arXiv subjects

Siqi Liu

Publications and source records attributed to Siqi Liu.

120 records · Page 7Linked to original sources

arXiv2018

Improved Image Captioning via Policy Gradient optimization of SPIDEr

Current image captioning methods are usually trained via (penalized) maximum likelihood estimation. However, the log-likelihood score of a caption does not correlate well with human assessments of quality. Standard syntactic evaluation metrics, such as BLEU, METEOR and ROUGE, are also not well correlated. The newer SPICE and CIDEr metrics are better correlated, but have traditionally been hard to optimize for. In this paper, we show how to use a policy gradient (PG) method to directly optimize a linear combination of SPICE and CIDEr (a combination we call SPIDEr): the SPICE score ensures our captions are semantically faithful to the image, while CIDEr score ensures our captions are syntactically fluent. The PG method we propose improves on the prior MIXER approach, by using Monte Carlo rollouts instead of mixing MLE training with PG. We show empirically that our algorithm leads to easier optimization and improved results compared to MIXER. Finally, we show that using our PG method we can optimize any of the metrics, including the proposed SPIDEr metric which results in image captions that are strongly preferred by human raters compared to captions generated by the same model but trained to optimize MLE or the COCO metrics.

cs.CV↗

arXiv2017

3D Anisotropic Hybrid Network: Transferring Convolutional Features from 2D Images to 3D Anisotropic Volumes

While deep convolutional neural networks (CNN) have been successfully applied for 2D image analysis, it is still challenging to apply them to 3D anisotropic volumes, especially when the within-slice resolution is much higher than the between-slice resolution and when the amount of 3D volumes is relatively small. On one hand, direct learning of CNN with 3D convolution kernels suffers from the lack of data and likely ends up with poor generalization; insufficient GPU memory limits the model size or representational power. On the other hand, applying 2D CNN with generalizable features to 2D slices ignores between-slice information. Coupling 2D network with LSTM to further handle the between-slice information is not optimal due to the difficulty in LSTM learning. To overcome the above challenges, we propose a 3D Anisotropic Hybrid Network (AH-Net) that transfers convolutional features learned from 2D images to 3D anisotropic volumes. Such a transfer inherits the desired strong generalization capability for within-slice information while naturally exploiting between-slice information for more effective modelling. The focal loss is further utilized for more effective end-to-end learning. We experiment with the proposed 3D AH-Net on two different medical image analysis tasks, namely lesion detection from a Digital Breast Tomosynthesis volume, and liver and liver tumor segmentation from a Computed Tomography volume and obtain the state-of-the-art results.

cs.CV↗

arXiv2017

Retention Time of Peptides in Liquid Chromatography Is Well Estimated upon Deep Transfer Learning

A fully automatic prediction for peptide retention time (RT) in liquid chromatography (LC), termed as DeepRT, was developed using deep learning approach, an ensemble of Residual Network (ResNet) and Long Short-Term Memory (LSTM). In contrast to the traditional predictor based on the hand-crafted features for peptides, DeepRT learns features from raw amino acid sequences and makes relatively accurate prediction of peptide RTs with 0.987 R2 for unmodified peptides. Furthermore, by virtue of transfer learning, DeepRT enables utilization of the peptides datasets generated from different LC conditions and of different modification status, resulting in the RT prediction of 0.992 R2 for unmodified peptides and 0.978 R2 for post-translationally modified peptides. Even though chromatographic behaviors of peptides are quite complicated, the study here demonstrated that peptide RT prediction could be largely improved by deep transfer learning. The DeepRT software is freely available at https://github.com/horsepurve/DeepRT, under Apache2 open source License.

q-bio.QM↗

arXiv2017

On the Competition Complexity of Dynamic Mechanism Design

The Competition Complexity of an auction measures how much competition is needed for the revenue of a simple auction to surpass the optimal revenue. A classic result from auction theory by Bulow and Klemperer [9], states that the Competition Complexity of VCG, in the case of n i.i.d. buyers and a single item, is 1, i.e., it is better to recruit one extra buyer and run a second price auction than to learn exactly the buyers' underlying distribution and run the revenue-maximizing auction tailored to this distribution. In this paper we study the Competition Complexity of dynamic auctions. Consider the following setting: a monopolist is auctioning off m items in m consecutive stages to n interested buyers. A buyer realizes her value for item k in the beginning of stage k. We prove that the Competition Complexity of dynamic auctions is at most 3n, and at least linear in n, even when the buyers' values are correlated across stages, under a monotone hazard rate assumption on the stage (marginal) distributions. We also prove results on the number of additional buyers necessary for VCG at every stage to be an α-approximation of the optimal revenue; we term this number the α-approximate Competition Complexity. As a corollary we provide the first results on prior-independent dynamic auctions. This is, to the best of our knowledge, the first non-trivial positive guarantees for simple ex-post IR dynamic auctions for correlated stages. A key step towards proving bounds on the Competition Complexity is getting a good benchmark/upper bound to the optimal revenue. To this end, we extend the recent duality framework of Cai et al. [12] to dynamic settings. As an aside to our approach we obtain a revenue non-monotonicity lemma for dynamic auctions, which may be of independent interest.

