Multi-center Assessment of CNN-Transformer with Belief Matching Loss for Patient-independent Seizure Detection in Scalp and Intracranial EEG

Neurologists typically identify epileptic seizures from electroencephalograms (EEGs) by visual inspection. This process is often time-consuming. To expedite the process, a reliable, automated, and patient-independent seizure detector is essential. However, developing such detector is challenging as seizures exhibit diverse morphologies across patients. In this study, we propose a patient-independent seizure detector to automatically detect seizures in both scalp EEG (sEEG) and intracranial EEG (iEEG). First, we deploy a convolutional neural network (CNN) with transformers and belief matching loss to detect seizures in single-channel EEG segments. Next, we utilized the channel-level outputs to detect seizures in multi-channel EEG segments. At last, we apply postprocessing filters to the segment-level outputs to determine the start and end points of seizures in multi-channel EEGs. We introduce the minimum overlap evaluation scoring (MOES) as an evaluation metric, improving upon existing metrics. We trained the seizure detector on the Temple University Hospital Seizure (TUH-SZ) sEEG dataset and evaluated it on five other independent sEEG and iEEG datasets. On the TUH-SZ dataset, the proposed patient-independent seizure detector achieves a sensitivity (SEN), precision (PRE), average and median false positive rate per hour (aFPR/h and mFPR/h), and median offset of 0.772, 0.429, 4.425, 0, and -2.125s, respectively. Across four adult datasets, we obtained SEN of 0.617-1.00, PRE of 0.534-1.00, aFPR/h of 0.425-2.002, and mFPR/h of 0-1.003. Meanwhile, on neonatal and paediatric datasets, we obtained SEN of 0.227-0.678, PRE of 0.377-0.818, aFPR/h of 0.253-0.421, and mFPR/h of 0.118-0.223. The proposed seizure detector takes less than 15s for a 30 minutes EEG, hence, it could potentially aid the clinicians in identifying seizures expeditiously, allocating more time for devising proper treatment.

Transformer Convolutional Neural Networks for Automated Artifact Detection in Scalp EEG

It is well known that electroencephalograms (EEGs) often contain artifacts due to muscle activity, eye blinks, and various other causes. Detecting such artifacts is an essential first step toward a correct interpretation of EEGs. Although much effort has been devoted to semi-automated and automated artifact detection in EEG, the problem of artifact detection remains challenging. In this paper, we propose a convolutional neural network (CNN) enhanced by transformers using belief matching (BM) loss for automated detection of five types of artifacts: chewing, electrode pop, eye movement, muscle, and shiver. Specifically, we apply these five detectors at individual EEG channels to distinguish artifacts from background EEG. Next, for each of these five types of artifacts, we combine the output of these channel-wise detectors to detect artifacts in multi-channel EEG segments. These segment-level classifiers can detect specific artifacts with a balanced accuracy (BAC) of 0.947, 0.735, 0.826, 0.857, and 0.655 for chewing, electrode pop, eye movement, muscle, and shiver artifacts, respectively. Finally, we combine the outputs of the five segment-level detectors to perform a combined binary classification (any artifact vs. background). The resulting detector achieves a sensitivity (SEN) of 60.4%, 51.8%, and 35.5%, at a specificity (SPE) of 95%, 97%, and 99%, respectively. This artifact detection module can reject artifact segments while only removing a small fraction of the background EEG, leading to a cleaner EEG for further analysis.

Enhance Language Identification using Dual-mode Model with Knowledge Distillation

In this paper, we propose to employ a dual-mode framework on the x-vector self-attention (XSA-LID) model with knowledge distillation (KD) to enhance its language identification (LID) performance for both long and short utterances. The dual-mode XSA-LID model is trained by jointly optimizing both the full and short modes with their respective inputs being the full-length speech and its short clip extracted by a specific Boolean mask, and KD is applied to further boost the performance on short utterances. In addition, we investigate the impact of clip-wise linguistic variability and lexical integrity for LID by analyzing the variation of LID performance in terms of the lengths and positions of the mimicked speech clips. We evaluated our approach on the MLS14 data from the NIST 2017 LRE. With the 3~s random-location Boolean mask, our proposed method achieved 19.23%, 21.52% and 8.37% relative improvement in average cost compared with the XSA-LID model on 3s, 10s, and 30s speech, respectively.

R3L: Connecting Deep Reinforcement Learning to Recurrent Neural Networks for Image Denoising via Residual Recovery

State-of-the-art image denoisers exploit various types of deep neural networks via deterministic training. Alternatively, very recent works utilize deep reinforcement learning for restoring images with diverse or unknown corruptions. Though deep reinforcement learning can generate effective policy networks for operator selection or architecture search in image restoration, how it is connected to the classic deterministic training in solving inverse problems remains unclear. In this work, we propose a novel image denoising scheme via Residual Recovery using Reinforcement Learning, dubbed R3L. We show that R3L is equivalent to a deep recurrent neural network that is trained using a stochastic reward, in contrast to many popular denoisers using supervised learning with deterministic losses. To benchmark the effectiveness of reinforcement learning in R3L, we train a recurrent neural network with the same architecture for residual recovery using the deterministic loss, thus to analyze how the two different training strategies affect the denoising performance. With such a unified benchmarking system, we demonstrate that the proposed R3L has better generalizability and robustness in image denoising when the estimated noise level varies, comparing to its counterparts using deterministic training, as well as various state-of-the-art image denoising algorithms.

