arXiv subjects

Deng Pan

Publications and source records attributed to Deng Pan.

46 records · Page 3Linked to original sources

arXiv2020

On the Learning Property of Logistic and Softmax Losses for Deep Neural Networks

Deep convolutional neural networks (CNNs) trained with logistic and softmax losses have made significant advancement in visual recognition tasks in computer vision. When training data exhibit class imbalances, the class-wise reweighted version of logistic and softmax losses are often used to boost performance of the unweighted version. In this paper, motivated to explain the reweighting mechanism, we explicate the learning property of those two loss functions by analyzing the necessary condition (e.g., gradient equals to zero) after training CNNs to converge to a local minimum. The analysis immediately provides us explanations for understanding (1) quantitative effects of the class-wise reweighting mechanism: deterministic effectiveness for binary classification using logistic loss yet indeterministic for multi-class classification using softmax loss; (2) disadvantage of logistic loss for single-label multi-class classification via one-vs.-all approach, which is due to the averaging effect on predicted probabilities for the negative class (e.g., non-target classes) in the learning process. With the disadvantage and advantage of logistic loss disentangled, we thereafter propose a novel reweighted logistic loss for multi-class classification. Our simple yet effective formulation improves ordinary logistic loss by focusing on learning hard non-target classes (target vs. non-target class in one-vs.-all) and turned out to be competitive with softmax loss. We evaluate our method on several benchmark datasets to demonstrate its effectiveness.

cs.LG↗

arXiv2020

Improve SGD Training via Aligning Mini-batches

Deep neural networks (DNNs) for supervised learning can be viewed as a pipeline of a feature extractor (i.e. last hidden layer) and a linear classifier (i.e. output layer) that is trained jointly with stochastic gradient descent (SGD). In each iteration of SGD, a mini-batch from the training data is sampled and the true gradient of the loss function is estimated as the noisy gradient calculated on this mini-batch. From the feature learning perspective, the feature extractor should be updated to learn meaningful features with respect to the entire data, and reduce the accommodation to noise in the mini-batch. With this motivation, we propose In-Training Distribution Matching (ITDM) to improve DNN training and reduce overfitting. Specifically, along with the loss function, ITDM regularizes the feature extractor by matching the moments of distributions of different mini-batches in each iteration of SGD, which is fulfilled by minimizing the maximum mean discrepancy. As such, ITDM does not assume any explicit parametric form of data distribution in the latent feature space. Extensive experiments are conducted to demonstrate the effectiveness of our proposed strategy.

cs.LG↗

arXiv2019

Drive-specific adaptation in disordered mechanical networks of bistable springs

Systems with many stable configurations abound in nature, both in living and inanimate matter. Their inherent nonlinearity and sensitivity to small perturbations make them challenging to study, particularly in the presence of external driving, which can alter the relative stability of different attractors. Under such circumstances, one may ask whether any clear relationship holds between the specific pattern of external driving and the particular attractor states selected by a driven multistable system. To gain insight into this question, we numerically study driven disordered mechanical networks of bistable springs which possess a vast number of stable configurations arising from the two stable rest lengths of each spring, thereby capturing the essential physical properties of a broad class of multistable systems. We find that the attractor states of driven disordered multistable mechanical networks are fine-tuned with respect to the pattern of external forcing to have low work absorption from it. Furthermore, we find that these drive-specific attractor states are even more stable than expected for a given level of work absorption. Our results suggest that the driven exploration of the vast configuration space of these systems is biased towards states with exceptional relationship to the driving environment, and could therefore be used to `discover' states with desired response properties in systems with a vast landscape of diverse configurations.

nlin.AO↗

arXiv2019

Rotational Doppler cooling and heating

Doppler cooling is a widely used technique to laser cool atoms and nanoparticles exploiting the Doppler shift involved in translational transformations. The rotational Doppler effect arising from rotational coordinate transformations should similarly enable optical manipulations of the rotational degrees of freedom in rotating nanosystems. Here, we show that rotational Doppler cooling and heating (RDC and RDH) effects embody rich and unexplored physics, such as a strong dependence on particle morphology. For geometrically confined particles, such as a nanorod that can represent diatomic molecules, RDC and RDH follow similar rules as their translational Doppler counterpart, where cooling and heating are always observed at red- or blue-detuned laser frequencies, respectively. Surprisingly, nanosystems that can be modeled as a solid particle shows a strikingly different response, where RDH appears in a frequency regime close to their resonances, while a detuned frequency produces cooling of rotation. We also predict that the RDH effect can lead to unprecedented spontaneous chiral symmetry breaking, whereby an achiral particle under linearly polarized illumination starts spontaneously rotating, rendering it nontrivial compared to the translational Doppler effect. Our results open up new exciting possibilities to control the rotational motion of molecules and nanoparticles.

physics.optics↗

arXiv2019

Magnetically Activated Rotational Vacuum Friction

We predict the existence of a torque acting on an isotropic neutral nanosphere activated by a static magnetic field when the particle temperature differs from the surrounding vacuum. This phenomenon originates in time-reversal symmetry breaking of the particle interaction with the vacuum electromagnetic field. We present a rigorous quantum treatment of photons and particle excitations that leads to a nonzero torque even in a motionless particle. We also find that the dynamical evolution of the particle temperature and rotation frequency follow an exotic dynamics, including spontaneous changes in the rotation direction. Magnetically activated thermal vacuum torques open a unique avenue for the investigation of the effect of time-reversal symmetry-breaking in thermal and Casimir physics.

