Adaptation Based on Generalized Discrepancy

Self-supervised learning based domain regularization for mask-wearing speaker verification

This work explores the usage of Fourier Transform for reducing the domain gap between the Source (e.g. Synthetic Image) and Target domain (e.g. Real Image) towards solving the Domain Adaptation problem. Most of the Unsupervised Domain Adaptation (UDA) algorithms reduce the global domain shift between labelled Source and unlabelled Target domain by matching the marginal distribution. UDA performance deteriorates for the cases where the domain gap between Source and Target is significant. To improve the overall performance of the existing UDA algorithms the proposed method attempts to bring the Source domain closer to the Target domain with the help of pseudo label based class consistent low-frequency swapping. This traditional image processing technique results in computational efficiency, especially compared to the state-of-the-art deep learning methods that use complex adversarial training. The proposed method Class Aware Frequency Transformation (CAFT <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> ) can easily be plugged into any existing UDA algorithm to improve its performance. We evaluate CAFT on various domain adaptation datasets and algorithms and have achieved performance gains across all the popular benchmarks.

Computer simulation studies of a frequency domain adaptive beamforming algorithm for planar arrays are presented. The algorithm is a modified, complex, least-mean-square (LMS) adaptive algorithm that provides estimates of the spherical coordinates (i.e., range, depression angle, and bearing angle) of multiple broadband targets in both the near-field and far-field regions of a planar array. Computer simulation results comparing the average range, depression angle, and bearing angle estimation errors as a function of the input signal-to-noise ratio (SNR) at a single element in the array, range (near field and far field), and harmonic number (frequency) are presented.

Domain adaptation has achieved promising results in many areas, such as image classification and object recognition. Although a lot of algorithms have been proposed to solve the task with different domain distributions, it remains a challenge for multi-source unsupervised domain adaptation. In addition, most of the existing algorithms learn a classifier on the source domain and predict the labels for the target data, which indicates that only the knowledge derived from the hyperplane is transferred to the target domain and the structure information is ignored. In light of this, we propose a novel algorithm for multi-source unsupervised domain adaptation. Generally speaking, we aim to preserve the whole structure from source domains and transfer it to serve the task on the target domain. The source and target data are put together for clustering, which simultaneously explores the structures of the source and target domains. The structure-preserved information from source domain further guides the clustering process on the target domain. Extensive experiments on two widely used databases on object recognition and face identification show the substantial improvement of our proposed approach over several state-of-the-art methods. Especially, our algorithm can take use of multi-source domains and achieve robust and better performance compared with the single source domain adaptation methods.

The application of convolutional neural networks (CNNs) for classification of human activities in frequency-modulated continuous-wave (FMCW) radar images has become widely adopted. However, complex measurement campaigns are needed for acquisition of large datasets - not only for varying activities but also for differing sensor settings such as e.g. deviating range and/or Doppler resolutions. In order to dramatically reduce cost and effort of those campaigns, this paper proposes deep domain adaptation techniques. These are based on four primary large datasets captured simultaneously with different sensor settings. The used datasets comprise five different human activities, each measured with the four varying settings, using Infineon's BGT60TR13C FMCW radar sensors operating at 60 GHz. This work emphasizes on the few-shot adversarial domain adaptation (FADA) algorithm, where we propose an improved version, showing better accuracy. The results are compared to the simple fine-tuning technique and the advanced Domain Adaptation using Stochastic Neighborhood Embedding (d-SNE) technique. We demonstrate the feasibility of the domain adaptation approach on radar data for human activity classification. Furthermore, we show that our improved FADA algorithm outperforms both conventional FADA and d-SNE while taking the same small amount of data in target domain captured with varying sensor settings.

Domain adaptation utilizes learned knowledge from an existing domain (source domain) to improve the classification performance of another related, but not identical, domain (target domain). Most existing domain adaptation methods first perform domain alignment, then apply standard classification algorithms. Transfer classifier induction is an emerging domain adaptation approach that incorporates the domain alignment into the process of building an adaptive classifier instead of using a standard classifier. Although transfer classifier induction approaches have achieved promising performance, they are mainly gradient-based approaches which can be trapped at local optima. In this article, we propose a transfer classifier induction algorithm based on evolutionary computation to address the above limitation. Specifically, a novel representation of the transfer classifier is proposed which has much lower dimensionality than the standard representation in existing transfer classifier induction approaches. We also propose a hybrid process to optimize two essential objectives in domain adaptation: 1) the manifold consistency and 2) the domain difference. Particularly, the manifold consistency is used in the main fitness function of the evolutionary search to preserve the intrinsic manifold structure of the data. The domain difference is reduced via a gradient-based local search applied to the top individuals generated by the evolutionary search. The experimental results show that the proposed algorithm can achieve better performance than seven state-of-the-art traditional domain adaptation algorithms and four state-of-the-art deep domain adaptation algorithms.

