arXiv subjects

Feng Sun

Publications and source records attributed to Feng Sun.

At least 37 records · Page 2Linked to original sources

arXiv2024

ASI++: Towards Distributionally Balanced End-to-End Generative Retrieval

Generative retrieval, a promising new paradigm in information retrieval, employs a seq2seq model to encode document features into parameters and decode relevant document identifiers (IDs) based on search queries. Existing generative retrieval solutions typically rely on a preprocessing stage to pre-define document IDs, which can suffer from a semantic gap between these IDs and the retrieval task. However, end-to-end training for both ID assignments and retrieval tasks is challenging due to the long-tailed distribution characteristics of real-world data, resulting in inefficient and unbalanced ID space utilization. To address these issues, we propose ASI++, a novel fully end-to-end generative retrieval method that aims to simultaneously learn balanced ID assignments and improve retrieval performance. ASI++ builds on the fully end-to-end training framework of vanilla ASI and introduces several key innovations. First, a distributionally balanced criterion addresses the imbalance in ID assignments, promoting more efficient utilization of the ID space. Next, a representation bottleneck criterion enhances dense representations to alleviate bottlenecks in learning ID assignments. Finally, an information consistency criterion integrates these processes into a joint optimization framework grounded in information theory. We further explore various module structures for learning ID assignments, including neural quantization, differentiable product quantization, and residual quantization. Extensive experiments on both public and industrial datasets demonstrate the effectiveness of ASI++ in improving retrieval performance and achieving balanced ID assignments.

cs.IR↗

arXiv2024

MoRA: High-Rank Updating for Parameter-Efficient Fine-Tuning

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 the low-rank updating mechanism may limit the ability of LLMs to effectively learn and memorize new knowledge. Inspired by this observation, we propose a new method called MoRA, which employs a square matrix to achieve high-rank updating while maintaining the same number of trainable parameters. To achieve it, we introduce the corresponding non-parameter operators to reduce the input dimension and increase the output dimension for the square matrix. Furthermore, these operators ensure that the weight can be merged back into LLMs, which makes our method can be deployed like LoRA. We perform a comprehensive evaluation of our method across five tasks: instruction tuning, mathematical reasoning, continual pretraining, memory and pretraining. Our method outperforms LoRA on memory-intensive tasks and achieves comparable performance on other tasks.

cs.CL↗

arXiv2024

A Deep Model for Partial Multi-Label Image Classification with Curriculum Based Disambiguation

In this paper, we study the partial multi-label (PML) image classification problem, where each image is annotated with a candidate label set consists of multiple relevant labels and other noisy labels. Existing PML methods typically design a disambiguation strategy to filter out noisy labels by utilizing prior knowledge with extra assumptions, which unfortunately is unavailable in many real tasks. Furthermore, because the objective function for disambiguation is usually elaborately designed on the whole training set, it can be hardly optimized in a deep model with SGD on mini-batches. In this paper, for the first time we propose a deep model for PML to enhance the representation and discrimination ability. On one hand, we propose a novel curriculum based disambiguation strategy to progressively identify ground-truth labels by incorporating the varied difficulties of different classes. On the other hand, a consistency regularization is introduced for model retraining to balance fitting identified easy labels and exploiting potential relevant labels. Extensive experimental results on the commonly used benchmark datasets show the proposed method significantly outperforms the SOTA methods.

cs.CV↗

arXiv2024

ResLoRA: Identity Residual Mapping in Low-Rank Adaption

As one of the most popular parameter-efficient fine-tuning (PEFT) methods, low-rank adaptation (LoRA) is commonly applied to fine-tune large language models (LLMs). However, updating the weights of LoRA blocks effectively and expeditiously is challenging due to the long calculation path in the original model. To address this, we propose ResLoRA, an improved framework of LoRA. By adding residual paths during training and using merging approaches to eliminate these extra paths during inference, our method can achieve better results in fewer training steps without any extra trainable parameters or inference cost compared to LoRA. The experiments on NLG, NLU, and text-to-image tasks demonstrate the effectiveness of our method. To the best of our knowledge, ResLoRA is the first work that combines the residual path with LoRA. The code of our method is available at https://github.com/microsoft/LMOps/tree/main/reslora .

cs.CL↗

arXiv2024

HD-Eval: Aligning Large Language Model Evaluators Through Hierarchical Criteria Decomposition

