The challenging yet relevant task of automated essay scoring (AES) continuously gains attention from multiple disciplines over the years. With the advent of pre-trained large language models such as BERT, fine-tuning those models has become the dominant technique in various natural language processing (NLP) tasks. Several studies fine-tune BERT for the AES task but only utilize the final pooled output from its last layer. With BERT’s multi-layer architecture that encodes hierarchical linguistic information, we believe we can improve overall essay scoring performance by leveraging information from its intermediate layers. In this study, we diverge from the canonical fine-tuning paradigm by exploring different combinations of model outputs and single- and multi-layer pooling strategies, as well as architecture modifications to the task-specific component of the model. Using a hybrid pooling strategy, experimental results show that our best essay representa- tion combined with a simple architectural modification outperforms the average QWK score of the basic fine-tuned BERT with default output on the ASAP AES dataset, suggesting its effectiveness for the AES task and potentially other long-text tasks.

Myocardial infarction is the leading cause of mortality worldwide. The most common detection methodology for this disease relies on the use of 2D echocardiography to capture left ventricular wall motions. However, these are prone to image quality issues which, in worst cases, exhibit high noise, blurriness, and incomplete extracted regions. In this paper, we tackle the problem of detecting myocardial infarction in low-quality 2D echocardiograms using the Hamad Medical Corporation and Qatar University benchmark dataset. The contributions of this paper are as follows: 1) use image enhancement techniques to improve segmentation results of the left ventricular wall with low variation; 2) apply handcrafted feature descriptors using Harris corner detector and Histogram of Oriented Gradients; and 3) use of data augmentations to improve generalizability of deep learning models on the limited number of echo frames. Our work outperformed the baseline results for the segmentation task yielding scores of 99.76%, 94.68%, 94.64%, and 99.54% for specificity, precision, F1, and accuracy, respectively. On the classification task, ResNest50 yielded the best specificity score with 96.64%.

In a more realistic setting, the class distribution of a dataset is very imbalanced. Some set of classes have abundant samples while other classes are very limited. Standard deep learning approach is employed when the samples are abundant but when there are very few labelled data, few-shot learning approach are popularly used. Our work is concerned in between these two extremes. Specifically we tackle the problem of incremental few-shot (IFS) learning, which requires an adaptable model that can both perform well with classes with abundant samples, called base classes while also able to adapt and incrementally learn novel or unseen classes with very few samples. We modify the model-agnostic meta-learning (MAML) algorithm to create a much simpler and modular incremental-few shot learner that mitigates inherent problems encountered in this domain notably model overfitting and catastrophic forgetting. We propose DynMAML, which enables compartmentalization of the base and novel tasks into two independent heads while preventing gradient flow that causes feature drift. However due to two independent heads, each sample would have two outputs. We introduce the separator module placed in parallel to both classifier heads to select and predict the correct output, i.e. whether the final output is from the base or novel head. Experiments on a standard benchmark dataset shows our approach achieves similar performance with state-of-the-art methods while being much simpler and modular.

We propose an asynchronous gradient sharing mechanism for the parallel actor-critic algorithms with improved exploration characteristics. The proposed algorithm (A3C-GS) has the property of automatically diversifying worker policies in the short term for exploration, thereby reducing the need for entropy loss terms. Despite policy diversification, the algorithm converges to the optimal policy in the long term. We show in our analysis that the gradient sharing operation is a composition of two contractions. The first contraction performs gradient computation, while the second contraction is a gradient sharing operation coordinated by locks. From these two contractions, certain short- and long-term properties result. For the short term, gradient sharing induces temporary heterogeneity in policies for performing needed exploration. In the long term, under a suitably small learning rate and gradient clipping, convergence to the optimal policy is theoretically guaranteed. We verify our results with several high-dimensional experiments and compare A3C-GS against other on-policy policy-gradient algorithms. Our proposed algorithm achieved the highest weighted score. Despite lower entropy weights, it performed well in high-dimensional environments that require exploration due to sparse rewards and those that need navigation in 3-D environments for long survival tasks. It consistently performed better than the base asynchronous advantage actor-critic (A3C) algorithm.

