Search

Contrastive loss has been increasingly used in learning representations from multiple modalities. In the limit, the nature of the contrastive loss encourages modalities to exactly match each other in the latent space. Yet it remains an open question how the modality alignment affects the downstream task performance. In this paper, based on an information-theoretic argument, we first prove that exact modality alignment is sub-optimal in general for downstream prediction tasks. Hence we advocate that the key of better performance lies in meaningful latent modality structures instead of perfect modality alignment. To this end, we propose three general approaches to construct latent modality structures. Specifically, we design 1) a deep feature separation loss for intra-modality regularization; 2) a Brownian-bridge loss for inter-modality regularization; and 3) a geometric consistency loss for both intra- and inter-modality regularization. Extensive experiments are conducted on two popular multi-modal representation learning frameworks: the CLIP-based two-tower model and the ALBEF-based fusion model. We test our model on a variety of tasks including zero/few-shot image classification, image-text retrieval, visual question answering, visual reasoning, and visual entailment. Our method achieves consistent improvements over existing methods, demonstrating the effectiveness and generalizability of our proposed approach on latent modality structure regularization., Comment: 14 pages, 8 figure, CVPR 2023 accepted

Aligning signals from different modalities is an important step in vision-language representation learning as it affects the performance of later stages such as cross-modality fusion. Since image and text typically reside in different regions of the feature space, directly aligning them at instance level is challenging especially when features are still evolving during training. In this paper, we propose to align at a higher and more stable level using cluster representation. Specifically, we treat image and text as two "views" of the same entity, and encode them into a joint vision-language coding space spanned by a dictionary of cluster centers (codebook). We contrast positive and negative samples via their cluster assignments while simultaneously optimizing the cluster centers. To further smooth out the learning process, we adopt a teacher-student distillation paradigm, where the momentum teacher of one view guides the student learning of the other. We evaluated our approach on common vision language benchmarks and obtain new SoTA on zero-shot cross modality retrieval while being competitive on various other transfer tasks., Comment: Accepted by CVPR 2022

Although a linear vibration absorber (LVA) or nonlinear energy sink (NES) can effectively mitigate the vibration of the main system in harmonic excitation, the amplitude of the absorber can be very large. Using the single-objective differential evolutionary (DE) algorithm, this paper pioneers the global control of the main system and a bistable NES to achieve decent vibration mitigation effects and decrease the global response of the system. Using the multi-objective DE algorithms and comparing the optimization results of the LVA, this paper proposes multi-objective and multi-parameter design criteria of a bistable NES. The mass ratio, the displacement amplitude, and the mechanical energy are optimization goals. The results show that the maximum amplitude of the main system and the absorber can be controlled at the same level, and the global control strategy does not change the resonance frequency of the main system. Compared with the LVA, the bistable NES has similar vibration mitigation effects with the variation of the mass ratio in the multi-objective optimization. However, the bistable NES can achieve better control over a larger spring stiffness range. Therefore, through single- and multi-objective optimization design, this paper proves the superiority of the bistable NES in the vibration mitigation. Meanwhile, this paper provides an optimization design method for global vibration control.

Ni layer is widely utilized in Bi2Te3-based thermoelectric generators (TEGs) as a diffusion barrier layer. However, its influences on the contact resistance of the solder joints and the power generation properties of the TEGs are still unclear, which are of great importance in practical applications. In this study, Bi2Te3-based thermoelectric (TE) modules with Ni (NiTM) and without Ni (TM) layer were fabricated by using Sn96.5Ag3Cu0.5 (SAC305) solder. Experimental results showed that Ni layer could effectively reduce 32% of contact resistances in P-type and N-type solder joints. P-type solder joint with Ni avoided voids forming on the border of the Bi0.5Sb1.5Te3 and SAC305 by suppressing severe Te and Sb elements diffusion. In N-type solder joint, since Ni inhibited the formation of Bi-Te self-diffusion barrier layer, no cracks generated at the interface between Bi2Te2.5Se0.5 and SAC305. In terms of power generation abilities, the inherently superior performance of TE legs was maintained due to effective suppression of element diffusion. When the temperature difference was 161 °C, the open circuit voltage of NiTM exhibited an enhancement of 13% compared with that of TM. The lower contact resistance and higher open circuit voltage resulted in a much higher maximum output power (an enhancement of 90%) in NiTM.