cs.GT↗

arXiv2017

Detection of Abnormal Input-Output Associations

We study a novel outlier detection problem that aims to identify abnormal input-output associations in data, whose instances consist of multi-dimensional input (context) and output (responses) pairs. We present our approach that works by analyzing data in the conditional (input--output) relation space, captured by a decomposable probabilistic model. Experimental results demonstrate the ability of our approach in identifying multivariate conditional outliers.

cs.AI↗

arXiv2017

DeepRT: deep learning for peptide retention time prediction in proteomics

Accurate predictions of peptide retention times (RT) in liquid chromatography have many applications in mass spectrometry-based proteomics. Herein, we present DeepRT, a deep learning based software for peptide retention time prediction. DeepRT automatically learns features directly from the peptide sequences using the deep convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) model, which eliminates the need to use hand-crafted features or rules. After the feature learning, principal component analysis (PCA) was used for dimensionality reduction, then three conventional machine learning methods were utilized to perform modeling. Two published datasets were used to evaluate the performance of DeepRT and we demonstrate that DeepRT greatly outperforms previous state-of-the-art approaches ELUDE and GPTime.

q-bio.QM↗

arXiv2016

A Performance Survey on Stack-based and Register-based Virtual Machines

Virtual machines have been widely adapted for high-level programming language implementations and for providing a degree of platform neutrality. As the overall use and adaptation of virtual machines grow, the overall performance of virtual machines has become a widely-discussed topic. In this paper, we present a survey on the performance differences of the two most widely adapted types of virtual machines - the stack-based virtual machine and the register-based virtual machine - using various benchmark programs. Additionally, we adopted a new approach of measuring performance by measuring the overall dispatch time, amount of dispatches, fetch time, and execution time while running benchmarks on custom-implemented, lightweight virtual machines. Finally, we present two lightweight, custom-designed, Turing-equivalent virtual machines that are specifically designed in benchmarking virtual machine performance - the "Conceptum" stack-based virtual machine, and the "Inertia" register-based virtual machine. Our result showed that while on average the register machine spends 20.39% less time in executing benchmarks than the stack machine, the stack-based virtual machine is still faster than the virtual machine regarding the instruction fetch time.

cs.PL↗

arXiv2015

Morphometry-Based Longitudinal Neurodegeneration Simulation with MR Imaging

We present a longitudinal MR simulation framework which simulates the future neurodegenerative progression by outputting the predicted follow-up MR image and the voxel-based morphometry (VBM) map. This framework expects the patients to have at least 2 historical MR images available. The longitudinal and cross-sectional VBM maps are extracted to measure the affinity between the target subject and the template subjects collected for simulation. Then the follow-up simulation is performed by resampling the latest available target MR image with a weighted sum of non-linear transformations derived from the best-matched templates. The leave-one-out strategy was used to compare different simulation methods. Compared to the state-of-the-art voxel-based method, our proposed morphometry-based simulation achieves better accuracy in most cases.

cs.CV↗

arXiv2015

Pulsar lensing geometry

Our analysis of archival VLBI data of PSR 0834+06 revealed that its scintillation properties can be precisely modelled using the inclined sheet model (Pen & Levin 2014), resulting in two distinct lens planes. These data strongly favour the grazing sheet model over turbulence as the primary source of pulsar scattering. This model can reproduce the parameters of the observed diffractive scintillation with an accuracy at the percent level. Comparison with new VLBI proper motion results in a direct measure of the ionized ISM screen transverse velocity. The results are consistent with ISM velocities local to the PSR 0834+06 sight-line (through the Galaxy). The simple 1D structure of the lenses opens up the possibility of using interstellar lenses as precision probes for pulsar lens mapping, precision transverse motions in the ISM, and new opportunities for removing scattering to improve pulsar timing. We describe the parameters and observables of this double screen system. While relative screen distances can in principle be accurately determined, a global conformal distance degeneracy exists that allows a rescaling of the absolute distance scale. For PSR B0834+06, we present VLBI astrometry results that provide (for the fist time) a direct measurement of the distance of the pulsar. For targets where independent distance measurements are not available, which are the cases for most of the recycled millisecond pulsars that are the targets of precision timing observations, the degeneracy presented in the lens modelling could be broken if the pulsar resides in a binary system.