cond-mat.mes-hall↗

arXiv2019

Circular Dichroism in Rotating Particles

Light interaction with rotating nanostructures gives rise to phenemona as varied as optical torques and quantum friction. Here we reveal that circular dichroism of rotating optically-isotropic particles has an unexpectedly strong dependence on their internal geometry. In particular, nanorings and nanocrosses exhibit a splitting of $2Ω$ in the particle optical resonances, while compact particles display weak circular dichroism at low rotation frequency $Ω$, but a strong circular dichroism at high $Ω$. We base our findings on a quantum-mechanical description of the polarizability of rotating particles, which has not been rigorously addressed so far. Specifically, we use the random-phase approximation and populate the particle electronic states according to the principle that they are thermally equilibrated in the rotating frame. We further provide insight into the rotational superradience effect and the ensuing optical gain, originating in population inversion as regarded from the lab frame, in which the particle is out of equilibrium. Surprisingly, we find the optical frequency cutoff for superradiance to deviate from the rotation frequency $Ω$. Our results unveil a rich, unexplored phenomenology of light interaction with rotating objects.

cond-mat.mes-hall↗

arXiv2018

Digital Biosensing By Foundry-Fabricated Graphene Sensors

Biomedical and environmental testing currently require access to highly specialized facilities or extensive training. Current biochemical tests require complex reagents to achieve a simple human readable result such as a color change. This results in low-information tests, with high reagent costs. To maximize information from testing while minimizing costs, biotechnology should leverage complex analysis enabled by advanced portable computing power and use simplified reagents. Nanotechnologists have created proof-of-concept devices to move diagnosis and monitoring from large core labs to hand-held, easy-to-use personal devices, but none have achieved large-scale manufacturability. We demonstrate the first economical nanoelectronics produced in a commercial foundry. The mass-manufacture of these graphene-based digital biosensors is achieved by successfully integrating delicate graphene material into standard electronic high-volume production processes. We describe critical manufacturing parameters and the results of biosensing measurements that establish graphene biosensors as a portable, low-cost method to gain precise data without labeled, complex reagents. The low power and resource usage of these biosensors enable individuals to gain personal, immediate control over precise biological and environmental data.

physics.app-ph↗

arXiv2017

Topologically protected Dirac plasmons in graphene

Topological optical states exhibit unique immunity to defects and the ability to propagate without losses rendering them ideal for photonic applications.A powerful class of such states is based on time-reversal symmetry breaking of the optical response.However, existing proposals either involve sophisticated and bulky structural designs or can only operate in the microwave regime. Here, we propose and provide a theoretical proof-of-principle demonstration for highly confined topologically protected optical states to be realized at infrared frequencies in a simple 2D material structure-a periodically patterned graphene monolayer-subject to a magnetic field below 1 tesla. In our graphene honeycomb superlattice structures plasmons exhibit substantial nonreciprocal behavior at the superlattice junctions, leading to the emergence of topologically protected edge states and localized bulk modes enabled by the strong magneto-optical response of this material, which leads to time-reversal-symmetry breaking already at moderate static magnetic fields. The proposed approach is simple and robust for realizing topologically nontrivial 2D optical states, not only in graphene, but also in other 2D atomic layers, and could pave the way for realizing fast, nanoscale, defect-immune devices for integrated photonics applications.

physics.optics↗

arXiv2017

Gravitational field around blackhole induces photonic spin-orbit interaction that twists light

The spin-orbit interaction (SOI) of light has been intensively studied in nanophotonics because it enables sensitive control of photons' spin degree of freedom and thereby the trajectory of the photons, which is useful for applications such as signal encoding and routing. A recent study showed that the SOI of photons manifests in the presence of a gradient in the permittivity of the medium through which the photons propagate; this enhances the scattering of circularly polarized light and results in the photons propagating along twisted trajectories. Here we theoretically predict that, because of the equivalence between an inhomogeneous dielectric medium and a gravitational field demonstrated in transformation optics, a significant SOI is induced onto circularly polarized light passing by the gravitational lens of a black hole. This leads to: i) the photons to propagate along chiral trajectories if the size of the blackhole is smaller than the wavelength of the incident photons, ii) the resulting image of the gravitational lens to manifest an azimuthal rotation because of these chiral trajectories. The findings open for a way to probe for and discover subwavelength-size black holes using circularly polarized light.

gr-qc↗

arXiv2010

Buffer Management Algorithm Design and Implementation Based on Network Processors

To solve the parameter sensitive issue of the traditional RED (random early detection) algorithm, an adaptive buffer management algorithm called PAFD (packet adaptive fair dropping) is proposed. This algorithm supports DiffServ (differentiated services) model of QoS (quality of service). In this algorithm, both of fairness and throughput are considered. The smooth buffer occupancy rate function is adopted to adjust the parameters. By implementing buffer management and packet scheduling on Intel IXP2400, the viability of QoS mechanisms on NPs (network processors) is verified. The simulation shows that the PAFD smoothes the flow curve, and achieves better balance between fairness and network throughput. It also demonstrates that this algorithm meets the requirements of fast data packet processing, and the hardware resource utilization of NPs is higher.

cs.NI↗