arXiv subjects

Jinfeng Zhang

Publications and source records attributed to Jinfeng Zhang.

27 records · Page 2Linked to original sources

arXiv2019

Multi-Granularity Self-Attention for Neural Machine Translation

Current state-of-the-art neural machine translation (NMT) uses a deep multi-head self-attention network with no explicit phrase information. However, prior work on statistical machine translation has shown that extending the basic translation unit from words to phrases has produced substantial improvements, suggesting the possibility of improving NMT performance from explicit modeling of phrases. In this work, we present multi-granularity self-attention (Mg-Sa): a neural network that combines multi-head self-attention and phrase modeling. Specifically, we train several attention heads to attend to phrases in either n-gram or syntactic formalism. Moreover, we exploit interactions among phrases to enhance the strength of structure modeling - a commonly-cited weakness of self-attention. Experimental results on WMT14 English-to-German and NIST Chinese-to-English translation tasks show the proposed approach consistently improves performance. Targeted linguistic analysis reveals that Mg-Sa indeed captures useful phrase information at various levels of granularities.

cs.CL↗

arXiv2019

Modeling Recurrence for Transformer

Recently, the Transformer model that is based solely on attention mechanisms, has advanced the state-of-the-art on various machine translation tasks. However, recent studies reveal that the lack of recurrence hinders its further improvement of translation capacity. In response to this problem, we propose to directly model recurrence for Transformer with an additional recurrence encoder. In addition to the standard recurrent neural network, we introduce a novel attentive recurrent network to leverage the strengths of both attention and recurrent networks. Experimental results on the widely-used WMT14 English-German and WMT17 Chinese-English translation tasks demonstrate the effectiveness of the proposed approach. Our studies also reveal that the proposed model benefits from a short-cut that bridges the source and target sequences with a single recurrent layer, which outperforms its deep counterpart.

cs.CL↗

arXiv2017

High performance position-sensitive-detector based on graphene-silicon heterojunction

Position-sensitive-detectors (PSDs) based on lateral photoeffect have been widely used in diverse applications, including optical engineering, aerospace and military fields. With increasing demands in long working distance, low energy consumption, and weak signal sensing systems, the poor responsivity of conventional Silicon-based PSDs has become a bottleneck limiting their applications. Herein, we propose a high-performance passive PSD based on graphene-Si heterostructure. The graphene is adapted as a photon absorbing and charge separation layer working together with Si as a junction, while the high mobility provides promising ultra-long carrier diffusion length and facilitates large active area of the device. A PSD with working area of 8 mm x 8 mm is demonstrated to present excellent position sensitivity to weak light at nWs level (much better than the limit of ~μWs of Si p-i-n PSDs). More importantly, it shows very fast response and low degree of non-linearity of ~3%, and extends the operating wavelength to the near infrared (IR) region (1319 and 1550 nm). This work therefore provides a new strategy for high performance and broadband PSDs.

physics.app-ph↗

arXiv2015

iSeg: an algorithm for segmentation of genomic data

Identification of functional elements of a genome often requires dividing a sequence of measurements along a genome into segments differing from adjacent segments. In many applications, the mean of the measured values at multiple genomic locations in a segment is used to make inference of the property of interest. The segments with non-zero means often correspond to genomic regions with certain biological events, such as changes between two conditions. This problem is often called the segmentation problem in the field of genomics, and the change-point problem in other scientific disciplines. We designed an efficient algorithm, called iSeg, for segmentation of high-throughput genomic profiles. iSeg first utilizes dynamic programming to compute the significance for a large number of candidate segments. It then uses tree-based data structures to detect overlapping significant regions and update them simultaneously. Refinement and merging of significant segments are performed at the end to generate the final segmentation. We evaluate iSeg using both simulated and experimental datasets and show that it performs quite well when compared with existing methods.

stat.AP↗

arXiv2006

Empirical Potential Function for Simplified Protein Models: Combining Contact and Local Sequence-Structure Descriptors

An effective potential function is critical for protein structure prediction and folding simulation. Simplified protein models such as those requiring only $C_α$ or backbone atoms are attractive because they enable efficient search of the conformational space. We show residue specific reduced discrete state models can represent the backbone conformations of proteins with small RMSD values. However, no potential functions exist that are designed for such simplified protein models. In this study, we develop optimal potential functions by combining contact interaction descriptors and local sequence-structure descriptors. The form of the potential function is a weighted linear sum of all descriptors, and the optimal weight coefficients are obtained through optimization using both native and decoy structures. The performance of the potential function in test of discriminating native protein structures from decoys is evaluated using several benchmark decoy sets. Our potential function requiring only backbone atoms or $C_α$ atoms have comparable or better performance than several residue-based potential functions that require additional coordinates of side chain centers or coordinates of all side chain atoms. By reducing the residue alphabets down to size 5 for local structure-sequence relationship, the performance of the potential function can be further improved. Our results also suggest that local sequence-structure correlation may play important role in reducing the entropic cost of protein folding.

