arXiv subjects

Junjie Liu

Publications and source records attributed to Junjie Liu.

At least 91 records · Page 5Linked to original sources

arXiv2021

Condensation-Net: Memory-Efficient Network Architecture with Cross-Channel Pooling Layers and Virtual Feature Maps

"Lightweight convolutional neural networks" is an important research topic in the field of embedded vision. To implement image recognition tasks on a resource-limited hardware platform, it is necessary to reduce the memory size and the computational cost. The contribution of this paper is stated as follows. First, we propose an algorithm to process a specific network architecture (Condensation-Net) without increasing the maximum memory storage for feature maps. The architecture for virtual feature maps saves 26.5% of memory bandwidth by calculating the results of cross-channel pooling before storing the feature map into the memory. Second, we show that cross-channel pooling can improve the accuracy of object detection tasks, such as face detection, because it increases the number of filter weights. Compared with Tiny-YOLOv2, the improvement of accuracy is 2.0% for quantized networks and 1.5% for full-precision networks when the false-positive rate is 0.1. Last but not the least, the analysis results show that the overhead to support the cross-channel pooling with the proposed hardware architecture is negligible small. The extra memory cost to support Condensation-Net is 0.2% of the total size, and the extra gate count is only 1.0% of the total size.

cs.CV↗

arXiv2021

Boosting quantum battery performance by structure engineering

Quantum coherences, correlations and collective effects can be harnessed to the advantage of quantum batteries. Here, we introduce a feasible structure engineering scheme that is applicable to spin-based open quantum batteries. Our scheme, which builds solely upon a modulation of spin energy gaps, allows engineered quantum batteries to exploit spin-spin correlations for mitigating environment-induced aging. As a result of this advantage, an engineered quantum battery can preserve relatively more energy as compared with its non-engineered counterpart over the course of the storage phase. Particularly, the excess in stored energy is independent of system size. This implies a scale-invariant passive protection strategy, which we demonstrate on an engineered quantum battery with staggered spin energy gaps. Our findings establish structure engineering as a useful route for advancing quantum batteries, and bring new perspectives on efficient quantum battery designs.

quant-ph↗

arXiv2020

Coherences and the thermodynamic uncertainty relation: Insights from quantum absorption refrigerators

The thermodynamic uncertainty relation, originally derived for classical Markov-jump processes, provides a trade-off relation between precision and dissipation, deepening our understanding of the performance of quantum thermal machines. Here, we examine the interplay of quantum system coherences and heat current fluctuations on the validity of the thermodynamics uncertainty relation in the quantum regime. To achieve the current statistics, we perform a full counting statistics simulation of the Redfield quantum master equation. We focus on steady-state quantum absorption refrigerators where nonzero coherence between eigenstates can either suppress or enhance the cooling power, compared with the incoherent limit. In either scenario, we find enhanced relative noise of the cooling power (standard deviation of the power over the mean) in the presence of system coherence, thereby corroborating the thermodynamic uncertainty relation. Our results indicate that fluctuations necessitate consideration when assessing the performance of quantum coherent thermal machines.

cond-mat.stat-mech↗

arXiv2020

Quantum Nondemolition Photon Counting With a Hybrid Electromechanical Probe

Quantum nondemolition (QND) measurements of photons is a much pursued endeavor in the field of quantum optics and quantum information processing. Here we propose a novel hybrid optoelectromechanical platform that integrates a cavity system with a hybrid electromechanical probe for QND photon counting. Building upon a mechanical-mode-mediated nonperturbative electro-optical dispersive coupling, our protocol performs the QND photon counting measurement by means of the current-voltage characteristics of the probe. In particular, we show that the peak voltage shift of the differential conductance is linearly dependent on the photon occupation number, thus providing a sensitive measure of the photon number, especially in the strong optomechanical coupling regime. Given that our proposed hybrid system is compatible with state-of-the-art experimental techniques, we discuss its implementations and anticipate applications in quantum optics and polariton physics.

