Automatic fact verification has become an increasingly popular topic in recent years and among datasets the Fact Extraction and VERification (FEVER) dataset is one of the most popular. In this work we present BEVERS, a tuned baseline system for the FEVER dataset. Our pipeline uses standard approaches for document retrieval, sentence selection, and final claim classification, however, we spend considerable effort ensuring optimal performance for each component. The results are that BEVERS achieves the highest FEVER score and label accuracy among all systems, published or unpublished. We also apply this pipeline to another fact verification dataset, Scifact, and achieve the highest label accuracy among all systems on that dataset as well. We also make our full code available.

Despite the effort toward benchmarking self-supervised learning (SSL) methods for image recognition transfer learning tasks, our understanding is still limited about their performance on noisy real-world datasets. This paper presents an extensive analysis of various types of SSL methods on real-world datasets containing noisy images of wildlife animals. These uncurated images are auto-captured by motion-activated cameras or camera traps installed in the wild. The camera-trap datasets exhibit different types of biases typically present in practical tasks. Using a set of variably-size biased datasets, we compare the supervised learning (SL) method to two types of SSL methods, i.e., instance discrimination and cluster discrimination. Our results reveal nuances in SSL’s performance. For example, we show that though SSL methods are often more generalizable than the SL method, the performance gain of some SSL methods diminishes with the reduction in the size of the target dataset. Also, there exists significant variability in the effectiveness of the two types of SSL methods. In addition to this, we show that, unlike SL, both types of SSL gain from increased model capacity.

We present a Bayesian deep structured semantic model (BDSSM) that is efficient in retrieval tasks with a large pool of candidates for real-time applications, e.g., in search engines and digital ads. Efficiency is achieved by indexing the items into groups based on sparse representation during offline pre-computation. In online retrieval, the algorithm only retrieves and ranks items from indices that are relevant to the query. We then explore optimization strategies in the algorithm to make sparse representations sparser. Finally, our algorithm is compared with other popular clustering-based, hashing-based, and tree-based retrieval methods. We measure the differences in multiple dimensions, including retrieval recall, storage of embeddings, and CPU time. We show that our approach outperforms other algorithms in both recall and CPU time with the same storage limit.

In order to satisfy safety conditions, an agent may be constrained from acting freely. A safe controller can be designed a priori if an environment is well understood, but not when learning is employed. In particular, reinforcement learned (RL) controllers require exploration, which can be hazardous in safety critical situations. We study the benefits of giving structure to the constraints of a constrained Markov decision process by specifying them in formal languages as a step towards using safety methods from software engineering and controller synthesis. We instantiate these constraints as finite automata to efficiently recognise constraint violations. Constraint states are then used to augment the underlying MDP state and to learn a dense cost function, easing the problem of quickly learning joint MDP/constraint dynamics. We empirically evaluate the effect of these methods on training a variety of RL algorithms over several constraints specified in Safety Gym, MuJoCo, and Atari environments.

The timing of flowering plays a critical role in determining the productivity of agricultural crops. If the crops flower too early, the crop would mature before the end of the growing season, losing the opportunity to capture and use large amounts of light energy. If the crops flower too late, the crop may be killed by the change of seasons before it is ready to harvest. Maize flowering is one of the most important periods where even small amounts of stress can significantly alter yield. In this work, we developed and compared two methods for automatic tassel detection based on the imagery collected from an unmanned aerial vehicle, using deep learning models. The first approach was a customized framework for tassel detection based on convolutional neural network (TD-CNN). The other method was a state-of-the-art object detection technique of the faster region-based CNN (Faster R-CNN), serving as baseline detection accuracy. The evaluation criteria for tassel detection were customized to correctly reflect the needs of tassel detection in an agricultural setting. Although detecting thin tassels in the aerial imagery is challenging, our results showed promising accuracy: the TD-CNN had an F1 score of 95.9% and the Faster R-CNN had 97.9% F1 score. More CNN-based model structures can be investigated in the future for improved accuracy, speed, and generalizability on aerial-based tassel detection.

