Publications

The multilayer perceptrons (MLPs) are widely used in many fields, however, singularities in the parameter space may seriously influence the learning dynamics of MLPs and cause strange learning behaviors. Given that the singularities are the subspaces of the parameter space where the Fisher information matrix (FIM) degenerates, the FIM plays a key role in the study of the singular learning dynamics of the MLPs. In this paper, we obtain the analytical form of the FIM for unipolar activation function-based MLPs where the input subjects to the Gaussian distribution with general covariance matrix and the unipolar error function is chosen as the activation function. Then three simulation experiments are taken to verify the validity of the obtained results.

The study of gender differences in emotions has attracted interest in psychology and neuroscience for decades. However, gender differences in the emotions evoked by haptic stimuli still need to be investigated further. In this paper, we applied visual-auditory-haptic fusion stimuli to evoke four target emotions (joy, sadness, fear, and neutral), and the electroen-cephalogram (EEG) signals were recorded synchronously. The differential entropy (DE) feature was extracted, and the support vector machine (SVM) was employed for emotion classification. Experimental results indicated that females generally behaved with lower brain activation changes than males in most frequency bands and brain regions. Meanwhile, haptic stimuli improved emotion classification accuracy, especially for females. In addition, females showed greater individual differences than males, and haptic stimuli reduced the differences in the same gender, particularly in males. This research promises to facilitate the application of emotion recognition in haptic interactions.

Myoelectric control interfaces, which map electromyographic (EMG) signals into control commands for external devices, have applications in active prosthesis control. However, the statistical characteristics of EMG signals change over time (e.g., because of changes in the electrode location), which makes interfaces based on static mapping unstable. Thus the user-decoder co-adaptation is needed during online operations. Nevertheless, current online decoder adaptation approaches present several practical challenges, such as expensive data labeling and slow convergence. Thus we introduce an unsupervised decoder adaptation method that converges rapidly. We use an autoencoder to extract motor intent representation in the latent manifold space rather than the sensor space, and further introduce an online unsupervised adaptation scheme based on Moore-Penrose Inverse, a noniterative approach suited for fast network re-training, to track the evolving manifold. A validation experiment first showed that the convergence time of the proposed adaptation scheme was reduced to about 50% of that for state-of-the-art methods. Online experiments further evaluated cursor and prosthetic hand control by the proposed myocontrol interface, where perturbations were representatively introduced by shifting the electrodes. Results showed that our scheme reached comparable improvements in robustness as supervised counterparts. Moreover, in a cup relocation test with a prosthetic hand, the completion time in the post-adaptation phase with electrode shift was comparable to that in the baseline phase without shift. These results suggest that our method effectively improves the accessibility and reliability of decoder adaptation, which has the potential to reduce the translational gap of myoelectric control interfaces by effective co-adaptation during operation.

Efficiently identifying activities of daily living (ADL) provides very important contextual information that is able to improve the effectiveness of various sports tracking and healthcare applications. Recently, attention mechanism that selectively focuses on time series signals has been widely adopted in sensor based human activity recognition (HAR), which can enhance interesting target activity and ignore irrelevant background activity. Several attention mechanisms have been investigated, which achieve remarkable performance in HAR scenario. Despite their success, prior these attention methods ignore the cross-interaction between different dimensions. In the paper, in order to avoid above shortcoming, we present a triplet cross-dimension attention for sensor-based activity recognition task, where three attention branches are built to capture the cross-interaction between sensor dimension, temporal dimension and channel dimension. The effectiveness of triplet attention method is validated through extensive experiments on four public HAR dataset namely UCI-HAR, PAMAP2, WISDM and UNIMIB-SHAR as well as the weakly labeled HAR dataset. Extensive experiments show consistent improvements in classification performance with various backbone models such as plain CNN and ResNet, demonstrating a good generality ability of the triplet attention. Visualization analysis is provided to support our conclusion, and actual implementation is evaluated on a Raspberry Pi platform.

Unlike image data, it is often hard to understand intricate sensor data for human activity, which generally contains heterogeneous sensor modalities from different body positions. The importance of every modality might also vary over time. Recent studies have witnessed the success of channel attention in boosting model performance. To maintain considerably low computational overhead, it utilizes global pooling operation to squeeze channel information, but neglects the importance of temporal-aware and modality-aware information that is very vital for activity recognition. In this paper, we propose a novel attention mechanism called TAMA (i.e., temporal-aware and modality-aware) to factorize global pooling operation into a pair of parallel activity feature embedding processes, which is able to simultaneously highlight the varying importance of temporal-aware and modality-aware information. Extensive ablation experiments verify that our TAMA attention can achieve competitive results on several standard HAR benchmarks without incurring extra computational burden. Moreover, a series of visualizing analysis is provided to show the improved interpretability by telling which temporal steps or which modalities is more determinant, which is in good line with human common intuition.

The robot end‐effecter positioning accuracy can be improved by the calibration of robot geometric parameters errors. According to the requirements of new generation geometrical product specification (GPS), the calibration uncertainty should be given when the calibration results are given. In this paper, the modified Denavit‐Hartenberg method (MDH) of six‐joint series robot is established and the joint movement trajectory method is applied to calibrate the robot geometric parameters. The uncertainty contributors significant are analyzed and the calibration uncertainty of robot geometric parameters is estimated based on the guide to the expression of uncertainty in measurement (GUM). In order to overcome the limitations of GUM for highly nonlinear model and reduce computational cost based on Monte Carlo Simulation (MCS) error estimation, an adaptive MCS (AMCS) is proposed to estimate the uncertainty distribution of robot end‐effector position. Simulation and practical example are illustrated and the experiments results confirm that not only can the proposed method evaluate the calibration uncertainty of geometric parameters, but also the uncertainty distribution of end‐effecter positions in the whole robot workspace can be estimated by AMCS in which the number of MCS trials can be selected adaptively and the quality of the numerical results can be controlled directly. The proposed method not only can evaluate the uncertainty of six‐joint series robot geometric parameters and end‐effecter position rapidly and accurately, but also can be popularized to the estimation of calibration uncertainty of other kinds of robot geometric parameters.