cond-mat.mes-hall↗

arXiv2020

BAMSProd: A Step towards Generalizing the Adaptive Optimization Methods to Deep Binary Model

Recent methods have significantly reduced the performance degradation of Binary Neural Networks (BNNs), but guaranteeing the effective and efficient training of BNNs is an unsolved problem. The main reason is that the estimated gradients produced by the Straight-Through-Estimator (STE) mismatches with the gradients of the real derivatives. In this paper, we provide an explicit convex optimization example where training the BNNs with the traditionally adaptive optimization methods still faces the risk of non-convergence, and identify that constraining the range of gradients is critical for optimizing the deep binary model to avoid highly suboptimal solutions. For solving above issues, we propose a BAMSProd algorithm with a key observation that the convergence property of optimizing deep binary model is strongly related to the quantization errors. In brief, it employs an adaptive range constraint via an errors measurement for smoothing the gradients transition while follows the exponential moving strategy from AMSGrad to avoid errors accumulation during the optimization. The experiments verify the corollary of theoretical convergence analysis, and further demonstrate that our optimization method can speed up the convergence about 1:2x and boost the performance of BNNs to a significant level than the specific binary optimizer about 3:7%, even in a highly non-convex optimization problem.

cs.CV↗

arXiv2020

QuantNet: Learning to Quantize by Learning within Fully Differentiable Framework

Despite the achievements of recent binarization methods on reducing the performance degradation of Binary Neural Networks (BNNs), gradient mismatching caused by the Straight-Through-Estimator (STE) still dominates quantized networks. This paper proposes a meta-based quantizer named QuantNet, which utilizes a differentiable sub-network to directly binarize the full-precision weights without resorting to STE and any learnable gradient estimators. Our method not only solves the problem of gradient mismatching, but also reduces the impact of discretization errors, caused by the binarizing operation in the deployment, on performance. Generally, the proposed algorithm is implemented within a fully differentiable framework, and is easily extended to the general network quantization with any bits. The quantitative experiments on CIFAR-100 and ImageNet demonstrate that QuantNet achieves the signifficant improvements comparing with previous binarization methods, and even bridges gaps of accuracies between binarized models and full-precision models.

cs.CV↗

arXiv2020

Sharp negative differential resistance from vibrational mode softening in molecular junctions

We unravel the critical role of vibrational mode softening in single-molecule electronic devices at high bias. Our theoretical analysis is carried out with a minimal model for molecular junctions, with mode softening arising due to quadratic electron-vibration couplings, and by developing a mean-field approach. We discover that the negative sign of the quadratic electron-vibration coupling coefficient can realize at high voltage a sharp negative differential resistance (NDR) effect with a large peak-to-valley ratio. Calculated current-voltage characteristics, obtained based on ab initio parameters for a nitro-substituted oligo(phenylene ethynylene) junction, agree very well with measurements. Our results establish that vibrational mode softening is a crucial effect at high voltage, underlying NDR, a substantial diode effect, and the breakdown of current-carrying molecular junctions.

cond-mat.mes-hall↗

arXiv2020

Dynamic Sparse Training: Find Efficient Sparse Network From Scratch With Trainable Masked Layers

We present a novel network pruning algorithm called Dynamic Sparse Training that can jointly find the optimal network parameters and sparse network structure in a unified optimization process with trainable pruning thresholds. These thresholds can have fine-grained layer-wise adjustments dynamically via backpropagation. We demonstrate that our dynamic sparse training algorithm can easily train very sparse neural network models with little performance loss using the same number of training epochs as dense models. Dynamic Sparse Training achieves the state of the art performance compared with other sparse training algorithms on various network architectures. Additionally, we have several surprising observations that provide strong evidence for the effectiveness and efficiency of our algorithm. These observations reveal the underlying problems of traditional three-stage pruning algorithms and present the potential guidance provided by our algorithm to the design of more compact network architectures.