Efficient Variational Bayesian Structure Learning of Dynamic Graphical Models

Estimating time-varying graphical models are of paramount importance in various social, financial, biological, and engineering systems, since the evolution of such networks can be utilized for example to spot trends, detect anomalies, predict vulnerability, and evaluate the impact of interventions. Existing methods require extensive tuning of parameters that control the graph sparsity and temporal smoothness. Furthermore, these methods are computationally burdensome with time complexity O(NP^3) for P variables and N time points. As a remedy, we propose a low-complexity tuning-free Bayesian approach, named BADGE. Specifically, we impose temporally-dependent spike-and-slab priors on the graphs such that they are sparse and varying smoothly across time. A variational inference algorithm is then derived to learn the graph structures from the data automatically. Owning to the pseudo-likelihood and the mean-field approximation, the time complexity of BADGE is only O(NP^2). Additionally, by identifying the frequency-domain resemblance to the time-varying graphical models, we show that BADGE can be extended to learning frequency-varying inverse spectral density matrices, and yields graphical models for multivariate stationary time series. Numerical results on both synthetic and real data show that that BADGE can better recover the underlying true graphs, while being more efficient than the existing methods, especially for high-dimensional cases.

Curb Your Normality: On the Quality Requirements of Demand Prediction for Dynamic Public Transport

As Public Transport (PT) becomes more dynamic and demand-responsive, it increasingly depends on predictions of transport demand. But how accurate need such predictions be for effective PT operation? We address this question through an experimental case study of PT trips in Metropolitan Copenhagen, Denmark, which we conduct independently of any specific prediction models. First, we simulate errors in demand prediction through unbiased noise distributions that vary considerably in shape. Using the noisy predictions, we then simulate and optimize demand-responsive PT fleets via a commonly used linear programming formulation and measure their performance. Our results suggest that the optimized performance is mainly affected by the skew of the noise distribution and the presence of infrequently large prediction errors. In particular, the optimized performance can improve under non-Gaussian vs. Gaussian noise. We also obtain that dynamic routing can reduce trip time by at least 23% vs. static routing. This reduction is estimated at 809,000 EUR per year in terms of Value of Travel Time Savings for the case study.

Modeling Sensing and Perception Errors towards Robust Decision Making in Autonomous Vehicles

Sensing and Perception (S&P) is a crucial component of an autonomous system (such as a robot), especially when deployed in highly dynamic environments where it is required to react to unexpected situations. This is particularly true in case of Autonomous Vehicles (AVs) driving on public roads. However, the current evaluation metrics for perception algorithms are typically designed to measure their accuracy per se and do not account for their impact on the decision making subsystem(s). This limitation does not help developers and third party evaluators to answer a critical question: is the performance of a perception subsystem sufficient for the decision making subsystem to make robust, safe decisions? In this paper, we propose a simulation-based methodology towards answering this question. At the same time, we show how to analyze the impact of different kinds of sensing and perception errors on the behavior of the autonomous system.

Factored Latent-Dynamic Conditional Random Fields for Single and Multi-label Sequence Modeling

Conditional Random Fields (CRF) are frequently applied for labeling and segmenting sequence data. Morency et al. (2007) introduced hidden state variables in a labeled CRF structure in order to model the latent dynamics within class labels, thus improving the labeling performance. Such a model is known as Latent-Dynamic CRF (LDCRF). We present Factored LDCRF (FLDCRF), a structure that allows multiple latent dynamics of the class labels to interact with each other. Including such latent-dynamic interactions leads to improved labeling performance on single-label and multi-label sequence modeling tasks. We apply our FLDCRF models on two single-label (one nested cross-validation) and one multi-label sequence tagging (nested cross-validation) experiments across two different datasets - UCI gesture phase data and UCI opportunity data. FLDCRF outperforms all state-of-the-art sequence models, i.e., CRF, LDCRF, LSTM, LSTM-CRF, Factorial CRF, Coupled CRF and a multi-label LSTM model in all our experiments. In addition, LSTM based models display inconsistent performance across validation and test data, and pose diffculty to select models on validation data during our experiments. FLDCRF offers easier model selection, consistency across validation and test performance and lucid model intuition. FLDCRF is also much faster to train compared to LSTM, even without a GPU. FLDCRF outshines the best LSTM model by ~4% on a single-label task on UCI gesture phase data and outperforms LSTM performance by ~2% on average across nested cross-validation test sets on the multi-label sequence tagging experiment on UCI opportunity data. The idea of FLDCRF can be extended to joint (multi-agent interactions) and heterogeneous (discrete and continuous) state space models.