs in the presence of (possible malicious) communication failures. Our algorithms are provably robust and scalable in the face of an adversary who interferes with (or jams) communications between the network devices (nodes). In WSNs, the communication costs incurred by nodes is a major concern because such devices are typically powered by on-board batteries with limited resources, and the actions of sending and listening on the communication channel dominates operational costs. These costs are modeled as a unit cost for sending or listening to a message in a single time slot, while the cost for staying silent in a time slot is zero. Similarly, the adversary incurs unit cost to jam a single slot as this equates to sending on the channel. Consider an adversary that incurs a cost of T by pursuing an arbitrary, worst-case jamming strategy. We demonstrate an algorithm that solves consensus, and another algorithm that solves leader election, with a per-node expected cost of O( p T=n lg2:5 n+ lg5 n). This implies that the adversary must incur a substantially large asymptotic cost in pursuing any jamming strategy. In the energy-constrained domain of WSNs, this result is particularly valuable as the adversary will rapidly deplete her resources in attempting to thwart consensus or leader election.

In order to filter out all kinds of interference noises in batching control and improve the precision of the ingredients, a morphology filtering method was used in data processing which is the combination of morphological open-closing and close-opening. When the morphology filter is used to de-noise the interference noise, it doesn't need to consider the dynamic signal spectrum characteristics. After data processing and application in practical production, it is proved that the random noise, pulse noise and mixture noise can be removed effectively. The experiment shows that the simple and realized easily algorithm can be used in facilitate industrial fields and real-time signal processing.

High precision acquisition, tracking, pointing (ATP) system is one of the key techniques of laser communication. The spot-detecting camera is used to detect the direction of beacon in laser communication link, so that it can get the position information of communication terminal for ATP system. The positioning accuracy of camera decides the capability of laser communication system directly. So the spot-detecting camera in satellite-to-earth laser communication ATP systems needs high precision on target detection. The positioning accuracy of cameras should be better than ±1μ rad . The spot-detecting cameras usually adopt centroid algorithm to get the position information of light spot on detectors. When the intensity of beacon is moderate, calculation results of centroid algorithm will be precise. But the intensity of beacon changes greatly during communication for distance, atmospheric scintillation, weather etc. The output signal of detector will be insufficient when the camera underexposes to beacon because of low light intensity. On the other hand, the output signal of detector will be saturated when the camera overexposes to beacon because of high light intensity. The calculation accuracy of centroid algorithm becomes worse if the spot-detecting camera underexposes or overexposes, and then the positioning accuracy of camera will be reduced obviously. In order to improve the accuracy, space-based cameras should regulate exposure time in real time according to light intensity. The algorithm of adaptive-exposure technique for spot-detecting camera based on metal-oxide-semiconductor (CMOS) detector is analyzed. According to analytic results, a CMOS camera in space-based laser communication system is described, which utilizes the algorithm of adaptive-exposure to adapting exposure time. Test results from imaging experiment system formed verify the design. Experimental results prove that this design can restrain the reduction of positioning accuracy for the change of light intensity. So the camera can keep stable and high positioning accuracy during communication.

Traditional algorithms for semantic similarity computation fall into three categories: distance-based, feature contrast and information-based methods. The former two methods ignore the objective statistics, while the last one cannot obtain enough domain data. In this paper, a new method for similarity computation based on Bayesian Estimation is proposed. First, the concept encountering probability is assumed to be a random variable with a priori Beta distribution. Second, its priori parameters are designated by the distance-based similarity algorithm. And the posteriori encountering probability is calculated based on Bayesian Estimation. Thereby, the semantic similarity integrating the subjective experience with the objective statistic is acquired based on information-based method. Finally, our method is implemented and the performance is analyzed by WordNet.

Merits and defects of note-based video retrieval and content-based video retrieval are investigated in this paper, and complementarity of two aspects each other are also discussed. A new theory of semantic processing in content-based video retrieval is explored. An object-oriented representation modeling of the video semantics, based on the multi-layers, is presented. And on this basis, a semantic extraction frame based on the multi-semantics is constructed.