astro-ph.SR↗

arXiv2015

Breaking through the high redshift bottleneck of Observational Hubble parameter Data: The Sandage-Loeb signal Scheme

We propose a valid scheme to measure the Hubble parameter $H(z)$ at high redshifts by detecting the Sandage-Loeb signal (SL signal) which can be realized by the next generation extremely large telescope. It will largely extend the current observational Hubble parameter data (OHD) towards the redshift region of $z \in [2.0,5.0]$, the so-called "redshift desert", where other dark energy probes are hard to provide useful information of the cosmic expansion. Quantifying the ability of this future measurement by simulating observational data for a CODEX (COsmic Dynamics and EXo-earth experiment)-like survey and constraining various cosmological models, we find that the SL signal scheme brings the redshift upper-limit of OHD from $z_\mathrm{max}=2.3$ to $z_\mathrm{max}\simeq 5.0$, provides more accurate constraints on different dark energy models, and greatly changes the degeneracy direction of the parameters. For the $Λ$CDM case, the accuracy of $Ω_m$ is improved by $58\%$ and the degeneracy between $Ω_m$ and $Ω_ Λ$ is rotated to the vertical direction of $Ω_k = 0$ line strongly; for the $w$CDM case, the accuracy of $w$ is improved by $15\%$. The Fisher matrix forecast on different time-dependent $w(z)$ is also performed.

astro-ph.CO↗

arXiv2014

Four New Observational $H(z)$ Data From Luminous Red Galaxies of Sloan Digital Sky Survey Data Release Seven

By adopting the differential age method, we utilize selected 17832 luminous red galaxies (LRGs) from Sloan Digital Sky Survey Data Release Seven (SDSS DR7) covering redshift $0<z<0.4$ to measure Hubble parameters. Using a full spectrum fitting package UlySS, these spectra are reduced with single stellar population (SSP) models and optimal age information of our selected sample are derived. With the decreasing age-redshift relation, four new observational $H(z)$ data (OHD) points are obtained, which are $H(z)=69.0\pm19.6$ km s$^{-1}$ Mpc$^{-1}$ at $z=0.07$, $H(z)=68.6\pm26.2$ km s$^{-1}$ Mpc$^{-1}$ at $z=0.12$, $H(z)$=$72.9\pm29.6$ km s$^{-1}$ Mpc$^{-1}$ at $z=0.2$ and $H(z)$=$88.8\pm36.6$ km s$^{-1}$ Mpc$^{-1}$ at $z=0.28$, respectively. Combined with other 21 available OHD data points, a performance of constraint on both flat and non-flat $Λ$CDM model is presented.

astro-ph.CO↗

arXiv2014

Distinguishing between inhomogeneous model and $Λ\textrm{CDM}$ model with the cosmic age method

Cosmological observables could be used to construct cosmological models, however, a fixed number of observables limited on the light cone is not enough to uniquely determine a certain model. A reconstructed spherically symmetric, inhomogeneous model that share the same angular-diameter-distance-redshift relationship $d_A(z)$ and Hubble parameter $H(z)$ besides $Λ\textrm{CDM}$ model (which we call LTB-$Λ\textrm{CDM}$ model in this paper), may provide another solution. Cosmic age, which is off the light cone, could be employed to distinguish these two models. We derive the formulae for age calculation with origin conditions. From the data given by 9-year WMAP measurement, we compute the likelihood of the parameters in these two models respectively by using the Distance Prior method and do likelihood analysis by generating Monte Carlo Markov Chain for the purpose of breaking the degeneracy of $Ω_m$ and $H_0$ (the parameters that we use for calculation). The results yield to be: $t_{Λ\textrm{CDM}} =13.76 \pm 0.09 ~\rm Gyr$, $t_{\rm {LTB}-Λ\textrm{CDM}} =11.38 \pm 0.15 ~\rm Gyr$, both in $1σ$ agreement with the constraint of cosmic age given by metal-deficient stars. The cosmic age method that is set in this paper enables us to distinguish between the inhomogeneous model and $Λ\textrm{CDM}$ model.

astro-ph.CO↗