Current state-of-the-art object detectors are at the expense of high computational costs and hard to deploy low-end devices. Knowledge distillation, which aims training a smaller student network by transferring knowledge from larger teacher model, is one promising solutions for model miniaturization. In this paper, we investigate each module typical detector in depth, propose general distillation framework that adaptively transfers according task specific priors. The intuition simply...

Terrain traversability analysis is a fundamental issue to achieve the autonomy of robot at off-road environments. Geometry-based and appearance-based methods have been studied in decades, while behavior-based exploiting learning from demonstration (LfD) are new trends. Behavior-based learn cost functions that guide trajectory planning compliance with experts' demonstrations, which can be more scalable various scenes driving behaviors. This research proposes method using Deep Maximum Entropy...

This paper introduces an innovative spectrum monitoring setup and a new performance metric, airtime utilization, which quantifies the extent of usage. Our design offers protocol-independent solution for detailed analysis across frequency, time, power dimensions, is essential effective management. The presented experimental monitor cost-effective large-scale deployment. Extensive measurements in 3.1 to 3.7 GHz frequency range are reported demonstrate practical application our setup....

Large language models rely on Supervised Fine-Tuning (SFT) to specialize in downstream tasks. Cross Entropy (CE) loss is the de facto choice SFT, but it often leads overfitting and limited output diversity due its aggressive updates data distribution. This paper aim address these issues by introducing maximum entropy principle, which favors with flatter distributions that still effectively capture data. Specifically, we develop a new distribution matching method called GEM, solves reverse...

Reinforcement Learning algorithms are widely applied in many fields, such as price index prediction, image recognition, and natural language processing. This paper introduces a novel algorithm based on the classical Deep Residual Shrinkage Neural Network for portfolio management. In this algorithm, Ensemble of Identical Independent Evaluators framework put forward by Jiang et al. is adopted policy function. Following this, we adopt to function identical independent evaluator optimize...

Autonomous driving is one of the current cutting edge technologies. For autonomous cars, their actions and trajectories should not only achieve autonomy safety, but also obey human drivers' behavior patterns, when sharing roads with other drivers on highway. Traditional methods, though robust interpretable, demands much labor in engineering complex mapping from situation to vehicle's future control. newly developed deep-learning they can automatically learn such data fewer humans'...

Scaling of deep neural networks, especially Transformers, is pivotal for their surging performance and has further led to the emergence sophisticated reasoning capabilities in foundation models. Such scaling generally requires training large models from scratch with random initialization, failing leverage knowledge acquired by smaller counterparts, which are already resource-intensive obtain. To tackle this inefficiency, we present $\textbf{L}$ossl$\textbf{E}$ss $\textbf{MO}$del...

Understanding of GAN training is still very limited. One major challenge its non-convex-non-concave min-max objective, which may lead to sub-optimal local minima. In this work, we perform a global landscape analysis the empirical loss GANs. We prove that class separable-GAN, including original JS-GAN, has exponentially many bad basins are perceived as mode-collapse. also study relativistic pairing (RpGAN) couples generated samples and true samples. RpGAN no basins. Experiments on synthetic...

Contrastive learning based on instance discrimination trains model to discriminate different transformations of the anchor sample from other samples, which does not consider semantic similarity among samples. This paper proposes a new kind contrastive method, named CLIM, uses positives samples in dataset. is achieved by searching local similar anchor, and selecting that are closer corresponding cluster center, we denote as center-wise image selection. The selected instantiated via an data...

Reinforcement learning algorithms are used in various fields widely, such as cryptocurrency market forecasting, image recognition, and natural language processing. In this research, we use the algorithm to solve portfolio management problem. algorithm, adopt squeeze-and-excitation neural network realize Ensemble of Identical Independent Evaluators proposed by Jiang et al. "The Squeeze-and-Excitation" block works adaptively recalibrating channel-wise feature responses, which improves ability...

This paper examines current controversies in the ethical problems of AI robots and discusses how to build good ethics trust computational intelligence, through lens philosophy science. By analyzing case study Xiaoice, this highlights currently existing robots. Machine is introduced as a framework address these issues examine feasibility imparting dimensions machines. The crux lies integrating human values applying machine creating ethically intelligent research suggests combining...

As a model-free algorithm, deep reinforcement learning (DRL) agent learns and makes decisions by interacting with the environment in an unsupervised way. In recent years, DRL algorithms have been widely applied scholars for portfolio optimization consecutive trading periods, since can dynamically adapt to market changes does not rely on specification of joint dynamics across assets. However, typical agents cannot learn policy that is aware dynamic correlation between asset returns. Since...

This paper introduces an innovative spectrum measurement setup and a new performance metric, airtime utilization, which quantifies the extent of usage. Our design offers protocol-independent solution for detailed analysis across frequency, time, power dimensions, is essential effective management. The presented experimental Wireless Community Testbed (WCT) cost-effective large-scale deployment. Extensive measurements in 3.1 to 3.7 GHz frequency range are reported demonstrate practical...

Reinforcement Learning algorithms are widely applied in many diverse fields, including portfolio management. Ensemble of Identical Independent Evaluators (EIIE) framework proposed by Jiang et al. achieved management based on their deep reinforcement learning algorithm. In the implementation EIIE framework, a neural network such as Convolutional Neural Network is policy network, to uncover more patterns data. However, this typology inefficient due its simple structure. To overcome shortcoming...

计算机视觉的任务目标是建立接近人类视觉系统的计算模型。随着深度神经网络（deep neural...

The objective of portfolio management is to realize optimization, i.e., maximizing the cumulative return over continuous trading periods. Using Artificial Intelligence algorithms, e.g., Deep Reinforcement Learning (DRL), optimization an emerging research trend. Jiang et al.'s Ensemble Identical Independent Evaluators (EIIE) framework achieves at least a four-fold improvement in indicator final value. Their has high flexibility allow us replace components achieve improvement. In EIIE, DRL...

Reinforcement learning algorithms have a wide range of applications in diverse areas, such as portfolio management, automatic driving, and visual object detection. This paper introduces novel network architecture Ghost expectation point (GXPT) embedded deep reinforcement framework based on GhostNet, which is constructed using convolutional neural networks ghost bottleneck modules. The module can generate many feature maps, improving the ability to extract information from real-world market....

针对双桥并联励磁功率单元的晶闸管开路故障，提出一种基于一维卷积神经网络(1D-convolutional neural networks，1D-CNN)和长短期记忆网络(long short-term memory，LSTM)混合模型的故障诊断方法。将1号整流桥共阴极侧、共阳极侧电流和AB相线电压构造时序特征向量作为输入，利用1D-CNN提取并重构样本空间特征；考虑到输入量本身是时间序列数据，采用LSTM网络进一步提取特征。根据特征向量与故障类别对应关系实现故障诊断。仿真结果表明，该模型能够有效地实现双桥并联励磁功率单元故障诊断，具有良好的抗噪能力。