Despite the recent advancements in information retrieval (IR), zero-shot IR remains a significant challenge, especially when dealing with new domains, languages, and newly-released use cases that lack historical query traffic from existing users.For such cases, it is common to use query augmentations followed by fine-tuning pre-trained models on the document data paired with synthetic queries.In this work, we propose a novel Universal Document Linking (UDL) algorithm, which links similar documents to enhance synthetic query generation across multiple datasets with different characteristics.UDL leverages entropy for the choice of similarity models and named entity recognition (NER) for the link decision of documents using similarity scores.Our empirical studies demonstrate the effectiveness and universality of the UDL across diverse datasets and IR models, surpassing state-of-the-art methods in zero-shot cases.The developed code for reproducibility is included in the supplementary material.1

Electroencephalography (EEG) based emotion recognition is gaining substantial interest because of its strong association with the area of brain-computer interface. Even though several works exist in the literature, it is still challenging to find discriminative features that can generalize well to different EEG datasets. In this work, we focus on developing a deep learning model that makes use of the spatial and temporal representations of the EEG signal to generate EEG embeddings for emotion recognition. The proposed model uses a self-attention mechanism along with a feature fusion approach to improve the discrimination power of the learned EEG embeddings. Comprehensive experiments are conducted on the DEAP dataset, which demonstrates the superiority of the proposed work, where the attained accuracies for the arousal and valence classification are 91.17% and 90.73% , respectively.

In this work, we investigate Data Augmentation methods to improve the performance of state-of-the-art models for four different downstream tasks.Specifically, we propose Generative Adversarial Network using Language Models (GAN-LM) approach that combines a deep generative model with a pre-trained language model to produce diverse augmentations.We compare the GAN-LM to various conventional methods in non-contextual-and contextuallevels on four public datasets: ZESHEL for zero-shot entity linking, TREC for question classification, STS-B for sentence pairs semantic textual similarity (STS), and mSTS for multilingual sentence pairs STS.Additionally, we subsample these datasets to study the impact of such augmentations in low-resource settings where limited amounts of training data is available.Compared to the state-of-the-art methods in downstream tasks, we mostly achieve the best performance using GAN-LM approach.Finally, we investigate the way of combining the GAN-LM with other augmentation methods to complement our proposed approach.The developed code for reproducibility is included in the supplementary material.1

This paper presents a human activity recognition (HAR) system with wearable devices. While various approaches have been suggested for HAR, most of them focus on either 1) the inertial sensors to capture the physical movement or 2) subject-dependent evaluations that are less practical to real world cases. To this end, our work integrates sensing in-puts from physiological sensors to compensate the limitation of inertial sensors in capturing the human activities with less physical movements. Physiological sensors can capture physiological responses reflecting human behaviors in executing daily activities. To simulate a realistic application, three different evaluation scenarios are considered, namely All-access, Cross-subject and Cross-activity. Lastly, we propose a Hierarchical Deep Learning (HDL) model, which improves the accuracy and stability of HAR, compared to conventional models. Our proposed HDL with fusion of inertial and physiological sensing inputs achieves 97.16%, 92.23%, 90.18% average accuracy in All-access, Cross-subject, Cross-activity scenarios, which confirms the effectiveness of our approach.

This paper investigates the employment of photoplethysmography (PPG) for user authentication systems. Time-stable and user-specific features are developed by stretching the signal, designing a convolutional neural network and performing a variation-stable approach with three score fusions. Two evaluation scenarios are explored, namely single-session and two-sessions. In the earlier, the training and testing are done solely on one session data to find the user-specific features, while the second scenario is performed on data from two different sessions to test the time permanence of the features. The verification system was tested on four databases achieving an accuracy of 100% for single-session and 87.3% for two-sessions cases. The simulation results confirm the effectiveness of proposed variation-stable fusion which can be extended to other biometrics. The code is available in [1].

The photoplethysmography (PPG) is a non-invasive physiological signal that captures the changes in blood volume resulted from heart activity. It carries unique person-specific characteristics that can be utilized for biometric systems. Currently, the use of a biometric system is crucial to ensure the security of the user’s identity. Due to the high sensitivity of the PPG signal, it suffers from extreme variations within the same subject when obtained at different time instances. These variations impose a challenge to employ the PPG signal and hinder the algorithm generalization for many applications including verification and identification systems. In this work, we propose a PPG Biometric Generative Adversarial Network (PBGAN) to create synthetic person-specific and time-stable PPG signals for genuine samples. Two types of classification models are employed with the PBGAN where the focus is on verification scenarios. In addition, we expand our previously recorded PPG dataset from 100 to 170 participants where the new size guarantees the generalization capability of the proposed system. This database along with three public ones are employed to evaluate the performances in terms of uniqueness and time stability. Furthermore, we consider three different training strategies to simulate practical scenarios. The best results acquired from our collected database in terms of Equal Error Rate (EER) is 1.3% for the single-session and 11.5% for the two-sessions scenarios which demonstrate the effectiveness of the proposed method in improving the verification system’s performance. Compared to our previous work, we achieve 1.3% and 1.4% EER improvements in two-sessions’ databases with small computational times which reveals the superiority of our proposed approach for real applications.

In this work, we demonstrates the feasibility of employing the biometric photoplethysmography (PPG) signal for human verification applications. The PPG signal has dominance in terms of accessibility and portability which makes its usage in many applications such as user access control very appealing. Therefore, we developed robust time-stable features using signal analysis and deep learning models to increase the robustness and performance of the verification system with the PPG signal. The proposed system focuses on utilizing different stretching mechanisms namely Dynamic Time Warping, zero padding and interpolation with Fourier transform, and fuses them at the data level to be then deployed with different deep learning models. The designed deep models consist of Convolutional Neural Network (CNN) and Long-Short Term Memory (LSTM) which are considered to build a user specific model for the verification task. We collected a dataset consisting of 100 participants and recorded at two different time sessions using Plux pulse sensor. This dataset along with another two public databases are deployed to evaluate the performance of the proposed verification system in terms of uniqueness and time stability. The final result demonstrates the superiority of our proposed system tested on the built dataset and compared with other two public databases. The best performance achieved from our collected two-sessions database in terms of accuracy is 98% for the single-session and 87.1% for the two-sessions scenarios.