Source author record

Yu Qian

Yu Qian appears in the imported research catalog. Authorship, coauthor and topic links are available while profile ownership is still unclaimed.

ResearcherUnclaimed source record

nlin.AO Computer Vision hep-ex nlin.PS physics.ins-det Software Engineering

Catalog footprint

What is connected

5works

6topics

4close collaborators

Actions

Connect this record

Open graph Browse works

Inspect adjacent papers, topics, institutions and collaborators without losing the researcher page.

Building this map preview

BZPEER is loading the nearby papers, people, topics and institutions for this page.

preprint2023arXiv

Boosting Pruned Networks with Linear Over-parameterization

Structured pruning compresses neural networks by reducing channels (filters) for fast inference and low footprint at run-time. To restore accuracy after pruning, fine-tuning is usually applied to pruned networks. However, too few remaining parameters in pruned networks inevitably bring a great challenge to fine-tuning to restore accuracy. To address this challenge, we propose a novel method that first linearly over-parameterizes the compact layers in pruned networks to enlarge the number of fine-tuning parameters and then re-parameterizes them to the original layers after fine-tuning. Specifically, we equivalently expand the convolution/linear layer with several consecutive convolution/linear layers that do not alter the current output feature maps. Furthermore, we utilize similarity-preserving knowledge distillation that encourages the over-parameterized block to learn the immediate data-to-data similarities of the corresponding dense layer to maintain its feature learning ability. The proposed method is comprehensively evaluated on CIFAR-10 and ImageNet which significantly outperforms the vanilla fine-tuning strategy, especially for large pruning ratio.

preprint2020arXiv

PSCS: A Path-based Neural Model for Semantic Code Search

To obtain code snippets for reuse, programmers prefer to search for related documents, e.g., blogs or Q&A, instead of code itself. The major reason is due to the semantic diversity and mismatch between queries and code snippets. Deep learning models have been proposed to address this challenge. Compared with approaches using information retrieval techniques, deep learning models do not suffer from the information loss caused by refining user intention into keywords. However, the performance of previous works is not satisfactory because they ignore the importance of code structure. When the semantics of code (e.g., identifier names, APIs) are ambiguous, code structure may be the only feature for the model to utilize. In that case, previous works relearn the structural information from lexical tokens of code, which is extremely difficult for a model without any domain knowledge. In this work, we propose PSCS, a path-based neural model for semantic code search. Our model encodes both the semantics and structures of code represented by AST paths. We train and evaluate our model over 330k-19k query-function pairs, respectively. The evaluation results demonstrate that PSCS achieves a SuccessRate of 47.6% and a Mean Reciprocal Rank (MRR) of 30.4% when considering the top-10 results with a match. The proposed approach significantly outperforms both DeepCS, the first approach that applies deep learning to code search task, and CARLCS, a state-of-the-art approach that introduces a co-attentive representation learning model on the basis of DeepCS. The importance of code structure is demonstrated with an ablation study on code features, which enlightens model design for further studies.

preprint2015arXiv

A Muti-channel Distributed DAQ for n-TPC

A new fast neutron spectrometer named n-TPC has been designed by LPRI (Key Laboratory of Particle & Radiation Imaging, Ministry of Education) at Tsinghua University. The neutron energy spectrum can be calculated from the recoil angle and energy of the recoil proton detected by a 704-pad GEM-TPC. In beam tests at IHIP (Institute of Heavy Ion Physics, Peking University) in 2014, n-TPC performed better than 6 percents at 6MeV energy resolution and 0.5 percents detection efficiency. To find the best working parameters (the component and proportion of the gas, the high voltage between each GEM layer, etc.) of the n-TPC and support its application in various conditions, a multichannel distributed DAQ has been design to read out the signals from the 704 channels. With over 25 Ms/s sampling rate and 12 bit resolution for each channel, it can record the time and amplitude information as well as traditional DAQs in the TPC application domain. The main design objective of this distributed DAQ, however, is more flexible parameter modulation and operation. It can support the n-TPC without the limitation of the chassis and categorize signals arriving from the 704 channels at the same time by different events without event triggers.

preprint2011arXiv

Universal driving structure of self-sustained oscillatory complex networks

Recently, self-sustained oscillations in complex networks consisting of nonoscillatory nodes (network oscillators) have attracted great interest in diverse natural and social fields. Due to complexity of network behaviors, little is known so far about the basic structures and fundamental rules underlying the oscillations, not to mention the principles of how to control it. In this article we propose a common design principle for oscillations; predict novel and universal Branched Circle (BC) structures of oscillatory networks based on this principle; and suggest an operable Complexity Reduction Method to reveal the BC structures. These ideas are applied to excitable cell networks (including neural cell networks), and genomic regulatory networks. Universal BC structures are identified clearly in these two considerably different systems. These BC structures reveal for the first time both oscillation sources and wave propagation pathways of complex networks, and guide us to control the oscillations with surprisingly high efficiency.

preprint2010arXiv

Structure and control of self-sustained target waves in excitable small-world networks

Small-world networks describe many important practical systems among which neural networks consisting of excitable nodes are the most typical ones. In this paper we study self-sustained oscillations of target waves in excitable small-world networks. A novel dominant phase-advanced driving (DPAD) method, which is generally applicable for analyzing all oscillatory complex networks consisting of nonoscillatory nodes, is proposed to reveal the self-organized structures supporting this type of oscillations. The DPAD method explicitly explores the oscillation sources and wave propagation paths of the systems, which are otherwise deeply hidden in the complicated patterns of randomly distributed target groups. Based on the understanding of the self-organized structure, the oscillatory patterns can be controlled with extremely high efficiency.