Calibration-Aware Margin Loss: Pushing the Accuracy-Calibration Consistency Pareto Frontier for Deep Metric Learning

The ability to use the same distance threshold across different test classes / distributions is highly desired for a frictionless deployment of commercial image retrieval systems. However, state-of-the-art deep metric learning losses often result in highly varied intra-class and inter-class embedding structures, making threshold calibration a non-trivial process in practice. In this paper, we propose a novel metric named Operating-Point-Incosistency-Score (OPIS) that measures the variance in the operating characteristics across different classes in a target calibration range, and demonstrate that high accuracy of a metric learning embedding model does not guarantee calibration consistency for both seen and unseen classes. We find that, in the high-accuracy regime, there exists a Pareto frontier where accuracy improvement comes at the cost of calibration consistency. To address this, we develop a novel regularization, named Calibration-Aware Margin (CAM) loss, to encourage uniformity in the representation structures across classes during training. Extensive experiments demonstrate CAM's effectiveness in improving calibration-consistency while retaining or even enhancing accuracy, outperforming state-of-the-art deep metric learning methods.

Generalized NOMP for Line Spectrum Estimation and Detection from Coarsely Quantized Samples

As radar systems accompanied by large numbers of antennas and scale up in bandwidth, the cost and power consumption of high-precision (e.g., 10-12 bits) analog-to-digital converter (ADC) become the limiting factor. As a remedy, line spectral estimation and detection (LSE\&D) from low resolution (e.g., 1-4 bits) quantization has been gradually drawn attention in recent years. As low resolution quantization reduces the dynamic range (DR) of the receiver, the theoretical detection probabilities for the multiple targets (especially for the weakest target) are analyzed, which reveals the effects of low resolution on weak signal detection and provides the guidelines for system design. The computation complexities of current methods solve the line spectral estimation from coarsely quantized samples are often high. In this paper, we propose a fast generalized Newtonized orthogonal matching pursuit (GNOMP) which has superior estimation accuracy and maintains a constant false alarm rate (CFAR) behaviour. Besides, such an approach are easily extended to handle the other measurement scenarios such as sign measurements from time-varying thresholds, compressive setting, multisnapshot setting, multidimensional setting and unknown noise variance. Substantial numerical simulations are conducted to demonstrate the effectiveness of GNOMP in terms of estimating accuracy, detection probability and running time. Besides, real data are also provided to demonstrate the effectiveness of the GNOMP.

Musketeer (All for One, and One for All): A Generalist Vision-Language Model with Task Explanation Prompts

We present a sequence-to-sequence vision-language model whose parameters are jointly trained on all tasks (all for one) and fully shared among multiple tasks (one for all), resulting in a single model which we named Musketeer. The integration of knowledge across heterogeneous tasks is enabled by a novel feature called Task Explanation Prompt (TEP). TEP reduces interference among tasks, allowing the model to focus on their shared structure. With a single model, Musketeer achieves results comparable to or better than strong baselines trained on single tasks, almost uniformly across multiple tasks.

Channel Estimation for RIS-aided mmWave Massive MIMO System Using Few-bit ADCs

Millimeter wave (mmWave) massive multiple-input multiple-output (massive MIMO) is one of the most promising technologies for the fifth generation and beyond wireless communication system. However, a large number of antennas incur high power consumption and hardware costs, and high-frequency communications place a heavy burden on the analog-to-digital converters (ADCs) at the base station (BS). Furthermore, it is too costly to equipping each antenna with a high-precision ADC in a large antenna array system. It is promising to adopt low-resolution ADCs to address this problem. In this paper, we investigate the cascaded channel estimation for a mmWave massive MIMO system aided by a reconfigurable intelligent surface (RIS) with the BS equipped with few-bit ADCs. Due to the low-rank property of the cascaded channel, the estimation of the cascaded channel can be formulated as a low-rank matrix completion problem. We introduce a Bayesian optimal estimation framework for estimating the user-RIS-BS cascaded channel to tackle with the information loss caused by quantization. To implement the estimator and achieve the matrix completion, we use efficient bilinear generalized approximate message passing (BiG-AMP) algorithm. Extensive simulation results verify that our proposed method can accurately estimate the cascaded channel for the RIS-aided mmWave massive MIMO system with low-resolution ADCs.

Target-Mounted Intelligent Reflecting Surface for Joint Location and Orientation Estimation

Intelligent reflecting surface (IRS) has been widely recognized as an efficient technique to reconfigure the electromagnetic environment in favor of wireless communication performance. In this paper, we propose a new application of IRS for device-free target sensing via joint location and orientation estimation. In particular, different from the existing works that use IRS as an additional anchor node for localization/sensing, we consider mounting IRS on the sensing target, whereby estimating the IRS's location and orientation as that of the target by leveraging IRS's controllable signal reflection. To this end, we first propose a tensor-based method to acquire essential angle information between the IRS and the sensing transmitter as well as a set of distributed sensing receivers. Next, based on the estimated angle information, we formulate two optimization problems to estimate the location and orientation of the IRS/target, respectively, and obtain the locally optimal solutions to them by invoking two iterative algorithms, namely, gradient descent method and manifold optimization. In particular, we show that the orientation estimation problem admits a closed-form solution in a special case that usually holds in practice. Furthermore, theoretical analysis is conducted to draw essential insights into the proposed sensing system design and performance. Simulation results verify our theoretical analysis and demonstrate that the proposed methods can achieve high estimation accuracy which is close to the theoretical bound.