Adaptor signatures, also known as scriptless scripts, have recently become an important tool in addressing the scalability and interoperability issues of blockchain applications such as cryptocurrencies. An adaptor signature extends a digital signature in a way that a complete signature reveals a secret based on a cryptographic condition. It brings about various advantages such as (i) low on-chain cost, (ii) improved fungibility of transactions, and (iii) advanced functionality beyond the limitation of the blockchain’s scripting language. In this work, we introduce the first post-quantum adaptor signature, named $${\mathsf {LAS}}$$. Our construction relies on the standard lattice assumptions, namely Module-SIS and Module-LWE. There are certain challenges specific to the lattice setting, arising mainly from the so-called knowledge gap in lattice-based proof systems, that makes the realization of an adaptor signature and its applications difficult. We show how to overcome these technical difficulties without introducing additional on-chain costs. Our evaluation demonstrates that $${\mathsf {LAS}}$$ is essentially as efficient as an ordinary lattice-based signature in terms of both communication and computation. We further show how to achieve post-quantum atomic swaps and payment channel networks using $${\mathsf {LAS}}$$.

Deep neural networks excel at comprehending complex visual signals, delivering on par or even superior performance to that of human experts. However, ad-hoc visual explanations of model decisions often reveal an alarming level of reliance on exploiting non-causal visual cues that strongly correlate with the target label in training data. As such, deep neural nets suffer compromised generalization to novel inputs collected from different sources, and the reverse engineering of their decision rules offers limited interpretability. To overcome these limitations, we present a novel contrastive learning strategy called {\it Proactive Pseudo-Intervention} (PPI) that leverages proactive interventions to guard against image features with no causal relevance. We also devise a novel causally informed salience mapping module to identify key image pixels to intervene, and show it greatly facilitates model interpretability. To demonstrate the utility of our proposals, we benchmark on both standard natural images and challenging medical image datasets. PPI-enhanced models consistently deliver superior performance relative to competing solutions, especially on out-of-domain predictions and data integration from heterogeneous sources. Further, our causally trained saliency maps are more succinct and meaningful relative to their non-causal counterparts.

Evolution of microstructure and electrical property in the conversion of high strength carbon fiber (HSCF) to high modulus carbon fiber (HMCF) and ultrahigh modulus carbon fiber (UHMCF) was investigated. Longitudinal grooves on fiber surfaces became less well-defined during high temperature graphitization. The tensile modulus of carbon fibers was affected by fiber crystalline structure and it increased with decreases in the value of interlayer spacing and improvements in the value of crystallite thickness. Increases in the crystallite size almost had little effect on the tensile strength. However, a lower interlayer spacing and a higher preferred orientation could result in a higher tensile strength. The crystal structure of carbon fibers became much more ordered during high temperature graphitization. It was found that the electrical resistivity gradually decreased from 14.69 × 10−4 Ω·cm to 9.70 × 10−4 Ω·cm and 8.80 × 10−4 Ω·cm, respectively, in the conversion of HSCF to HMCF and UHMCF.

This paper presents a high-speed data transmission plan based on VPX bus. The aim is to establish a high-speed data transmission structure which could be used for dealing with massive echo data from the PAR (Phased array radar) in the future. The frequency is up to gigahertz so that it is hard to design the PCB which is used for high-speed serial data transmission between the boards. Stable and reliable interconnection between payload boards is a main problem. This paper uses channel simulation method to design the hardware. A channel simulation model is established between high-speed backplane and payload modules. This design extracts key information from channel tub curve and eye diagram, then design the hardware based on channel capacity and bit error rate. This paper designs a dual-channel highspeed serial RapidIO annular communication between payload modules based on VPX bus. Single channel theory transmission rate is 3.125Gbps. And the result shows that the actual speed of dual-channel is up to 540.70MB/s. It’s nearly the theory speed of the RapidIO communication protocol.