Large language models (LLMs) have emerged as a promising alternative to expensive human evaluations. However, the alignment and coverage of LLM-based evaluations are often limited by the scope and potential bias of the evaluation prompts and criteria. To address this challenge, we propose HD-Eval, a novel framework that iteratively aligns LLM-based evaluators with human preference via Hierarchical Criteria Decomposition. HD-Eval inherits the essence from the evaluation mindset of human experts and enhances the alignment of LLM-based evaluators by decomposing a given evaluation task into finer-grained criteria, aggregating them according to estimated human preferences, pruning insignificant criteria with attribution, and further decomposing significant criteria. By integrating these steps within an iterative alignment training process, we obtain a hierarchical decomposition of criteria that comprehensively captures aspects of natural language at multiple levels of granularity. Implemented as a white box, the human preference-guided aggregator is efficient to train and more explainable than relying solely on prompting, and its independence from model parameters makes it applicable to closed-source LLMs. Extensive experiments on three evaluation domains demonstrate the superiority of HD-Eval in further aligning state-of-the-art evaluators and providing deeper insights into the explanation of evaluation results and the task itself.

cs.CL↗

arXiv2024

Text Diffusion with Reinforced Conditioning

Diffusion models have demonstrated exceptional capability in generating high-quality images, videos, and audio. Due to their adaptiveness in iterative refinement, they provide a strong potential for achieving better non-autoregressive sequence generation. However, existing text diffusion models still fall short in their performance due to a challenge in handling the discreteness of language. This paper thoroughly analyzes text diffusion models and uncovers two significant limitations: degradation of self-conditioning during training and misalignment between training and sampling. Motivated by our findings, we propose a novel Text Diffusion model called TREC, which mitigates the degradation with Reinforced Conditioning and the misalignment by Time-Aware Variance Scaling. Our extensive experiments demonstrate the competitiveness of TREC against autoregressive, non-autoregressive, and diffusion baselines. Moreover, qualitative analysis shows its advanced ability to fully utilize the diffusion process in refining samples.

cs.CL↗

arXiv2024

Re-understanding of the deformation potential constant in the single crystal silicon

The mobility formula based on deformation potential (DP) theory is of great importance in semiconductor physics. However, the related calculations for the DP constant are controversial. It is necessary to redo in-depth and comprehensive research on the mobility of single crystal silicon and the related parameters such as the effective mass and the DP constant. In this work the conductivity effective mass is redefined and a method based on the first principles is presented to evaluate the correction of the DP constant. It is found that the effective mass is closer to experimental data and the correction of the DP is a negligible value of about 0.3 eV. Using these parameters, we obtain the mobilities of the single crystal silicon in reasonable agreement with the experimental values. Our method can be effectively applied to the prediction for the mobility in bulk materials.

cond-mat.mtrl-sci↗

arXiv2024

Improving Domain Adaptation through Extended-Text Reading Comprehension

To enhance the domain-specific capabilities of large language models, continued pre-training on a domain-specific corpus is a prevalent method. Recent work demonstrates that adapting models using reading comprehension data formatted by regex-based patterns can significantly improve performance on domain-specific tasks. However, regex-based patterns are incapable of parsing raw corpora using domain-specific knowledge. Furthermore, the question and answer pairs are extracted directly from the corpus in predefined formats offers limited context. To address this limitation, we improve reading comprehension via LLM and clustering. LLM focuses on leveraging domain knowledge within the corpus to refine comprehension stage, while clustering supplies relevant knowledge by extending the context to enrich reading stage. Additionally, our method incorporates parameter-efficient fine-tuning to improve the efficiency of domain adaptation. In comparison to AdaptLLM, our method achieves an improvement exceeding 5% in domain-specific tasks. Our code will available at https://github.com/microsoft/LMOps.

cs.CL↗

arXiv2023

Auto Search Indexer for End-to-End Document Retrieval

Generative retrieval, which is a new advanced paradigm for document retrieval, has recently attracted research interests, since it encodes all documents into the model and directly generates the retrieved documents. However, its power is still underutilized since it heavily relies on the "preprocessed" document identifiers (docids), thus limiting its retrieval performance and ability to retrieve new documents. In this paper, we propose a novel fully end-to-end retrieval paradigm. It can not only end-to-end learn the best docids for existing and new documents automatically via a semantic indexing module, but also perform end-to-end document retrieval via an encoder-decoder-based generative model, namely Auto Search Indexer (ASI). Besides, we design a reparameterization mechanism to combine the above two modules into a joint optimization framework. Extensive experimental results demonstrate the superiority of our model over advanced baselines on both public and industrial datasets and also verify the ability to deal with new documents.

