Depthwise separable convolutions and frequency-domain convolutions are two recent ideas for building efficient convolutional neural networks. They are seemingly incompatible: the vast majority of operations in depthwise separable CNNs are in pointwis
e convolutional layers, but pointwise layers use 1x1 kernels, which do not benefit from frequency transformation. This paper unifies these two ideas by transforming the activations, not the kernels. Our key insights are that 1) pointwise convolutions commute with frequency transformation and thus can be computed in the frequency domain without modification, 2) each channel within a given layer has a different level of sensitivity to frequency domain pruning, and 3) each channels sensitivity to frequency pruning is approximately monotonic with respect to frequency. We leverage this knowledge by proposing a new technique which wraps each pointwise layer in a discrete cosine transform (DCT) which is truncated to selectively prune coefficients above a given threshold as per the needs of each channel. To learn which frequencies should be pruned from which channels, we introduce a novel learned parameter which specifies each channels pruning threshold. We add a new regularization term which incentivizes the model to decrease the number of retained frequencies while still maintaining task accuracy. Unlike weight pruning techniques which rely on sparse operators, our contiguous frequency band pruning results in fully dense computation. We apply our technique to MobileNetV2 and in the process reduce computation time by 22% and incur <1% accuracy degradation.
For both visible and infrared images have their own advantages and disadvantages, RGBT tracking has attracted more and more attention. The key points of RGBT tracking lie in feature extraction and feature fusion of visible and infrared images. Curren
t RGBT tracking methods mostly pay attention to both individual features (features extracted from images of a single camera) and common features (features extracted and fused from an RGB camera and a thermal camera), while pay less attention to the different and dynamic contributions of individual features and common features for different sequences of registered image pairs. This paper proposes a novel RGBT tracking method, called Dynamic Fusion Network (DFNet), which adopts a two-stream structure, in which two non-shared convolution kernels are employed in each layer to extract individual features. Besides, DFNet has shared convolution kernels for each layer to extract common features. Non-shared convolution kernels and shared convolution kernels are adaptively weighted and summed according to different image pairs, so that DFNet can deal with different contributions for different sequences. DFNet has a fast speed, which is 28.658 FPS. The experimental results show that when DFNet only increases the Mult-Adds of 0.02% than the non-shared-convolution-kernel-based fusion method, Precision Rate (PR) and Success Rate (SR) reach 88.1% and 71.9% respectively.
Dynamics of ethylene autoignition and Deflagration-to-Detonation Transition (DDT) in a one-dimensional shock tube are numerically investigated using a skeletal chemistry including 10 species and 10 reactions. Different combustion modes are investigat
ed through considering various premixed gas equivalence ratios (0.2 to 2.0) and incident shock wave Mach numbers (1.8 to 3.2). Four ignition and DDT modes are observed from the studied cases, i.e., no ignition, deflagration combustion, detonation after reflected shock and deflagration behind the incident shock. For detonation development behind the reflected shock, three autoignition hot spots are formed. The first one occurs at the wall surface after the re-compression of the reflected shock and contact surface, which further develops to a reaction shock because of the explosion in the explosion regime. The other two are off the wall, respectively caused by the reflected shock rarefaction wave interaction and reaction induction in the compressed mixture. The last hot spot develops to a reaction wave and couples with the reflected shock after a DDT process, which eventually leads to detonation combustion. For deflagration development behind the reflected shock, the wave interactions, wall surface autoignition hot spot as well as its induction of reaction shock are qualitatively similar to the mode of detonation after incident shock reflection, before the reflected shock rarefaction wave collision point. However, only one hot spot is induced after the collision, which also develops to a reaction wave but cannot catch up with the reflected shock. For deflagration behind the incident shock, deflagration combustion is induced by the incident shock compression whereas detonation occurs after the shock reflection.
The discovery of two-dimensional (2D) ferroelectrics with switchable out-of-plane polarization such as monolayer $alpha$-In$_2$Se$_3$ offers a new avenue for ultrathin high-density ferroelectric-based nanoelectronics such as ferroelectric field effec
t transistors and memristors. The functionality of ferroelectrics depends critically on the dynamics of polarization switching in response to an external electric/stress field. Unlike the switching dynamics in bulk ferroelectrics that have been extensively studied, the mechanisms and dynamics of polarization switching in 2D remain largely unexplored. Molecular dynamics (MD) using classical force fields is a reliable and efficient method for large-scale simulations of dynamical processes with atomic resolution. Here we developed a deep neural network-based force field of monolayer In$_2$Se$_3$ using a concurrent learning procedure that efficiently updates the first-principles-based training database. The model potential has accuracy comparable with density functional theory (DFT), capable of predicting a range of thermodynamic properties of In$_2$Se$_3$ polymorphs and lattice dynamics of ferroelectric In$_2$Se$_3$. Pertinent to the switching dynamics, the model potential also reproduces the DFT kinetic pathways of polarization reversal and 180$^circ$ domain wall motions. Moreover, isobaric-isothermal ensemble MD simulations predict a temperature-driven $alpha rightarrow beta$ phase transition at the single-layer limit, as revealed by both local atomic displacement and Steinhardts bond orientational order parameter $Q_4$. Our work paves the way for further research on the dynamics of ferroelectric $alpha$-In$_2$Se$_3$ and related systems.