cs.LG↗

arXiv2020

Dissipation-engineering of nonreciprocal quantum dot circuits: An input-output approach

Nonreciprocal effects in nanoelectronic devices offer unique possibilities for manipulating electron transport and engineering quantum electronic circuits for information processing purposes. However, a lack of rigorous theoretical tools is hindering this development. Here, we provide a general input-output description of nonreciprocal transport in solid-state quantum dot architectures, based on quantum optomechanical analogs. In particular, we break reciprocity between coherently-coupled quantum dots by dissipation-engineering in which these (so-called) primary dots are mutually coupled to auxiliary, damped quantum dots. We illustrate the general framework in two representative multiterminal noninteracting models, which can be used as building blocks for larger circuits. Importantly, the identified optimal conditions for nonreciprocal behavior hold even in the presence of additional dissipative effects that result from local electron-phonon couplings. Besides the analysis of the scattering matrix, we show that a nonreciprocal coupling induces unidirectional electron flow in the resonant transport regime. Altogether, our analysis provides the formalism and working principles towards the realization of nonreciprocal nanoelectronic devices.

quant-ph↗

arXiv2019

Generalized input-output method: A new route to quantum transport junctions

In this work, we put forward a generalized input-output method (GIOM) for studying charge transport in molecular junctions accounting for strong electron-vibration interactions and including electronic and phononic environments. The method radically expands the scope of the input-output theory, which was originally proposed to treat quantum optic problems. Based on the GIOM we derive a Langevin-type equation of motion for molecular operators, which posses a great generality and accuracy, and permits the derivation of a stationary charge current expression involving only two types of transfer rates. Furthermore, we devise the so-called "Polaron Transport in Electronic Resonance" (PoTER) approximation, which allows to feasibly simulate electron dynamics in generic tight-binding models with strong electron-vibration interactions. For short chains, the charge current reduces to known limits and reasonably agrees with exact numerical simulations. For extended junctions the current displays a turnover from phonon-assisted to phonon-suppressed transport. Nevertheless, the onset of ohmic behavior requires extensions beyond the PoTER approximation. As an additional application, we consider a cavity-coupled molecule junction. Here we identify a cavity-induced suppression of charge current in the single-site case, and observe signatures of polariton formation in the current-voltage characteristics in the strong light-matter coupling regime. A critical understanding gained from the GIOM-PoTER scheme is that the single-site vibrationally-coupled model is deceptively simpler, and amenable to approximations than multi-site models. Therefore, benchmarking of methods should not be concluded with the single-site case. The work manifests that the input-output framework, which is normally employed in quantum optics, can serve as a powerful and feasible tool in the realm of electron transport junctions.

quant-ph↗

arXiv2019

Knowledge Representing: Efficient, Sparse Representation of Prior Knowledge for Knowledge Distillation

Despite the recent works on knowledge distillation (KD) have achieved a further improvement through elaborately modeling the decision boundary as the posterior knowledge, their performance is still dependent on the hypothesis that the target network has a powerful capacity (representation ability). In this paper, we propose a knowledge representing (KR) framework mainly focusing on modeling the parameters distribution as prior knowledge. Firstly, we suggest a knowledge aggregation scheme in order to answer how to represent the prior knowledge from teacher network. Through aggregating the parameters distribution from teacher network into more abstract level, the scheme is able to alleviate the phenomenon of residual accumulation in the deeper layers. Secondly, as the critical issue of what the most important prior knowledge is for better distilling, we design a sparse recoding penalty for constraining the student network to learn with the penalized gradients. With the proposed penalty, the student network can effectively avoid the over-regularization during knowledge distilling and converge faster. The quantitative experiments exhibit that the proposed framework achieves the state-ofthe-arts performance, even though the target network does not have the expected capacity. Moreover, the framework is flexible enough for combining with other KD methods based on the posterior knowledge.