We propose a Leaked Motion Video Predictor (LMVP) to predict future frames by capturing the spatial and temporal dependencies from given inputs. The motion is modeled by a newly proposed component, motion guider, which plays the role of both learner and teacher. Specifically, it {\em learns} the temporal features from real data and {\em guides} the generator to predict future frames. The spatial consistency in video is modeled by an adaptive filtering network. To further ensure the spatio-temporal consistency of the prediction, a discriminator is also adopted to distinguish the real and generated frames. Further, the discriminator leaks information to the motion guider and the generator to help the learning of motion. The proposed LMVP can effectively learn the static and temporal features in videos without the need for human labeling. Experiments on synthetic and real data demonstrate that LMVP can yield state-of-the-art results.

The performance of many network learning applications crucially hinges on the success of network embedding algorithms, which aim to encode rich network information into low-dimensional vertex-based vector representations. This paper considers a novel variational formulation of network embeddings, with special focus on textual networks. Different from most existing methods that optimize a discriminative objective, we introduce Variational Homophilic Embedding (VHE), a fully generative model that learns network embeddings by modeling the semantic (textual) information with a variational autoencoder, while accounting for the structural (topology) information through a novel homophilic prior design. Homophilic vertex embeddings encourage similar embedding vectors for related (connected) vertices. The proposed VHE promises better generalization for downstream tasks, robustness to incomplete observations, and the ability to generalize to unseen vertices. Extensive experiments on real-world networks, for multiple tasks, demonstrate that the proposed method consistently achieves superior performance relative to competing state-of-the-art approaches., Comment: Accepted to NeurIPS 2019

There has been recent interest in developing scalable Bayesian sampling methods such as stochastic gradient MCMC (SG-MCMC) and Stein variational gradient descent (SVGD) for big-data analysis. A standard SG-MCMC algorithm simulates samples from a discrete-time Markov chain to approximate a target distribution, thus samples could be highly correlated, an undesired property for SG-MCMC. In contrary, SVGD directly optimizes a set of particles to approximate a target distribution, and thus is able to obtain good approximations with relatively much fewer samples. In this paper, we propose a principle particle-optimization framework based on Wasserstein gradient flows to unify SG-MCMC and SVGD, and to allow new algorithms to be developed. Our framework interprets SG-MCMC as particle optimization on the space of probability measures, revealing a strong connection between SG-MCMC and SVGD. The key component of our framework is several particle-approximate techniques to efficiently solve the original partial differential equations on the space of probability measures. Extensive experiments on both synthetic data and deep neural networks demonstrate the effectiveness and efficiency of our framework for scalable Bayesian sampling., Comment: UAI 2018

A new form of the variational autoencoder (VAE) is proposed, based on the symmetric Kullback-Leibler divergence. It is demonstrated that learning of the resulting symmetric VAE (sVAE) has close connections to previously developed adversarial-learning methods. This relationship helps unify the previously distinct techniques of VAE and adversarially learning, and provides insights that allow us to ameliorate shortcomings with some previously developed adversarial methods. In addition to an analysis that motivates and explains the sVAE, an extensive set of experiments validate the utility of the approach.

Fully homomorphic encryption (FHE) allows one to perform arbitrary computation on encrypted data. From the cloud computing perspective, a FHE scheme can be used to outsource a computation to a remote cloud server without revealing the computed data to the server. This is a very attractive cryptographic primitive. Existing methods used to construct FHE schemes are complicated, and often these methods are difficult to understand except by experts in the field. This paper aims to explain how to construct a FHE scheme in an easily understandable way. We first present a general overview of FHE scheme construction. We then take the van Dijk et al FHE scheme as an example and explain how a FHE scheme is obtained from a simple integer-based encryption scheme. Our explanation gives details of the noise bound, how noise growth can be controlled and how the vanDijk scheme is implemented.

Contains fulltext : 72073.pdf (author's version ) (Open Access) 4th International Conference, ISPEC 2008, 21 april 2008

Polyacrylonitrile-based carbon fibers were electrochemically oxidized under a wide variety of current densities, and then surface changes in the morphological structures and the chemical compositions were characterized by scanning electron microscopy and X-ray photoelectron spectroscopy. Finally, the effects of carbon fiber surface chemistry on the interfacial properties of carbon fiber reinforced composites were researched. The results showed that surface contaminants generated during the production process were removed and the relative contents of oxygen and nitrogen elements increased by 322% and 432%, respectively, after the electrochemical oxidation. Increases in the O/C and N/C ratios resulted in the increase in the interlaminar shear strength values of composites. However, much higher O/C and N/C ratios led to decline in the interlaminar shear strength values. Carboxyl groups were responsible for strong interactions between carbon fiber surfaces and the epoxy resin. When the current densities were more than 5 A/m2, the interfacial bonding between carbon fibers and the epoxy resin became weak and the interlaminar shear strength values also decreased.