Motivation Empowered by advanced genomics discovery tools, recent biomedical research has produced a massive amount of genomic data on (post-)transcriptional regulations related to transcription factors, microRNAs, long non-coding RNAs, epigenetic modifications and genetic variations. Computational modeling, as an essential research method, has generated promising testable quantitative models that represent complex interplay among different gene regulatory mechanisms based on these data in many biological systems. However, given the dynamic changes of interactome in chaotic systems such as cancers, and the dramatic growth of heterogeneous data on this topic, such promise has encountered unprecedented challenges in terms of model complexity and scalability. In this study, we introduce a new integrative machine learning approach that can infer multifaceted gene regulations in cancers with a particular focus on microRNA regulation. In addition to new strategies for data integration and graphical model fusion, a supervised deep learning model was integrated to identify conditional microRNA-mRNA interactions across different cancer stages. Results In a case study of human breast cancer, we have identified distinct gene regulatory networks associated with four progressive stages. The subsequent functional analysis focusing on microRNA-mediated dysregulation across stages has revealed significant changes in major cancer hallmarks, as well as novel pathological signaling and metabolic processes, which shed light on microRNAs' regulatory roles in breast cancer progression. We believe this integrative model can be a robust and effective discovery tool to understand key regulatory characteristics in complex biological systems. Availability http://sbbi-panda.unl.edu/pin/

The (inverse) discrete Fourier transform (DFT/IDFT) is often perceived as essential to orthogonal frequency-division multiplexing (OFDM) systems. In this paper, a deep complex-valued convolutional network (DCCN) is developed to recover bits from time-domain OFDM signals without relying on any explicit DFT/IDFT. The DCCN can exploit the cyclic prefix (CP) of OFDM waveform for increased SNR by replacing DFT with a learned linear transform, and has the advantage of combining CP-exploitation, channel estimation, and intersymbol interference (ISI) mitigation, with a complexity of $\mathcal{O}(N^2)$. Numerical tests show that the DCCN receiver can outperform the legacy channel estimators based on ideal and approximate linear minimum mean square error (LMMSE) estimation and a conventional CP-enhanced technique in Rayleigh fading channels with various delay spreads and mobility. The proposed approach benefits from the expressive nature of complex-valued neural networks, which, however, currently lack support from popular deep learning platforms. In response, guidelines of exact and approximate implementations of a complex-valued convolutional layer are provided for the design and analysis of convolutional networks for wireless PHY. Furthermore, a suite of novel training techniques are developed to improve the convergence and generalizability of the trained model in fading channels. This work demonstrates the capability of deep neural networks in processing OFDM waveforms and the results suggest that the FFT processor in OFDM receivers can be replaced by a hardware AI accelerator.

Automating the detection of the corn tassels during flowering time is important in corn breeding. To control pollination, after a tassel is visible, the plant should be checked daily for emerging ears. The conventional methods are labor-intensive and time-consuming. In this study, we developed a technique for automatic detecting and locating corn tassel in unmanned aerial vehicle (UAV) imagery with the state-of-the art Faster Region based Convolutional Neural Network (Faster R-CNN). Each raw image was divided into 1000x1000 pixels sub-images, and 2000 sub-images were manually annotated for tassel locations with bounding boxes as ground-truth data. 80% of the annotated sub-images were used as training data and the remaining 20% were used for testing. The performance of the trained Faster R-CNN model was evaluated by customized evaluation criteria. The model achieved good performance on tassel detection with mean average precision of 91.78% and F1 score up to 97.98%.

This chapter introduces the basic concepts of machine learning (ML) in general and deep learning in particular, especially as they relate to computer vision and their applications to plant phenotyping. In short, machine learning is often employed when one needs to generalize from experience in a non-obvious way. In contrast, one would typically not use machine learning in more straightforward tasks like calculating payroll, sorting a list of words, serving web pages, word processing, monitoring CPU usage, and querying a database. In reinforcement learning, the performance of a policy is typically measured by repeatedly using the policy to control an agent in a simulated environment and computing the total reward per episode. In ML, there are many possible sets of functions scriptF that the learning algorithm can choose from, including decision trees, support vector machines, Bayesian networks, and hidden Markov models.

Towards the goal of presenting contrastive explanations for the output of a classifier, we present a method for finding contrastive foils in feature space and then visualizing in data space for easy interpretation. To produce a simple classifier for this method, we present a model that combines decision trees with supervised variational autoencoders using our new differentiable decision tree. This allows for end-to-end optimization of a deep network to perform feature extraction from structured, high-dimensional data for classification by a single decision tree. Our experiments demonstrate that the resulting model is satisfactory in both classification and image generation and can explain a model’s reasoning in answering a number of questions on classification and instance generation. Further experiments demonstrate that the model produces explanations using features that align with human-produced labels without any prior access to those labels.