cs.CV↗

arXiv2019

DupNet: Towards Very Tiny Quantized CNN with Improved Accuracy for Face Detection

Deploying deep learning based face detectors on edge devices is a challenging task due to the limited computation resources. Even though binarizing the weights of a very tiny network gives impressive compactness on model size (e.g. 240.9 KB for IFQ-Tinier-YOLO), it is not tiny enough to fit in the embedded devices with strict memory constraints. In this paper, we propose DupNet which consists of two parts. Firstly, we employ weights with duplicated channels for the weight-intensive layers to reduce the model size. Secondly, for the quantization-sensitive layers whose quantization causes notable accuracy drop, we duplicate its input feature maps. It allows us to use more weights channels for convolving more representative outputs. Based on that, we propose a very tiny face detector, DupNet-Tinier-YOLO, which is 6.5X times smaller on model size and 42.0% less complex on computation and meanwhile achieves 2.4% higher detection than IFQ-Tinier-YOLO. Comparing with the full precision Tiny-YOLO, our DupNet-Tinier-YOLO gives 1,694.2X and 389.9X times savings on model size and computation complexity respectively with only 4.0% drop on detection rate (0.880 vs. 0.920). Moreover, our DupNet-Tinier-YOLO is only 36.9 KB, which is the tiniest deep face detector to our best knowledge.

cs.CV↗

arXiv2019

Unusual Transport Properties with Non-Commutative System-Bath Coupling Operators

Understanding non-equilibrium heat transport is crucial for controling heat flow in nano-scale systems. We study thermal energy transfer in a generalized non-equilibrium spin-boson model (NESB) with non-commutative system-bath coupling operators and discover unusual transport properties. Compared to the conventional NESB, the heat current is greatly enhanced by rotating the coupling operators. Constructive contribution to thermal rectification can be optimized when two sources of asymmetry, system-bath coupling strength and coupling operators, coexist. At the weak coupling and the adiabatic limit, the scaling dependence of heat current on the coupling strength and the system energy gap changes drastically when the coupling operators become non-commutative. These scaling relations can further be explained analytically by the non-equilibrium polaron-transformed Redfield equation. These novel transport properties, arising from the pure quantum effect of non-commutative coupling operators, should generally appear in other non-equilibrium set-ups and driven-systems.

cond-mat.stat-mech↗

arXiv2019

A Nonequilibrium Variational Polaron Theory to Study Quantum Heat Transport

We propose a nonequilibrium variational polaron transformation, based on an ansatz for nonequilibrium steady state (NESS) with an effective temperature, to study quantum heat transport at the nanoscale. By combining the variational polaron transformed master equation with the full counting statistics, we have extended the applicability of the polaron-based framework to study nonequilibrium process beyond the super-Ohmic bath models. Previously, the polaron-based framework for quantum heat transport reduces exactly to the non-interacting blip approximation (NIBA) formalism for Ohmic bath models due to the issue of the infrared divergence associated with the full polaron transformation. The nonequilibrium variational method allows us to appropriately treat the infrared divergence in the low-frequency bath modes and explicitly include cross-bath correlation effects. These improvements provide more accurate calculation of heat current than the NIBA formalism for Ohmic bath models. We illustrate the aforementioned improvements with the nonequilibrium spin-boson model in this work and quantitatively demonstrate the cross-bath correlation, current turnover, and rectification effects in quantum heat transfer.

