In collider experiments, the kinematic reconstruction of heavy, short-lived particles is vital for precision tests Standard Model and in searches physics beyond it. Performing events with many final-state jets, such as all-hadronic decay top-antitop quark pairs, challenging. We present (HyPER), a novel architecture based on graph neural networks that uses hypergraph representation learning to build more powerful efficient representations events. HyPER used reconstruct parent from sets...

Low-rank adaptation is a popular parameter-efficient fine-tuning method for large language models. In this paper, we analyze the impact of low-rank updating, as implemented in LoRA. Our findings suggest that updating mechanism may limit ability LLMs to effectively learn and memorize new knowledge. Inspired by observation, propose called MoRA, which employs square matrix achieve high-rank while maintaining same number trainable parameters. To it, introduce corresponding non-parameter...

Personalized conversation models (PCMs) generate responses according to speaker preferences. Existing personalized tasks typically require extract preferences from user descriptions or their histories, which are scarce for newcomers and inactive users. In this paper, we propose a few-shot task with an auxiliary social network. The requires given few conversations the methods mainly designed incorporate histories. Those can hardly model speakers so connections between speakers. To better...

In ICASSP 2023 speech signal improvement challenge, we developed a dual-stage neural model which improves quality induced by different distortions in stage-wise divide-and-conquer fashion. Specifically, the first stage, network focuses on recovering missing components of spectrum, while second our aims to further suppress noise, reverberation, and artifacts introduced first-stage model. Achieving 0.446 final score 0.517 P.835 score, system ranks 4th non-real-time track.

Ziheng Li, Shaohan Huang, Zihan Zhang, Zhi-Hong Deng, Qiang Lou, Haizhen Jian Jiao, Furu Wei, Weiwei Qi Zhang. Proceedings of the 61st Annual Meeting Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Accurate and efficient measurements for evaluating the similarity between mathematical formulae play an important role in information retrieval. Most previous studies have focused on representing different types to catch their features combining traditional structure matching algorithms. This paper presents a new unsupervised model called N-ary Tree-based Formula Embedding Model (NTFEM) task of evaluation. Using n-ary tree represent formula, we convert formula into linear sequence that can...

The real-world data distribution is essentially long-tailed, which poses great challenge to the deep model. In this work, we propose a new method, Gradual Balanced Loss and Adaptive Feature Generator (GLAG) alleviate imbalance. GLAG first learns balanced robust feature model with Loss, then fixes augments under-represented tail classes on level knowledge from well-represented head classes. And generated samples are mixed up real training during epochs. general loss it can combine different...

The ROOT I/O (RIO) subsystem is foundational to most HEP experiments - it provides a file format, set of APIs/semantics, and reference implementation in C++. It often found at the base an experiment's framework used serialize data; case LHC experiment, this may be hundreds petabytes files! Individual physicists will further use RIO perform their end-stage analysis, reading from intermediate files they generate experiment data.

Packet loss is a common and unavoidable problem in voice over internet phone (VoIP) systems. To deal with the problem, we propose band-split packet concealment network (BS-PLCNet). Specifically, split full-band signal into wide-band (0-8kHz) high-band (8-24kHz). The signals are processed by gated convolutional recurrent (GCRN), while counterpart simple GRU network. ensure high speech quality automatic recognition (ASR) compatibility, multi-task learning (MTL) framework including fundamental...

Language models (LMs) have recently shown superior performances in various speech generation tasks, demonstrating their powerful ability for semantic context modeling. Given the intrinsic similarity between and enhancement, harnessing information is advantageous enhancement tasks. In light of this, we propose SELM, a novel paradigm that integrates discrete tokens leverages language models. SELM comprises three stages: encoding, modeling, decoding. We transform continuous waveform signals...

In collider experiments, the kinematic reconstruction of heavy, short-lived particles is vital for precision tests Standard Model and in searches physics beyond it. Performing events with many final-state jets, such as all-hadronic decay topantitop quark pairs, challenging. We present HyPER, a graph neural network that uses blended graph-hypergraph representation learning to reconstruct parent from sets objects. HyPER tested on simulation shown perform favorably when compared existing...

Multimodal large language models (MLLMs) have shown promising advancements in general visual and understanding. However, the representation of multimodal information using MLLMs remains largely unexplored. In this work, we introduce a new framework, E5-V, designed to adapt for achieving universal embeddings. Our findings highlight significant potential representing inputs compared previous approaches. By leveraging with prompts, E5-V effectively bridges modality gap between different types...

Multi-distribution learning (MDL), which seeks to learn a shared model that minimizes the worst-case risk across $k$ distinct data distributions, has emerged as unified framework in response evolving demand for robustness, fairness, multi-group collaboration, etc. Achieving data-efficient MDL necessitates adaptive sampling, also called on-demand throughout process. However, there exist substantial gaps between state-of-the-art upper and lower bounds on optimal sample complexity. Focusing...

Graph neural networks have been widely used in recent recommender systems, where negative sampling plays an important role. Existing methods restrict the relationship between nodes as either hard positive pairs or pairs. This leads to loss of structural information, and lacks mechanism generate for with few neighbors. To overcome limitations, we propose a novel soft link-based method, namely MixDec Sampling, which consists Mixup Sampling module Decay module. The augments node features by...

When processing large amounts of data, the rate at which reading and writing can take place is a critical factor. High energy physics data relying on ROOT no exception. The recent parallelisation LHC experiments' software frameworks analysis ever increasing amount collision collected by experiments further emphasised this issue underlying need implicit parallelism expressed within I/O.

Abstract The research on upgrading and modification of the existing railway signal system found that line basic datasheet data engineering cannot satisfy application requirement, it is necessary to study optimization line. This paper presents a method for generation process verification. Based analysis update requirements, hierarchical block storage model established by expanding datasheet, design process. formal verification combining UML with NuSMV model, which verifies activity, certainty...

In ICASSP 2023 speech signal improvement challenge, we developed a dual-stage neural model which improves quality induced by different distortions in stage-wise divide-and-conquer fashion. Specifically, the first stage, network focuses on recovering missing components of spectrum, while second our aims to further suppress noise, reverberation, and artifacts introduced first-stage model. Achieving 0.446 final score 0.517 P.835 score, system ranks 4th non-real-time track.

The aim of this paper is to solve the some specific integrals such as , and . traditional method that people normally decompose polynomial into several partial fractions first. This process involves adding it all up, expanding brackets, doing matrices computation, which takes too many steps calculation. fraction part requires using Euler’s formula large amounts brackets prove multiplication those denominators equal denominator given in original equation. Once making one little mistake, next...

Recent studies have shown that dual encoder models trained with the sentence-level translation ranking task are effective methods for cross-lingual sentence embedding. However, our research indicates token-level alignment is also crucial in multilingual scenarios, which has not been fully explored previously. Based on findings, we propose a dual-alignment pre-training (DAP) framework embedding incorporates both and alignment. To achieve this, introduce novel representation learning (RTL)...