Recent studies on few-shot classification using transfer learning pose challenges to the effectiveness and efficiency of episodic meta-learning algorithms. Transfer learning approaches are a natural ...alternative, but they are restricted to few-shot classification. Moreover, little attention has been on the development of probabilistic models with well-calibrated uncertainty from few-shot samples, except for some Bayesian episodic learning algorithms. To tackle the aforementioned issues, we propose a new transfer learning method to obtain accurate and reliable models for few-shot regression and classification. The resulting method does not require episodic meta-learning and is called meta-free representation learning (MFRL). MFRL first finds low-rank representation generalizing well on meta-test tasks. Given the learned representation, probabilistic linear models are fine-tuned with few-shot samples to obtain models with well-calibrated uncertainty. The proposed method not only achieves the highest accuracy on a wide range of few-shot learning benchmark datasets but also correctly quantifies the prediction uncertainty. In addition, weight averaging and temperature scaling are effective in improving the accuracy and reliability of few-shot learning in existing meta-learning algorithms with a wide range of learning paradigms and model architectures.
The concept of C-RAN is attractive for both academic and industry to bring innovative ideas for advanced wireless processing and reduce cost directly in mobile infrastructure. C-RAN could be one of ...direction of future mobile infrastructure, however, to achieve the goal of C-RAN is quite challenging, since it's domain specific cloud computing technology, in which real-time processing and huge amount of antenna data make it more difficult than traditional cloud computing. In the paper, together with the load analysis of C-RAN, one of GPP (e.g. x86) based C-RAN architecture is proposed targeted to scalable, large-scale BBU pool. The design philosophy is to efficiently reduce computation cost and power consumption by allocating computation/switch resource fitting the traffic need.
The transduction of sequence has been mostly done by recurrent networks, which are computationally demanding and often underestimate uncertainty severely. We propose a computationally efficient ...attention-based network combined with the Gaussian process regression to generate real-valued sequence, which we call the Attentive-GP. The proposed model not only improves the training efficiency by dispensing recurrence and convolutions but also learns the factorized generative distribution with Bayesian representation. However, the presence of the GP precludes the commonly used mini-batch approach to the training of the attention network. Therefore, we develop a block-wise training algorithm to allow mini-batch training of the network while the GP is trained using full-batch, resulting in a scalable training method. The algorithm has been proved to converge and shows comparable, if not better, quality of the found solution. As the algorithm does not assume any specific network architecture, it can be used with a wide range of hybrid models such as neural networks with kernel machine layers in the scarcity of resources for computation and memory.
Batch processes are characterized by inherent nonlinearity, multiplicity of operating phases, between-phase transient dynamics and batch-to-batch uncertainty that pose significant challenges for ...accurate state estimation and quality prediction. Conventional multi-model strategies, however, may be ill-suited for multiphase batch processes because the localized models do not specifically characterize the complex transient dynamics between two consecutive operating phases. In this study, a novel Bayesian model averaging based multi-kernel Gaussian process regression (BMA-MKGPR) approach is proposed for state estimation and quality prediction of nonlinear batch processes with multiple operating phases and between-phase transient dynamics. The new approach is applied to a simulated batch polymerization process and the result comparison shows that it can effective handle multiple nonlinear operating phases, between-phase transient dynamics and process uncertainty with high prediction accuracies.
As the main trend of development in new energy vehicles field, electric vehicle is bound to have a great development all over the world. However, it must have convenient public charging facilities ...and services before the vehicles become popular. As the vehicles need a long time when they charge, and the charging stations are widespread and unattended operation, it requires the drivers accomplish the charging operation by self-service, and the charging system should automatically complete the identification of the users' identity, battery charging and fee management. Therefore, it is imperative to develop a smart card which could have a bidirectional interaction with the charging post in order to provide a safe, friendly and convenient service for the users. In accordance with ISO/IEC7816 and ISO/IEC14443-A protocol, this paper accomplished the design and implementation of smart card based on CPU core for secure payment, and the smart card has both contact and contactless interface for communication, which could be well applicable to the secure payment of electric vehicles' charging.