physics.chem-ph↗

arXiv2019

Thermodynamic uncertainty relation in quantum thermoelectric junctions

Recently, a thermodynamic uncertainty relation (TUR) has been formulated for classical Markovian systems demonstrating trade-off between precision (current fluctuation) and cost (dissipation). Systems that violate the TUR are interesting as they overcome another trade-off relation concerning the efficiency of a heat engine, its power, and its stability (power fluctuations). Here, we analyze the root, extent, and impact on performance of TUR violations in quantum thermoelectric junctions at steady state. Considering noninteracting electrons, first we show that only the "classical" component of the current noise, arising from single-electron transfer events follows the TUR. The remaining, "quantum" part of current noise is therefore responsible for the potential violation of TUR in such quantum systems. Next, focusing on the resonant transport regime we determine the parameter range in which the violation of the TUR can be observed---for both voltage-biased junctions and thermoelectric engines. We illustrate our findings with exact numerical simulations of a serial double quantum dot system. Most significantly, we demonstrate that the TUR always holds in noninteracting thermoelectric generators when approaching the thermodynamic efficiency limit.

cond-mat.mes-hall↗

arXiv2018

Electric field control of spins in molecular magnets

Coherent control of individual molecular spins in nano-devices is a pivotal prerequisite for fulfilling the potential promised by molecular spintronics. By applying electric field pulses during time-resolved electron spin resonance measurements, we measure the sensitivity of the spin in several antiferromagnetic molecular nanomagnets to external electric fields. We find a linear electric field dependence of the spin states in Cr$_7$Mn, an antiferromagnetic ring with a ground-state spin of $S=1$, and in a frustrated Cu$_3$ triangle, both with coefficients of about $2~\mathrm{rad}\, \mathrm{s}^{-1} / \mathrm{V} \mathrm{m}^{-1}$. Conversely, the antiferromagnetic ring Cr$_7$Ni, isomorphic with Cr$_7$Mn but with $S=1/2$, does not exhibit a detectable effect. We propose that the spin-electric field coupling may be used for selectively controlling individual molecules embedded in nanodevices.

cond-mat.str-el↗

arXiv2018

Hybrid quantum-classical simulation of quantum speed limits in open quantum systems

The quantum speed limit (QSL) provides a fundamental upper bound on the speed of quantum evolution, but its evaluation in generic open quantum systems still presents a formidable computational challenge. Herein, we introduce a hybrid quantum-classical method for computing QSL times in multi-level open quantum systems. The method is based on a mixed Wigner-Heisenberg representation of the composite quantum dynamics, in which the open subsystem of interest is treated quantum mechanically and the bath is treated in a classical-like fashion. By solving a set of coupled first-order deterministic differential equations for the quantum and classical degrees of freedom, one can compute the QSL time. To demonstrate the utility of the method, we study the unbiased spin-boson model and provide a detailed analysis of the effect of the subsystem-bath coupling strength and bath temperature on the QSL time. In particular, we find a turnover of the QSL time in the strong coupling regime, which is indicative of a speed-up in the quantum evolution. We also apply the method to the Fenna- Matthews-Olson complex model and identify a potential connection between the QSL time and the efficiency of the excitation energy transfer at different temperatures.

quant-ph↗

arXiv2018

Heat transfer statistics in mixed quantum-classical systems

The modelling of quantum heat transfer processes at the nanoscale is crucial for the development of energy harvesting and molecular electronics devices. Herein, we adopt a mixed quantum-classical description of a device, in which the open subsystem of interest is treated quantum mechanically and the surrounding heat baths are treated in a classical-like fashion. By introducing such a mixed quantum-classical description of the composite system, one is able to study the heat transfer between the subsystem and bath from a closed system point of view, thereby avoiding simplifying assumptions related to the bath time scale and subsystem-bath coupling strength. In particular, we adopt the full counting statistics approach to derive a general expression for the moment generating function of heat in systems whose dynamics are described by the quantum-classical Liouville equation (QCLE). From this expression, one can deduce expressions for the dynamics of the average heat and heat current, which may be evaluated using numerical simulations. Due to the approximate nature of the QCLE, we also find that the steady state fluctuation symmetry holds up to order $\hbar$ for systems whose subsystembath couplings and baths go beyond bilinear and harmonic, respectively. To demonstrate the approach, we consider the nonequilibrium spin boson model and simulate its time-dependent average heat and heat current under various conditions.

physics.chem-ph↗