Effect of low current density electrochemical oxidation on the properties of carbon fiber-reinforced epoxy resin composites

An announcement scheme is a system that facilitates vehicles to broadcast road-related information in vehicular ad hoc networks (VANETs) in order to improve road safety and efficiency. Here we propose a new cryptographic primitive for public updating of reputation score based on the Boneh-Boyen-Shacham short group signature scheme. This allows private reputation score retrieval without a secure channel. Using this we devise a privacy-aware announcement scheme using reputation systems which is reliable, auditable and robust., Comment: This paper is a preprint of a paper submitted to IET Information Security and is subject to Institution of Engineering and Technology Copyright. If accepted, the copy of record will be available at IET Digital Library

Controlling chaos in the Lorenz system with a controllable Rayleigh number is investigated by the state space exact linearization method. Based on proving the exact linearizability, the nonlinear feedback is utilized to design the transformation changing the original chaotic system into a linear controllable one so that the control is realized. Numerical examples of control are presented.

Cross-domain alignment between two sets of entities (e.g., objects in an image, words in a sentence) is fundamental to both computer vision and natural language processing. Existing methods mainly focus on designing advanced attention mechanisms to simulate soft alignment, with no training signals to explicitly encourage alignment. The learned attention matrices are also dense and lacks interpretability. We propose Graph Optimal Transport (GOT), a principled framework that germinates from recent advances in Optimal Transport (OT). In GOT, cross-domain alignment is formulated as a graph matching problem, by representing entities into a dynamically-constructed graph. Two types of OT distances are considered: (i) Wasserstein distance (WD) for node (entity) matching; and (ii) Gromov-Wasserstein distance (GWD) for edge (structure) matching. Both WD and GWD can be incorporated into existing neural network models, effectively acting as a drop-in regularizer. The inferred transport plan also yields sparse and self-normalized alignment, enhancing the interpretability of the learned model. Experiments show consistent outperformance of GOT over baselines across a wide range of tasks, including image-text retrieval, visual question answering, image captioning, machine translation, and text summarization.

In this paper, we introduce Attribute-Based Signatures with User-Controlled Linkability (ABS-UCL). Attribute-based signatures allow a signer who has enough credentials/attributes to anonymously sign a message w.r.t. some public policy revealing neither the attributes used nor his identity. User-controlled linkability is a new feature which allows a user to make some of his signatures directed at the same recipient linkable while still retaining anonymity. Such a feature is useful for many real-life applications. We give a general framework for constructing ABS-UCL and present an efficient instantiation of the construction that supports multiple attribute authorities.

Deployment of DNSSEC, although increasing, still suffers from many practical issues that results in a false sense of security. While many domains outsource zone management, they also have to outsource DNSSEC key management to the DNS operator, making the operator an attractive target for attackers. Moreover, DNSSEC does not provide any sort of protection in the case the operator itself decides to serve false information, for example, if it gets compromised.In this work, we show how to use techniques from threshold ECDSA: (1) to protect keys such that domains do not reveal their signing keys to a DNS operator, and (2) to protect the operational integrity of DNS operator. As a result of being highly specialized, prior work on threshold ECDSA has focused on a limited set of threat models, and none have so far considered techniques to amortize signature generation. Our work takes a different approach and presents a generic technique for obtaining a threshold ECDSA protocol from any secure multiparty computation protocol that works over an appropriate finite field. We show how this technique lends itself to very efficient threshold signing protocols by comparing it against state-of-the-art protocols from both academia and industry. For similar threat models, our protocols are as fast as the previous best protocol in terms of signing, and up to an order of magnitude faster for key generation on a fast network. Finally, we show how to integrate our application into a widely used DNS management software and demonstrate through experiments the overhead compared to traditional DNSSECs.