In recent research, Tensor Product Representation (TPR) is applied for the systematic generalization task of deep neural networks by learning the compositional structure of data. However, such prior works show limited performance in discovering and representing the symbolic structure from unseen test data because their decomposition to the structural representations was incomplete. In this work, we propose an Attention-based Iterative Decomposition (AID) module designed to enhance the decomposition operations for the structured representations encoded from the sequential input data with TPR. Our AID can be easily adapted to any TPR-based model and provides enhanced systematic decomposition through a competitive attention mechanism between input features and structured representations. In our experiments, AID shows effectiveness by significantly improving the performance of TPR-based prior works on the series of systematic generalization tasks. Moreover, in the quantitative and qualitative evaluations, AID produces more compositional and well-bound structural representations than other works.

Intracranial hemorrhage (ICH) is a type of stroke with a high mortality rate and failing to localize even minor ICH can put a patient's life at risk. However, its patterns are diverse in shapes and sizes and, sometimes, even hard to recognize its existence. Therefore, it is challenging to accurately detect and localize diverse ICH patterns. In this article, we propose a novel Perihematomal Edema Guided Scale Adaptive R-CNN (PESA R-CNN) for accurate segmentation of various size hemorrhages with the goal of minimizing missed hemorrhage regions. In our approach, we design a Center Surround Difference U-Net (CSD U-Net) to incorporate Perihematomal Edema (PHE) for more accurate Region of Interest (RoI) generation. We trained CSD U-Net to predict PHE and hemorrhage regions as targets in a weakly supervised manner and utilized its prediction results to generate RoI. By including more informative features of PHE around hemorrhage, this enhanced RoI generation allows a model to reduce the false-negative rate. Furthermore, these expanded RoIs are aligned with the Scale Adaptive RoI Align (SARA) module based on their size to prevent the loss of fine-scale information and small hemorrhage patterns. Each scale adaptively aligned RoI is processed with the corresponding separate segmentation network of Multi-Scale Segmentation Network (MSSN), which integrates the results from each scale's segmentation network. In experiments, our model shows significant improvement on dice coefficient (0.697) and Hausdorff distance (12.918), compared to all other segmentation models. It also minimizes the number of missing small hemorrhage regions and enhances overall segmentation performance on diverse ICH patterns.

Motor imagery (MI) based Brain-Computer Interfaces (BCIs) controlled Functional Electrical Stimulation (FES) can help people with severe neuromuscular impairments to control their limbs by bypassing peripheral nerves and muscle pathways. However, there are still four major limitations with current MI-based BCIs for FES control: 1) They require relatively longer training and the training procedures are not clear. 2) Classification of different MI tasks within the same limb is difficult 3) MI features cannot be utilized during passive hand-motions induced by FES due to movement artifacts. 4) Few FES units are available which have real-time parameter control functionality. This study addresses these limitations by applying a hybrid BCI paradigm with a modified low-cost commercial off-the-shelf Transcutaneous Electrical Nerve Stimulation (TENS) unit. Four subjects were asked to mimic visual cues to imagine either closing or opening their dominant hand at different rates, such as fast or slow. After FES was initiated, the subjects were asked to attend to visual stimulus to elicit Steady-State Visual Evoked Potential (SSVEP) to stop FES. Results of this study showed that the modified TENS unit was able to successfully control hand motion in real-time. The classification results of different MI tasks within the same hand were promising. Furthermore, all subjects could stop the FES within 6 seconds and the average completion time was 2 seconds. The results of this study could provide insights towards future research of rehabilitation for stroke patients.

The goal of this study was to systematically compare signal characteristics and performance of three stimulation patterns that have been used to elicit steady-state somatosensory evoked potentials (SSSEPs): no pulses, random pulses, and rhythmic pulses with a consistent pattern. These three different vibrotactile stimulation patterns were provided to the fingertips of five healthy subjects by a small solenoid-type vibrating tactor. The five subjects, who were blindfolded, were asked to selectively focus their attention to either left or right fingertip flutter sensations, according to the audible cue given. Results of this study showed that small solenoid-type haptic tactors can elicit SSSEPs near the contra lateral central brain areas. There were significant differences in the resulting signals between the three paradigms, with the rhythmic pattern showing the highest classification accuracy. Moreover, the accuracy of the rhythmic pulse pattern was significantly higher than with or without random pulse patterns for the majority of subjects. The results of this study should provide insights to future research of SSSEP based BCIs and hybrid BCIs that use SSSEPS as one type of brain signal for users who are unable to use visual BCIs.