Eye movements are intricate and dynamic events that contain a wealth of information about the subject and the stimuli. We propose an abstract representation of eye movements that preserve the important nuances in gaze behavior while being stimuli-agnostic. We consider eye movements as raw position and velocity signals and train separate deep temporal convolutional autoencoders. The autoencoders learn micro-scale and macroscale representations that correspond to the fast and slow features of eye movements. We evaluate the joint representations with a linear classifier fitted on various classification tasks. Our work accurately discriminates between gender and age groups and outperforms previous works on biometrics and stimuli classification. Further experiments highlight the validity and generalizability of this method, bringing eye-tracking research closer to real-world applications.

In this paper, we seek to extend multi-object tracking research on a relatively less explored domain, that of, underwater multi-object tracking in the wild. Multi-object fish tracking is an important task because it can provide fish monitoring systems with richer information (e.g. multiple views of the same fish) as compared to detections and it can be an invaluable input to fish behavior analysis. However, there is a lack of an annotated benchmark dataset with enough samples for this task. To circumvent the need for manual ground truth tracking annotation, we craft a synthetic dataset. Using this synthetic dataset, we train an integrated detector and tracker called SynDHN. SynDHN uses the Deep Hungarian Network (DHN), which is a differentiable approximation of the Hungarian assignment algorithm. We repurpose DHN to become the tracking component of our algorithm by performing the task of affinity estimation between detector predictions. We consider both spatial and appearance features for affinity estimation. Our results show that despite being trained on a synthetic dataset, SynDHN generalizes well to real underwater video tracking and performs better against our baseline algorithms.

Schistosomiasis, trichuriasis, and ascariasis are few of the many neglected tropical diseases that still affect populations in poor countries. These diseases cause a variety of symptoms such as abdominal pain, may lead to complications, and may even result in death in severe schistosomiasis cases. To complement the efforts of governments and health organizations in mitigating the morbidity and transmission of neglected tropical diseases, several applications utilizing machine learning techniques have been developed in recent years to automate the detection of parasites in microscopy samples. In this paper, we explore the use of YOLO, a convolutional neural network framework, in the detection of helminth eggs in stool samples. We collected and labelled a dataset with varying imaging conditions due to different staining conditions and acquisition by smartphone cameras with different parameters. We demonstrate that the approach works well despite this variance in imaging conditions in the dataset, achieving high sensitivity in the detection of helminth eggs and high accuracy in the identification of egg species. The trained model operates in real-time, making it suitable for automated diagnosis and real-time annotation.

Protein-protein interactions (PPIs)are vital to numerous biological processes. Computational methods have been used to predict PPIs from protein sequences. Several studies utilize popular algorithms such as Support Vector Machines (SVM)and Random Forest (RF)for detecting PPIs. The hypothesis of this study is that Extreme Gradient Boosting (XGBoost), which uses gradient boosted decision trees as the base classifier, can produce comparable results to those produced by SVM and RF. Based on the experimental results for the assembled protein interaction dataset, XGBoost produced better results than SVM and RF for the majority of the metrics used.

The following topics are dealt with: learning (artificial intelligence); feature extraction; medical image processing; convolutional neural nets; image segmentation; image classification; diseases; support vector machines; pattern classification; Internet of Things.

Monocular depth estimation is extremely challenging because it is inherently an ambiguous and ill-posed problem. The unsupervised approach to monocular depth estimation using convolutional neural networks is gaining a lot of interest since learning from a set of rectified stereo image pairs without ground truth depths and predicting scene geometry from a single image have become feasible. The proposed approach requires training an encoder-decoder network architecture, referred to as autoencoders in autoencoder (AsiANet), in an unsupervised fashion to discover the implicit relationship between a single image and its corresponding depth map. AsiANet uses a unique Inception-like pooling module based on fractional max-pooling for dimensionality reduction. Experiments on the KITTI benchmark dataset show that the proposed architecture trained using the Charbonnier loss function achieved superior performance on depth map prediction compared to previous unsupervised monocular depth estimation methods.

Reinforcement learning algorithms that use policy gradient methods approach an optimal policy faster than Q-learning but at the cost of incurring high variances in gradients. Among variance reduction techniques are actor-critic methods that use value and advantage functions to train a policy actor. We propose an algorithm under the actor-critic family that further reduces gradient variance through estimation of advantage distributions from K deep network critics. We combine outputs of the K critics into an advantage distribution using a histogram approach followed by kernel convolution. We show in our analysis that using the K-critic advantage distribution provides variance reduction properties that results in more stable performance even on long training runs. We test our algorithm on a set of high-dimensional VizDoom experiments. Our experimental results show that our proposed algorithm attains the most average rewards compared to other methods, and with less noise compared to the 1-critic method.