The inventories carried in a supply chain as a strategic tool to influence the competing firms are considered to be strategic inventories (SI). We present a two-period game-theoretic supply chain model, in which a singular manufacturer supplies produ
cts to a pair of identical Cournot duopolistic retailers. We show that the SI carried by the retailers under dynamic contract is Pareto-dominating for the manufacturer, retailers, consumers, the channel, and the society as well. We also find that retailers SI, however, can be eliminated when the manufacturer commits wholesale contract or inventory holding cost is too high. In comparing the cases with and without downstream competition, we also show that the downstream Cournot duopoly undermines the profits for the retailers, but benefits all others.
While Deep Reinforcement Learning (DRL) provides transformational capabilities to the control of Robotics and Autonomous Systems (RAS), the black-box nature of DRL and uncertain deployment-environments of RAS pose new challenges on its dependability.
Although there are many existing works imposing constraints on the DRL policy to ensure a successful completion of the mission, it is far from adequate in terms of assessing the DRL-driven RAS in a holistic way considering all dependability properties. In this paper, we formally define a set of dependability properties in temporal logic and construct a Discrete-Time Markov Chain (DTMC) to model the dynamics of risk/failures of a DRL-driven RAS interacting with the stochastic environment. We then do Probabilistic Model Checking based on the designed DTMC to verify those properties. Our experimental results show that the proposed method is effective as a holistic assessment framework, while uncovers conflicts between the properties that may need trade-offs in the training. Moreover, we find the standard DRL training cannot improve dependability properties, thus requiring bespoke optimisation objectives concerning them. Finally, our method offers a novel dependability analysis to the Sim-to-Real challenge of DRL.
Procedural modeling is now the de facto standard of material modeling in industry. Procedural models can be edited and are easily extended, unlike pixel-based representations of captured materials. In this paper, we present a semi-automatic pipeline
for general material proceduralization. Given Spatially-Varying Bidirectional Reflectance Distribution Functions (SVBRDFs) represented as sets of pixel maps, our pipeline decomposes them into a tree of sub-materials whose spatial distributions are encoded by their associated mask maps. This semi-automatic decomposition of material maps progresses hierarchically, driven by our new spectrum-aware material matting and instance-based decomposition methods. Each decomposed sub-material is proceduralized by a novel multi-layer noise model to capture local variations at different scales. Spatial distributions of these sub-materials are modeled either by a by-example inverse synthesis method recovering Point Process Texture Basis Functions (PPTBF) or via random sampling. To reconstruct procedural material maps, we propose a differentiable rendering-based optimization that recomposes all generated procedures together to maximize the similarity between our procedural models and the input material pixel maps. We evaluate our pipeline on a variety of synthetic and real materials. We demonstrate our methods capacity to process a wide range of material types, eliminating the need for artist designed material graphs required in previous work. As fully procedural models, our results expand to arbitrary resolution and enable high level user control of appearance.
In this paper, we propose to use the general $L^2$-based Sobolev norms (i.e., $H^s$ norms, $sin mathbb{R}$) to measure the data discrepancy due to noise in image processing tasks that are formulated as optimization problems. As opposed to a popular t
rend of developing regularization methods, we emphasize that an textit{implicit} regularization effect can be achieved through the class of Sobolev norms as the data-fitting term. Specifically, we analyze that the implicit regularization comes from the weights that the $H^s$ norm imposes on different frequency contents of an underlying image. We also build the connections of such norms with the optimal transport-based metrics and the Sobolev gradient-based methods, leading to a better understanding of functional spaces/metrics and the optimization process involved in image processing. We use the fast Fourier transform to compute the $H^s$ norm efficiently and combine it with the total variation regularization in the framework of the alternating direction method of multipliers (ADMM). Numerical results in both denoising and deblurring support our theoretical findings.
In this paper, we propose a highly accurate continuous-time trajectory estimation framework dedicated to SLAM (Simultaneous Localization and Mapping) applications, which enables fuse high-frequency and asynchronous sensor data effectively. We apply t
he proposed framework in a 3D LiDAR-inertial system for evaluations. The proposed method adopts a non-rigid registration method for continuous-time trajectory estimation and simultaneously removing the motion distortion in LiDAR scans. Additionally, we propose a two-state continuous-time trajectory correction method to efficiently and efficiently tackle the computationally-intractable global optimization problem when loop closure happens. We examine the accuracy of the proposed approach on several publicly available datasets and the data we collected. The experimental results indicate that the proposed method outperforms the discrete-time methods regarding accuracy especially when aggressive motion occurs. Furthermore, we open source our code at url{https://github.com/APRIL-ZJU/clins} to benefit research community.
In this work, we are concerned with a Fokker-Planck equation related to the nonlinear noisy leaky integrate-and-fire model for biological neural networks which are structured by the synaptic weights and equipped with the Hebbian learning rule. The eq
uation contains a small parameter $varepsilon$ separating the time scales of learning and reacting behavior of the neural system, and an asymptotic limit model can be derived by letting $varepsilonto 0$, where the microscopic quasi-static states and the macroscopic evolution equation are coupled through the total firing rate. To handle the endowed flux-shift structure and the multi-scale dynamics in a unified framework, we propose a numerical scheme for this equation that is mass conservative, unconditionally positivity preserving, and asymptotic preserving. We provide extensive numerical tests to verify the schemes properties and carry out a set of numerical experiments to investigate the models learning ability, and explore the solutions behavior when the neural network is excitatory.
