arXiv subjects

Fu Li

Publications and source records attributed to Fu Li.

103 records · Page 6Linked to original sources

arXiv2018

StNet: Local and Global Spatial-Temporal Modeling for Action Recognition

Despite the success of deep learning for static image understanding, it remains unclear what are the most effective network architectures for the spatial-temporal modeling in videos. In this paper, in contrast to the existing CNN+RNN or pure 3D convolution based approaches, we explore a novel spatial temporal network (StNet) architecture for both local and global spatial-temporal modeling in videos. Particularly, StNet stacks N successive video frames into a \emph{super-image} which has 3N channels and applies 2D convolution on super-images to capture local spatial-temporal relationship. To model global spatial-temporal relationship, we apply temporal convolution on the local spatial-temporal feature maps. Specifically, a novel temporal Xception block is proposed in StNet. It employs a separate channel-wise and temporal-wise convolution over the feature sequence of video. Extensive experiments on the Kinetics dataset demonstrate that our framework outperforms several state-of-the-art approaches in action recognition and can strike a satisfying trade-off between recognition accuracy and model complexity. We further demonstrate the generalization performance of the leaned video representations on the UCF101 dataset.

cs.CV↗

arXiv2018

Colloidal Tin Sulfide Nanosheets: Formation Mechanism, Ligand-mediated Shape Tuning and Photo-detection

Colloidal materials of tin(II) sulfide (SnS), as a layered semiconductor with a narrow band gap, are emerging as a potential alternative to the more toxic metal chalcogenides (PbS, PbSe, CdS, CdSe) for various applications such as electronic and optoelectronic devices. We describe a new and simple pathway to produce colloidal SnS nanosheets with large lateral sizes and controllable thickness, as well as single-crystallinity. The synthesis of the nanosheets is achieved by employing tin(II) acetate as tin precursor instead of harmful precursors such as bis[bis(trimethylsilyl)amino] tin(II) and halogen-involved precursors like tin chloride, which limits the large-scale production. We successfully tuned the morphology between squared nanosheets with lateral dimensions from 150 to about 500 nm and a thickness from 24 to 29 nm, and hexagonal nanosheets with lateral sizes from 230 to 1680 nm and heights ranging from 16 to 50 nm by varying the ligands oleic acid and trioctylphosphine. The formation mechanism of both shapes has been investigated in depth, which is also supported by DFT simulations. The optoelectronic measurements show their relatively high conductivity with a pronounced sensitivity to light, which is promising in terms of photo-switching, photo-sensing, and photovoltaic applications also due to their reduced toxicity.

cond-mat.mtrl-sci↗

arXiv2018

Combinatorial Multi-Armed Bandit with General Reward Functions

In this paper, we study the stochastic combinatorial multi-armed bandit (CMAB) framework that allows a general nonlinear reward function, whose expected value may not depend only on the means of the input random variables but possibly on the entire distributions of these variables. Our framework enables a much larger class of reward functions such as the $\max()$ function and nonlinear utility functions. Existing techniques relying on accurate estimations of the means of random variables, such as the upper confidence bound (UCB) technique, do not work directly on these functions. We propose a new algorithm called stochastically dominant confidence bound (SDCB), which estimates the distributions of underlying random variables and their stochastically dominant confidence bounds. We prove that SDCB can achieve $O(\log{T})$ distribution-dependent regret and $\tilde{O}(\sqrt{T})$ distribution-independent regret, where $T$ is the time horizon. We apply our results to the $K$-MAX problem and expected utility maximization problems. In particular, for $K$-MAX, we provide the first polynomial-time approximation scheme (PTAS) for its offline problem, and give the first $\tilde{O}(\sqrt T)$ bound on the $(1-ε)$-approximation regret of its online problem, for any $ε>0$.

cs.LG↗

arXiv2018

Exploiting Spatial-Temporal Modelling and Multi-Modal Fusion for Human Action Recognition

In this report, our approach to tackling the task of ActivityNet 2018 Kinetics-600 challenge is described in detail. Though spatial-temporal modelling methods, which adopt either such end-to-end framework as I3D \cite{i3d} or two-stage frameworks (i.e., CNN+RNN), have been proposed in existing state-of-the-arts for this task, video modelling is far from being well solved. In this challenge, we propose spatial-temporal network (StNet) for better joint spatial-temporal modelling and comprehensively video understanding. Besides, given that multi-modal information is contained in video source, we manage to integrate both early-fusion and later-fusion strategy of multi-modal information via our proposed improved temporal Xception network (iTXN) for video understanding. Our StNet RGB single model achieves 78.99\% top-1 precision in the Kinetics-600 validation set and that of our improved temporal Xception network which integrates RGB, flow and audio modalities is up to 82.35\%. After model ensemble, we achieve top-1 precision as high as 85.0\% on the validation set and rank No.1 among all submissions.