We propose AC2, a policy gradient algorithm that employs a primary and a secondary critic to manage both bias and variance in policy gradients. We present through analyses and experiments that performance becomes more stable if a secondary critic concentrates on few problematic states (upper 95-percentile) that cause extreme changes in value estimates. This scheme can keep biases tolerable while lowering variances. We relate our algorithm with critic ensembles that have more components and show that ensemble averaging may not significantly reduce gradient variances in more difficult environments. We test our algorithm in a series of high-dimensional experiments and report better performance than ensembles with more critic components especially in harder environments. In addition, performance is more stable if the secondary critic trains on a few problematic states than by random sampling. Our algorithm reports better reward performance than single critic and other RL models.

Sign language is very important for people who have impaired hearing and speaking inabilities. In this work, we present a method to classify RGB images of static letter hand poses in Sign Language using a Convolutional Neural Netowrk (CNN) inspired by Densely Connected Convolutional Neural Networks (DenseNet). It was further implemented to classify sign languages in real time using a web camera. DenseNet has been widely used for classification tasks due to the advantages it introduces such as alleviating the vanishing gradient - a common problem encountered with deep networks. Since a deep network is proposed to be used for our sign language classification task, this characteristic is useful. Our proposed network was able to achieve an accuracy of 90.3 % which is comparable to other works including those that used depth images in addition to RGB images. Our network was also able to achieve prediction rates of 50 to 100 Hz which makes it capable of real-time prediction.

High-throughput analytical techniques such as nuclear magnetic resonance, protein kinase phosphorylation, and mass spectroscopic methods generate time dense profiles of metabolites or proteins that are replete with structural and kinetic information about the underlying system that produced them. Experimentalists are in urgent need of computational tools that will allow efficient extraction of this information from these time series data. A new parameter estimation method for biochemical systems formulated as Generalized Mass Action (GMA) models known to capture the nonlinear dynamics of complex biological systems such as gene regulatory, signal transduction and metabolic networks, is described. For such models, it is known that parameter estimation algorithm performance deteriorates rapidly with increasing network size. We propose a decomposition strategy that breaks up the system equations into terms whose rate constants and kinetic order parameters are estimated one term at a time resulting in dramatic parameter space dimensionality reductions. This approach is demonstrated in a hybrid algorithm based on Regularized Least Squares Regression and Multi-objective Particle Swarm Optimization. We validate our proposed strategy through the efficient and accurate extraction of GMA model parameter values from noise-free and noisy simulated data for Saccharomyces cerevisiae and actual Nuclear Magnetic Resonance (NMR) data for Lactoccocus lactis.

MOTIVATION: High-throughput technologies now allow the acquisition of biological data, such as comprehensive biochemical time-courses at unprecedented rates. These temporal profiles carry topological and kinetic information regarding the biochemical network from which they were drawn. Retrieving this information will require systematic application of both experimental and computational methods. RESULTS: S-systems are non-linear mathematical approximative models based on the power-law formalism. They provide a general framework for the simulation of integrated biological systems exhibiting complex dynamics, such as genetic circuits, signal transduction and metabolic networks. We describe how the heuristic optimization technique simulated annealing (SA) can be effectively used for estimating the parameters of S-systems from time-course biochemical data. We demonstrate our methods using three artificial networks designed to simulate different network topologies and behavior. We then end with an application to a real biochemical network by creating a working model for the cadBA system in Escherichia coli. AVAILABILITY: The source code written in C++ is available at http://www.engg.upd.edu.ph/~naval/bioinformcode.html. All the necessary programs including the required compiler are described in a document archived with the source code. SUPPLEMENTARY INFORMATION: Supplementary material is available at Bioinformatics online.

In this paper, we present an approach that extends the Particle Swarm Optimization (PSO) algorithm to handle multiobjective optimization problems by incorporating the mechanism of crowding distance computation into the algorithm of PSO, specifically on global best selection and in the deletion method of an external archive of nondominated solutions. The crowding distance mechanism together with a mutation operator maintains the diversity of nondominated solutions in the external archive. The performance of this approach is evaluated on test functions and metrics from literature. The results show that the proposed approach is highly competitive in converging towards the Pareto front and generates a well distributed set of nondominated solutions.