Numerical study of an adaptive domain decomposition algorithm based on Chebyshev tau method for solving singular perturbed problems

CSF: Closed-mask-guided semantic fusion method for semantic perception of unknown scenes

A Two-Way alignment approach for unsupervised multi-Source domain adaptation

Gait phase recognition based on surface electromyography (sEMG) signals provides a human–robot interface (HRI) for rehabilitation robots. However, a high-performance deep learning model requires to be calibrated on a large amount of data and frequently recalibrated for new gait speeds, which raises a burden on users. In this article, a domain adaptive convolutional neural network (DACNN) model for sEMG-based gait phase recognition against speed changes is proposed, which only needs to be pretrained on a comfortable gait speed (C.G.) and quick recalibrated on new gait speeds. Specifically, a CNN-based backbone model (BM) is first constructed and pretrained using C.G. data. Then, a domain adaptation (DA) algorithm is applied for recalibration, which allows the BM to learn speed-invariant representation with a small amount of new speed data. Four subjects participated in the experiment and were required to walk at four gait speeds, and three BMs based on AlexNet, LeNet, and fusion-data-CNN (FDCNN) were set as architectural options. Compared with the fine-tuning (FT) and non-recalibration (NoRC) strategies, the results show that the proposed method significantly outperforms both baseline methods. By recalibrating on 6-s data, the proposed method can achieve an accuracy improvement of 1.5%–6.1% compared with FT and 58%–81% compared with NoRC. The average gait phase recognition accuracies are 58.13% for AlexNet, 81.56% for LeNet, and 81.53% for FDCNN, respectively. The average mean absolute errors for gait event identification are 48, 85, and 66 ms, respectively. These results indicate that the DACNN with low recalibration burden can improve the usability of sEMG-based pattern recognition systems.

Chaotic routing [4, 13, 14] is a randomized, nonminimal adaptive routing algorithm for multicomputers. An adaptive routing algorithm is one in which the path a packet takes from its source to its destination may depend on other packets it encounters. Such algorithms potentially avoid network bottlenecks by routing packets around “hot spots.” Minimal adaptive routing algorithms have the additional advantage that the path each packet takes is a shortest one.Chinn, Leighton, and Tompa [6] provide a lower bound for permutation routing problems on the n×n mesh for a large class of deterministic minimal adaptive algorithms. Specifically, they prove that for any such routing algorithm, there exists a permutation that requires Ω(n 2/k 2) steps to route all the packets in the permutation, where k is the number of packets a node can contain.We present experimental results showing the performance of the Chaos router on permutations for which a deterministic minimal adaptive version of the Chaos router performs poorly. The results show that on these worst case permutations, the time the Chaos router takes to deliver all packets in the permutation closely fits a polynomial in n whose degree is 3/2. From these experiments, we conjecture that no practical router for the n×n mesh can route arbitrary permutations in time proportional to n, even though the mesh topology has the bandwidth to do so.

Wearable sensors provide enormous opportunities to identify activities and events of interest for various applications. However, a major limitation of the current systems is the fact that machine learning algorithms trained on particular sensors need to be retrained upon any changes in configuration of the system, such as adding a new sensor. In this paper, we aim to seamlessly train machine learning algorithms for the new sensors to identify activities and observations that are detectable by the pre-existing sensors. We create a domain adaptation method to expand training algorithms from known wearable sensors to new sensors, eliminating the need for manual training of machine learning algorithms. Specifically, our proposed approach eliminates the need for capturing substantial amount of data on new sensors. We propose the concept of stochastic features for human activity recognition, and design a new architecture of deep neural network to approximate the posterior distribution of the features. This approximation aligns the feature space of the new and old sensors by using limited, unlabeled data from the new sensor so that the previously defined classifier can be used with the new sensor. The experimental results show that (i) stochastic features are more robust against additive noise compared to typical convolutional neural networks based on deterministic features (ii) our framework outperforms the state-of-the-art domain adaptation algorithms. It can also achieve 10% improvement when training new sensors with limited unlabeled training data compared to training a model from scratch for the new sensor.

A multi-representation-based domain adaptation network for fault diagnosis

Technological developments call for increasing perception and action capabilities of robots. Among other skills, vision systems that can adapt to any possible change in the working conditions are needed. Since these conditions are unpredictable, we need benchmarks which allow to assess the generalization and robustness capabilities of our visual recognition algorithms. In this work we focus on robotic kitting in unconstrained scenarios. As a first contribution, we present a new visual dataset for the kitting task. Differently from standard object recognition datasets, we provide images of the same objects acquired under various conditions where camera, illumination and background are changed. This novel dataset allows for testing the robustness of robot visual recognition algorithms to a series of different domain shifts both in isolation and unified. Our second contribution is a novel online adaptation algorithm for deep models, based on batch-normalization layers, which allows to continuously adapt a model to the current working conditions. Differently from standard domain adaptation algorithms, it does not require any image from the target domain at training time. We benchmark the performance of the algorithm on the proposed dataset, showing its capability to fill the gap between the performances of a standard architecture and its counterpart adapted offline to the given target domain.

This paper introduces a novel classification algorithm for heterogeneous domain adaptation. The algorithm projects both the target and source data into a common feature space of the class decomposition scheme used. The distinctive features of the algorithm are: (1) it does not impose any assumptions on the data other than sharing the same class labels; (2) it allows adaptation of multiple source domains at once; and (3) it can help improving the topology of the projected data for class separability. The algorithm provides two built-in classification rules and allows applying any other classification model.