CFAR based NOMP for Line Spectral Estimation and Detection

The line spectrum estimation problem is considered in this paper. We propose a CFAR-based Newtonized OMP (NOMP-CFAR) method which can maintain a desired false alarm rate without the knowledge of the noise variance. The NOMP-CFAR consists of two steps, namely, an initialization step and a detection step. In the initialization step, NOMP is employed to obtain candidate sinusoidal components. In the detection step, CFAR detector is applied to detect each candidate frequency, and remove the most unlikely frequency component. Then, the Newton refinements are used to refine the remaining parameters. The relationship between the false alarm rate and the required threshold is established. By comparing with the NOMP, NOMP-CFAR has only $1$ dB performance loss in additive white Gaussian noise scenario with false alarm probability $10^{-2}$ and detection probability $0.8$ without knowledge of noise variance. For varied noise variance scenario, NOMP-CFAR still preserves its CFAR property, while NOMP violates the CFAR. Besides, real experiments are also conducted to demonstrate the detection performance of NOMP-CFAR, compared to CFAR and NOMP.

Compressed CPD-Based Channel Estimation and Joint Beamforming for RIS-Assisted Millimeter Wave Communications

We consider the problem of channel estimation and joint active and passive beamforming for reconfigurable intelligent surface (RIS) assisted millimeter wave (mmWave) multiple-input multiple-output (MIMO) orthogonal frequency division multiplexing (OFDM) systems. We show that, with a well-designed frame-based training protocol, the received pilot signal can be organized into a low-rank third-order tensor that admits a canonical polyadic decomposition (CPD). Based on this observation, we propose two CPD-based methods for estimating the cascade channels associated with different subcarriers. The proposed methods exploit the intrinsic low-rankness of the CPD formulation, which is a result of the sparse scattering characteristics of mmWave channels, and thus have the potential to achieve a significant training overhead reduction. Specifically, our analysis shows that the proposed methods have a sample complexity that scales quadratically with the sparsity of the cascade channel. Also, by utilizing the singular value decomposition-like structure of the effective channel, this paper develops a joint active and passive beamforming method based on the estimated cascade channels. Simulation results show that the proposed CPD-based channel estimation methods attain mean square errors that are close to the Cramer-Rao bound (CRB) and present a clear advantage over the compressed sensing-based method. In addition, the proposed joint beamforming method can effectively utilize the estimated channel parameters to achieve superior beamforming performance.

An In-depth Study of Stochastic Backpropagation

In this paper, we provide an in-depth study of Stochastic Backpropagation (SBP) when training deep neural networks for standard image classification and object detection tasks. During backward propagation, SBP calculates the gradients by only using a subset of feature maps to save the GPU memory and computational cost. We interpret SBP as an efficient way to implement stochastic gradient decent by performing backpropagation dropout, which leads to considerable memory saving and training process speedup, with a minimal impact on the overall model accuracy. We offer some good practices to apply SBP in training image recognition models, which can be adopted in learning a wide range of deep neural networks. Experiments on image classification and object detection show that SBP can save up to 40% of GPU memory with less than 1% accuracy degradation.

Towards Regression-Free Neural Networks for Diverse Compute Platforms

With the shift towards on-device deep learning, ensuring a consistent behavior of an AI service across diverse compute platforms becomes tremendously important. Our work tackles the emergent problem of reducing predictive inconsistencies arising as negative flips: test samples that are correctly predicted by a less accurate model, but incorrectly by a more accurate one. We introduce REGression constrained Neural Architecture Search (REG-NAS) to design a family of highly accurate models that engender fewer negative flips. REG-NAS consists of two components: (1) A novel architecture constraint that enables a larger model to contain all the weights of the smaller one thus maximizing weight sharing. This idea stems from our observation that larger weight sharing among networks leads to similar sample-wise predictions and results in fewer negative flips; (2) A novel search reward that incorporates both Top-1 accuracy and negative flips in the architecture search metric. We demonstrate that \regnas can successfully find desirable architectures with few negative flips in three popular architecture search spaces. Compared to the existing state-of-the-art approach, REG-NAS enables 33-48% relative reduction of negative flips.

Confederated Learning: Federated Learning with Decentralized Edge Servers

Federated learning (FL) is an emerging machine learning paradigm that allows to accomplish model training without aggregating data at a central server. Most studies on FL consider a centralized framework, in which a single server is endowed with a central authority to coordinate a number of devices to perform model training in an iterative manner. Due to stringent communication and bandwidth constraints, such a centralized framework has limited scalability as the number of devices grows. To address this issue, in this paper, we propose a ConFederated Learning (CFL) framework. The proposed CFL consists of multiple servers, in which each server is connected with an individual set of devices as in the conventional FL framework, and decentralized collaboration is leveraged among servers to make full use of the data dispersed throughout the network. We develop an alternating direction method of multipliers (ADMM) algorithm for CFL. The proposed algorithm employs a random scheduling policy which randomly selects a subset of devices to access their respective servers at each iteration, thus alleviating the need of uploading a huge amount of information from devices to servers. Theoretical analysis is presented to justify the proposed method. Numerical results show that the proposed method can converge to a decent solution significantly faster than gradient-based FL algorithms, thus boasting a substantial advantage in terms of communication efficiency.