cs.IR↗

arXiv2023

Democratizing Reasoning Ability: Tailored Learning from Large Language Model

Large language models (LLMs) exhibit impressive emergent abilities in natural language processing, but their democratization is hindered due to huge computation requirements and closed-source nature. Recent research on advancing open-source smaller LMs by distilling knowledge from black-box LLMs has obtained promising results in the instruction-following ability. However, the reasoning ability which is more challenging to foster, is relatively rarely explored. In this paper, we propose a tailored learning approach to distill such reasoning ability to smaller LMs to facilitate the democratization of the exclusive reasoning ability. In contrast to merely employing LLM as a data annotator, we exploit the potential of LLM as a reasoning teacher by building an interactive multi-round learning paradigm. This paradigm enables the student to expose its deficiencies to the black-box teacher who then can provide customized training data in return. Further, to exploit the reasoning potential of the smaller LM, we propose self-reflection learning to motivate the student to learn from self-made mistakes. The learning from self-reflection and LLM are all tailored to the student's learning status, thanks to the seamless integration with the multi-round learning paradigm. Comprehensive experiments and analysis on mathematical and commonsense reasoning tasks demonstrate the effectiveness of our method. The code will be available at https://github.com/Raibows/Learn-to-Reason.

cs.CL↗

arXiv2023

Calibrating LLM-Based Evaluator

Recent advancements in large language models (LLMs) on language modeling and emergent capabilities make them a promising reference-free evaluator of natural language generation quality, and a competent alternative to human evaluation. However, hindered by the closed-source or high computational demand to host and tune, there is a lack of practice to further calibrate an off-the-shelf LLM-based evaluator towards better human alignment. In this work, we propose AutoCalibrate, a multi-stage, gradient-free approach to automatically calibrate and align an LLM-based evaluator toward human preference. Instead of explicitly modeling human preferences, we first implicitly encompass them within a set of human labels. Then, an initial set of scoring criteria is drafted by the language model itself, leveraging in-context learning on different few-shot examples. To further calibrate this set of criteria, we select the best performers and re-draft them with self-refinement. Our experiments on multiple text quality evaluation datasets illustrate a significant improvement in correlation with expert evaluation through calibration. Our comprehensive qualitative analysis conveys insightful intuitions and observations on the essence of effective scoring criteria.

cs.CL↗

arXiv2022

MMINR: Multi-frame-to-Multi-frame Inference with Noise Resistance for Precipitation Nowcasting with Radar

Precipitation nowcasting based on radar echo maps is essential in meteorological research. Recently, Convolutional RNNs based methods dominate this field, but they cannot be solved by parallel computation resulting in longer inference time. FCN based methods adopt a multi-frame-to-single-frame inference (MSI) strategy to avoid this problem. They feedback into the model again to predict the next time step to get multi-frame nowcasting results in the prediction phase, which will lead to the accumulation of prediction errors. In addition, precipitation noise is a crucial factor contributing to high prediction errors because of its unpredictability. To address this problem, we propose a novel Multi-frame-to-Multi-frame Inference (MMI) model with Noise Resistance (NR) named MMINR. It avoids error accumulation and resists precipitation noiseś negative effect in parallel computation. NR contains a Noise Dropout Module (NDM) and a Semantic Restore Module (SRM). NDM deliberately dropout noise simple yet efficient, and SRM supplements semantic information of features to alleviate the problem of semantic information mistakenly lost by NDM. Experimental results demonstrate that MMINR can attain competitive scores compared with other SOTAs. The ablation experiments show that the proposed NDM and SRM can solve the aforementioned problems.

cs.CV↗

arXiv2021

RRLFSOR: An Efficient Self-Supervised Learning Strategy of Graph Convolutional Networks

Graph Convolutional Networks (GCNs) are widely used in many applications yet still need large amounts of labelled data for training. Besides, the adjacency matrix of GCNs is stable, which makes the data processing strategy cannot efficiently adjust the quantity of training data from the built graph structures.To further improve the performance and the self-learning ability of GCNs,in this paper, we propose an efficient self-supervised learning strategy of GCNs,named randomly removed links with a fixed step at one region (RRLFSOR).RRLFSOR can be regarded as a new data augmenter to improve over-smoothing.RRLFSOR is examined on two efficient and representative GCN models with three public citation network datasets-Cora,PubMed,and Citeseer.Experiments on transductive link prediction tasks show that our strategy outperforms the baseline models consistently by up to 21.34% in terms of accuracy on three benchmark datasets.