Recent studies on few-shot classification using transfer learning pose challenges to the effectiveness and efficiency of episodic meta-learning algorithms. Transfer learning approaches are a natural ...alternative, but they are restricted to few-shot classification. Moreover, little attention has been on the development of probabilistic models with well-calibrated uncertainty from few-shot samples, except for some Bayesian episodic learning algorithms. To tackle the aforementioned issues, we propose a new transfer learning method to obtain accurate and reliable models for few-shot regression and classification. The resulting method does not require episodic meta-learning and is called meta-free representation learning (MFRL). MFRL first finds low-rank representation generalizing well on meta-test tasks. Given the learned representation, probabilistic linear models are fine-tuned with few-shot samples to obtain models with well-calibrated uncertainty. The proposed method not only achieves the highest accuracy on a wide range of few-shot learning benchmark datasets but also correctly quantifies the prediction uncertainty. In addition, weight averaging and temperature scaling are effective in improving the accuracy and reliability of few-shot learning in existing meta-learning algorithms with a wide range of learning paradigms and model architectures.
Long Term Evolution (LTE) system supports high-speed data transmission, thus high demand for real-time is a challenging and crucial problem. Meanwhile, channel estimation is the important module for ...the whole physical layer process. In this paper, the Single Instruction Multiple Data (SIMD) technology as a new method for signal processing is applied to implement this module based on general purpose processor (GPP). Through theoretical analysis and comparison between different algorithms for channel estimation, least square (LS) with linear interpolation is adopted to meet the requirement of real-time. The latency of the proposed processing method is evaluated by simulations. The simulation results show that the proposed method is much more efficient than the conventional processing without SIMD.
The latest 3GPP LTE release enables coordinated multi-point (CoMP) operation in heterogeneous networks. In this context, a multitude of low power nodes can be deployed within the coverage area of a ...high-power macro node, allowing to configure different sets of network nodes for the uplink reception from and the downlink transmission to a target mobile terminal. The resulting asymmetry between the uplink composite channel for joint reception and the downlink channel for joint transmission precludes the mobile terminal to accurately estimate the uplink pathloss using standard estimation techniques based on measuring the signal strength of downlink reference signals. In this paper, we introduce the concept of virtual propagation loss for an uplink signal jointly received by multiple geographically separated points. We demonstrate how this new metric can readily be computed at a network controller (.g., the eNodeB in LTE systems), hence enabling a fast correction of the erroneous pathloss estimate at the mobile terminal side through a simple closed-loop feedback for power control. We evaluate our method with extensive system-level simulations, showing that the proposed power control scheme achieves 20% cell-average and 50% cell-edge throughput gain over the LTE Rel.-10.
GPP-based design of soft MMSE MIMO detection Jiang Zhou; Tao Peng; Ran Duan ...
7th International Conference on Communications and Networking in China,
2012-Aug.
Conference Proceeding
General-purpose processor (GPP), which is suitable for software radio processing because of its powerful processing capability and flexibility, has aroused industry-wide attention and exploration. By ...taking advantage of streaming single instruction multiple data extensions (SSE), this paper presents the GPP-based design of multi-input multi-output (MIMO) detection which is of high complexity in LTE downlink receiver. The performance simulation results illustrate that with dramatically reduced complexity soft minimum mean square error (MMSE) detection is able to approximate ML performance in low signal-to-noise-ratio (SNR), and about 4 dB lose in high SNR. The complexity comparison results show that the implementation of soft MMSE detection in parallel can reduce the processing time by about 86% compared with not in parallel.
The great development of general-purpose processor technology makes the GPP possess the efficiently massive data processing abilities. In the solution(s) of 3GPP LTE system based on GPP platform, ...Turbo decoding algorithms play a vital role for high decoding complexity. This paper presents the design and implementation of SIMD technique based parallel turbo decoder suitable for GPP architecture. We improve the throughput through a) adopting fix-point design, b) taking full advantage of SIMD technique, c) using multiple soft-input soft-output decoders that operate in parallel, d) introducing multi-core technique to parallelize workload across cores. With 4 threads, the throughput of our SIMD and multi-threading technology based turbo decoder for GPP platform achieves 150 Mbps which can meet the requirement of LTE system.