cs.CV↗

arXiv2017

Investigating the `Past of a Particle' without disturbing it

In a recent article [Chin. Phys. Lett. 34, 020301 (2017)], Ben-Israel et al. have claimed that the experiment proposed in [Chin. Phys. Lett. 32, 050303 (2015)] to determine the past of a quantum particle in a nested Mach-Zehnder interferometer does not work, and they have proposed a modification to the experiment. We show that their claim is false, and the modification is not required.

quant-ph↗

arXiv2017

Revisiting the Effectiveness of Off-the-shelf Temporal Modeling Approaches for Large-scale Video Classification

This paper describes our solution for the video recognition task of ActivityNet Kinetics challenge that ranked the 1st place. Most of existing state-of-the-art video recognition approaches are in favor of an end-to-end pipeline. One exception is the framework of DevNet. The merit of DevNet is that they first use the video data to learn a network (i.e. fine-tuning or training from scratch). Instead of directly using the end-to-end classification scores (e.g. softmax scores), they extract the features from the learned network and then fed them into the off-the-shelf machine learning models to conduct video classification. However, the effectiveness of this line work has long-term been ignored and underestimated. In this submission, we extensively use this strategy. Particularly, we investigate four temporal modeling approaches using the learned features: Multi-group Shifting Attention Network, Temporal Xception Network, Multi-stream sequence Model and Fast-Forward Sequence Model. Experiment results on the challenging Kinetics dataset demonstrate that our proposed temporal modeling approaches can significantly improve existing approaches in the large-scale video recognition tasks. Most remarkably, our best single Multi-group Shifting Attention Network can achieve 77.7% in term of top-1 accuracy and 93.2% in term of top-5 accuracy on the validation set.

cs.CV↗

arXiv2017

Temporal Modeling Approaches for Large-scale Youtube-8M Video Understanding

This paper describes our solution for the video recognition task of the Google Cloud and YouTube-8M Video Understanding Challenge that ranked the 3rd place. Because the challenge provides pre-extracted visual and audio features instead of the raw videos, we mainly investigate various temporal modeling approaches to aggregate the frame-level features for multi-label video recognition. Our system contains three major components: two-stream sequence model, fast-forward sequence model and temporal residual neural networks. Experiment results on the challenging Youtube-8M dataset demonstrate that our proposed temporal modeling approaches can significantly improve existing temporal modeling approaches in the large-scale video recognition tasks. To be noted, our fast-forward LSTM with a depth of 7 layers achieves 82.75% in term of GAP@20 on the Kaggle Public test set.

cs.CV↗

arXiv2015

Characterizing Propositional Proofs as Non-Commutative Formulas

Does every Boolean tautology have a short propositional-calculus proof? Here, a propositional calculus (i.e. Frege) proof is a proof starting from a set of axioms and deriving new Boolean formulas using a set of fixed sound derivation rules. Establishing any super-polynomial size lower bound on Frege proofs (in terms of the size of the formula proved) is a major open problem in proof complexity, and among a handful of fundamental hardness questions in complexity theory by and large. Non-commutative arithmetic formulas, on the other hand, constitute a quite weak computational model, for which exponential-size lower bounds were shown already back in 1991 by Nisan [Nis91] who used a particularly transparent argument. In this work we show that Frege lower bounds in fact follow from corresponding size lower bounds on non-commutative formulas computing certain polynomials (and that such lower bounds on non-commutative formulas must exist, unless NP=coNP). More precisely, we demonstrate a natural association between tautologies $T$ to non-commutative polynomials $p$, such that: if $T$ has a polynomial-size Frege proof then $p$ has a polynomial-size non-commutative arithmetic formula; and conversely, when $T$ is a DNF, if $p$ has a polynomial-size non-commutative arithmetic formula over $GF(2)$ then $T$ has a Frege proof of quasi-polynomial size.

cs.CC↗

arXiv2015

Combining Traditional Marketing and Viral Marketing with Amphibious Influence Maximization