cs.LG↗

arXiv2020

Network experimentation at scale

We describe our framework, deployed at Facebook, that accounts for interference between experimental units through cluster-randomized experiments. We document this system, including the design and estimation procedures, and detail insights we have gained from the many experiments that have used this system at scale. We introduce a cluster-based regression adjustment that substantially improves precision for estimating global treatment effects as well as testing for interference as part of our estimation procedure. With this regression adjustment, we find that imbalanced clusters can better account for interference than balanced clusters without sacrificing accuracy. In addition, we show how logging exposure to a treatment can be used for additional variance reduction. Interference is a widely acknowledged issue with online field experiments, yet there is less evidence from real-world experiments demonstrating interference in online settings. We fill this gap by describing two case studies that capture significant network effects and highlight the value of this experimentation framework.

cs.SI↗

arXiv2020

Influence of different exchange-correlation potentials on twisted structures of bilayer XS2 (X= Mo, Cr)

In this work, we employ the LDA, GGA and GGA with four vdW corrections to study crystal and electronic structures of bilayer transition metal dichalcogenides (TMDs) with different twist angles. We find the GGA interlayer distance of bilayer MoS2 has good agreement with experimental value while vdW correction method still needs to be further improved. Our results indicate the GGA interlayer distances of bilayer XS2 (X= Mo, Cr) with twist angles are smaller than that of normal bilayer, which is the opposite in the LDA case. The GGA results show that reduced bandgap is due to the reduction of interlayer distance and, flat valley and conductivity bands appear owing to twist angle. Our study not only supports valuable information for application possibility of twisted two-dimensional (2D) materials but also stimulates more related research.

cond-mat.mtrl-sci↗

arXiv2020

Hybrid Modeling Application in Control Valve

In view of the serious nonlinearity, time-varying and parameter uncertainty in the physical model of regulating valve, a prediction model of flow rate and pressure of regulating valve based on mixed model was proposed.According to the physical model of the regulator, the parameters that can represent the operation state of the regulator are analyzed, and the relevant parameters are identified by unbiased LSSVM method.The DAMADICS simulation results show that the model can predict the output of flow rate and pressure with better accuracy, which can provide guidance for the design of automatic control valve or fault diagnosis system.

eess.SY↗

arXiv2020

Prediction for structure stability and ultrahigh hydrogen evolution performance of monolayer 2H-CrS2

By a combination of the first-principles calculations and climbing image nudged elastic band method (ciNEB) we investigate structure stabilities and hydrogen evolution reaction (HER) performance of monolayer 2H-CrS2. The results suggest the free energy for the Volmer reaction in the monolayer 2H-CrS2 with S vacancy is 0.07 eV, comparable with Pt-based catalyst, and HER on the surface of the monolayer is prone to the Volmer-Heyrovsky mechanism with no energy barrier. We propose that high HER performance stems from the reduction of the energy level of d-band center. Additionally, the S vacancy leads to defect states in the middle of electronic bandgap and the reduction of potential barrier between the S atom layer and the vacuum, which is conducive to improve HER performance.

cond-mat.mtrl-sci↗

arXiv2020

Effects of biaxial strain and local constant potential on electronic structure of monolayer SnSe

We use the modified Becke-Johnson exchange potential (mBJ) with the spin-orbit coupling effect (SOC) to study effects of biaxial strain and local constant potential on electronic structure of monolayer SnSe. Our results show the fundamental band gap size can be tuned via biaxial strain. Compressive strain (tensile strain) can narrow (enlarge) band gap, and compressive strain causes the transition from quasi-direct to indirect band gap. Moreover, considering that any tuning of electronic structure is realized by changing the periodic potential distribution in the crystalline, we directly add constant potential (CP) to muffin-tin spheres. The results demonstrate that positive and negative CPs can narrow and enlarge band gap, respectively. At CP of 0.9 Ry, semiconductor-metal transition appears, and interestingly a new type of nearly linear dispersions occur at band edge. Our work is good for inspiring more experimental and further theoretical research works.

cond-mat.mtrl-sci↗