q-bio.BM↗

arXiv2004

Potential function of simplified protein models for discriminating native proteins from decoys: Combining contact interaction and local sequence-dependent geometry

An effective potential function is critical for protein structure prediction and folding simulation. For simplified models of proteins where coordinates of only $C_α$ atoms need to be specified, an accurate potential function is important. Such a simplified model is essential for efficient search of conformational space. In this work, we present a formulation of potential function for simplified representations of protein structures. It is based on the combination of descriptors derived from residue-residue contact and sequence-dependent local geometry. The optimal weight coefficients for contact and local geometry is obtained through optimization by maximizing margins among native and decoy structures. The latter are generated by chain growth and by gapless threading. The performance of the potential function in blind test of discriminating native protein structures from decoys is evaluated using several benchmark decoy sets. This potential function have comparable or better performance than several residue-based potential functions that require in addition coordinates of side chain centers or coordinates of all side chain atoms.

q-bio.BM↗

arXiv2004

Importance of chirality and reduced flexibility of protein side chains: A study with square and tetrahedral lattice models

In simple models side chains are often represented implicitly (e.g., by spin-states) or simplified as one atom. We study side chain effects using square lattice and tetrahedral lattice models, with explicitly side chains of two atoms. We distinguish effects due to chirality and effects due to side chain flexibilities, since residues in proteins are L-residues, and their side chains adopt different rotameric states. Short chains are enumerated exhaustively. For long chains, we sample effectively rare events (eg, compact conformations) and obtain complete pictures of ensemble properties of these models at all compactness region. We find that both chirality and reduced side chain flexibility lower the folding entropy significantly for globally compact conformations, suggesting that they are important properties of residues to ensure fast folding and stable native structure. This corresponds well with our finding that natural amino acid residues have reduced effective flexibility, as evidenced by analysis of rotamer libraries and side chain rotatable bonds. We further develop a method calculating the exact side-chain entropy for a given back bone structure. We show that simple rotamer counting often underestimates side chain entropy significantly, and side chain entropy does not always correlate well with main chain packing. Among compact backbones with maximum side chain entropy, helical structures emerges as the dominating configurations. Our results suggest that side chain entropy may be an important factor contributing to the formation of alpha helices for compact conformations.

q-bio.BM↗

arXiv2003

Origin of Scaling Behavior of Protein Packing Density: A Sequential Monte Carlo Study of Compact Long Chain Polymers

Single domain proteins are thought to be tightly packed. The introduction of voids by mutations is often regarded as destabilizing. In this study we show that packing density for single domain proteins decreases with chain length. We find that the radius of gyration provides poor description of protein packing but the alpha contact number we introduce here characterize proteins well. We further demonstrate that protein-like scaling relationship between packing density and chain length is observed in off-lattice self-avoiding walks. A key problem in studying compact chain polymer is the attrition problem: It is difficult to generate independent samples of compact long self-avoiding walks. We develop an algorithm based on the framework of sequential Monte Carlo and succeed in generating populations of compact long chain off-lattice polymers up to length $N=2,000$. Results based on analysis of these chain polymers suggest that maintaining high packing density is only characteristic of short chain proteins. We found that the scaling behavior of packing density with chain length of proteins is a generic feature of random polymers satisfying loose constraint in compactness. We conclude that proteins are not optimized by evolution to eliminate packing voids.

cond-mat.soft↗

arXiv2002

Statistical Geometry of Packing Defects of Lattice Chain Polymer from Enumeration and Sequential Monte Carlo Method

Voids exist in proteins as packing defects and are often associated with protein functions. We study the statistical geometry of voids in two-dimensional lattice chain polymers. We define voids as topological features and develop a simple algorithm for their detection. For short chains, void geometry is examined by enumerating all conformations. For long chains, the space of void geometry is explored using sequential Monte Carlo importance sampling and resampling techniques. We characterize the relationship of geometric properties of voids with chain length, including probability of void formation, expected number of voids, void size, and wall size of voids. We formalize the concept of packing density for lattice polymers, and further study the relationship between packing density and compactness, two parameters frequently used to describe protein packing. We find that both fully extended and maximally compact polymers have the highest packing density, but polymers with intermediate compactness have low packing density. To study the conformational entropic effects of void formation, we characterize the conformation reduction factor of void formation and found that there are strong end-effect. Voids are more likely to form at the chain end. The critical exponent of end-effect is twice as large as that of self-contacting loop formation when existence of voids is not required. We also briefly discuss the sequential Monte Carlo sampling and resampling techniques used in this study.

physics.bio-ph↗