In this paper, we propose the amphibious influence maximization (AIM) model that combines traditional marketing via content providers and viral marketing to consumers in social networks in a single framework. In AIM, a set of content providers and consumers form a bipartite network while consumers also form their social network, and influence propagates from the content providers to consumers and among consumers in the social network following the independent cascade model. An advertiser needs to select a subset of seed content providers and a subset of seed consumers, such that the influence from the seed providers passing through the seed consumers could reach a large number of consumers in the social network in expectation. We prove that the AIM problem is NP-hard to approximate to within any constant factor via a reduction from Feige's k-prover proof system for 3-SAT5. We also give evidence that even when the social network graph is trivial (i.e. has no edges), a polynomial time constant factor approximation for AIM is unlikely. However, when we assume that the weighted bi-adjacency matrix that describes the influence of content providers on consumers is of constant rank, a common assumption often used in recommender systems, we provide a polynomial-time algorithm that achieves approximation ratio of $(1-1/e-ε)^3$ for any (polynomially small) $ε> 0$. Our algorithmic results still hold for a more general model where cascades in social network follow a general monotone and submodular function.

cs.SI↗

arXiv2014

An ideal experiment to determine the 'past of a particle' in the nested Mach-Zehnder Interferometer

An ideal experiment is designed to determine the past of a particle in the nested Mach-Zehnder interferometer (MZI) by using standard quantum mechanics with quantum non-demolition measurements. We find that when the photon reaches the detector, it only follows one arm of the outer interferometer and leaves no trace in the inner MZI; while when it goes through the inner MZI, it cannot reach the detector. Our result obtained from the standard quantum mechanics is contradict to the statement based on two state vector formulism, "the photon did not enter the (inner) interferometer, the photon never left the interferometer, but it was there". Therefore, the statement and also the overlap claim are incorrect.

quant-ph↗

arXiv2014

The effect of dissipation in direct communication scheme

The effect of the dissipation and finite number of beam splitters are discussed. A method using balanced dissipation to improve the communication for finite beam splitters, which greatly increases communication reliability with an expense of decreasing communication efficiency.

quant-ph↗

arXiv2014

Generating Matrix Identities and Proof Complexity

Motivated by the fundamental lower bounds questions in proof complexity, we initiate the study of matrix identities as hard instances for strong proof systems. A matrix identity of $d \times d$ matrices over a field $\mathbb{F}$, is a non-commutative polynomial $f(x_1,\ldots,x_n)$ over $\mathbb{F}$ such that $f$ vanishes on every $d \times d$ matrix assignment to its variables. We focus on arithmetic proofs, which are proofs of polynomial identities operating with arithmetic circuits and whose axioms are the polynomial-ring axioms (these proofs serve as an algebraic analogue of the Extended Frege propositional proof system; and over $GF(2)$ they constitute formally a sub-system of Extended Frege [HT12]). We introduce a decreasing in strength hierarchy of proof systems within arithmetic proofs, in which the $d$th level is a sound and complete proof system for proving $d \times d$ matrix identities (over a given field). For each level $d>2$ in the hierarchy, we establish a proof-size lower bound in terms of the number of variables in the matrix identity proved: we show the existence of a family of matrix identities $f_n$ with $n$ variables, such that any proof of $f_n=0$ requires $Ω(n^{2d})$ number of lines. The lower bound argument uses fundamental results from the theory of algebras with polynomial identities together with a generalization of the arguments in [Hru11]. We then set out to study matrix identities as hard instances for (full) arithmetic proofs. We present two conjectures, one about non-commutative arithmetic circuit complexity and the other about proof complexity, under which up to exponential-size lower bounds on arithmetic proofs (in terms of the arithmetic circuit size of the identities proved) hold. Finally, we discuss the applicability of our approach to strong propositional proof systems such as Extended Frege.

cs.CC↗

arXiv2013

Energy-Aware Aggregation of Dynamic Temporal Workload in Data Centers

Data center providers seek to minimize their total cost of ownership (TCO), while power consumption has become a social concern. We present formulations to minimize server energy consumption and server cost under three different data center server setups (homogeneous, heterogeneous, and hybrid hetero-homogeneous clusters) with dynamic temporal workload. Our studies show that the homogeneous model significantly differs from the heterogeneous model in computational time (by an order of magnitude). To be able to compute optimal configurations in near real-time for large scale data centers, we propose two modes, aggregation by maximum and aggregation by mean. In addition, we propose two aggregation methods, static (periodic) aggregation and dynamic (aperiodic) aggregation. We found that in the aggregation by maximum mode, the dynamic aggregation resulted in cost savings of up to approximately 18% over the static aggregation. In the aggregation by mean mode, the dynamic aggregation by mean could save up to approximately 50% workload rearrangement compared to the static aggregation by mean mode. Overall, our methodology helps to understand the trade-off in energy-aware aggregation.

cs.DC↗