Mid- to late-season weeds that escape from the routine early-season weed management threaten agricultural production by creating a large number of seeds for several future growing seasons. Rapid and accurate detection of weed patches in field is the first step of site-specific weed management. In this study, object detection-based convolutional neural network models were trained and evaluated over low-altitude unmanned aerial vehicle (UAV) imagery for mid- to late-season weed detection in soybean fields. The performance of two object detection models, Faster RCNN and the Single Shot Detector (SSD), were evaluated and compared in terms of weed detection performance using mean Intersection over Union (IoU) and inference speed. It was found that the Faster RCNN model with 200 box proposals had similar good weed detection performance to the SSD model in terms of precision, recall, f1 score, and IoU, as well as a similar inference time. The precision, recall, f1 score and IoU were 0.65, 0.68, 0.66 and 0.85 for Faster RCNN with 200 proposals, and 0.66, 0.68, 0.67 and 0.84 for SSD, respectively. However, the optimal confidence threshold of the SSD model was found to be much lower than that of the Faster RCNN model, which indicated that SSD might have lower generalization performance than Faster RCNN for mid- to late-season weed detection in soybean fields using UAV imagery. The performance of the object detection model was also compared with patch-based CNN model. The Faster RCNN model yielded a better weed detection performance than the patch-based CNN with and without overlap. The inference time of Faster RCNN was similar to patch-based CNN without overlap, but significantly less than patch-based CNN with overlap. Hence, Faster RCNN was found to be the best model in terms of weed detection performance and inference time among the different models compared in this study. This work is important in understanding the potential and identifying the algorithms for an on-farm, near real-time weed detection and management.

We study the problem of deep reinforcement learning where the agent's action sequences are constrained, e.g., prohibition of dithering or overactuating action sequences that might damage a robot, drone, or other physical device. Our model focuses on constraints that can be described by automata such as DFAs or PDAs. We then propose multiple approaches to augment the state descriptions of the Markov decision process (MDP) with summaries of recent action histories. We empirically evaluate these methods applying DQN to three Atari games, training with reward shaping. We found that our approaches are effective in significantly reducing, and even eliminating, constraint violations while maintaining high reward. We also observed that the total reward achieved by an agent can be highly sensitive to how much the constraints encourage or discourage exploration of potentially effective actions during training, and, in addition to helping ensure safe policies, the use of constraints can enhance exploration during training.

By expressing web page content in a format that machines can understand, the semantic web provides huge possibilities for the Internet and for machine reasoning. Unfortunately, there is a considerable distance between the present-day World Wide Web and the semantic web of the future. The process of annotating the Web to make it semantic web-ready is quite long and not without resistance. In this paper one mechanism for semanticizing the Web is presented. This system is known as AutoSHOE, and it is capable of categorizing pages according to one of the present HTML semantic representations (Simple HTML Ontology Extensions) by Heflin et al. We are also extending this system to other semantic web representations, such as the Resource Description Framework (RDF). The AutoSHOE system includes mechanisms to train classifiers to identify web pages that belong in an ontology, as well as methods to classify pages within an ontology and to learn relations between pages with respect to an ontology. The modular design of AutoSHOE allows for the addition of new ontologies as well as algorithms for feature extraction, classifier learning, and rule learning. This system has the promise to help transparently bridge traditional web technology to the semantic web using contemporary machine learning techniques rather than tedious manual annotation.

Many classification tasks target high-level concepts that can be decomposed into a hierarchy of finer-grained sub-concepts. For example, some string entities that are Locations are also Attractions, some Attractions are Museums, etc. Such hierarchies are common in named entity recognition (NER), document classification, and biological sequence analysis. We present a new approach for learning hierarchically decomposable concepts. The approach learns a high-level classifier (e.g., location vs. non-location) by seperately learning multiple finer-grained classifiers (e.g., museum vs. non-museum), and then combining the results. Soliciting labels at a finer level of granularity than that of the target concept is a new approach to active learning, which we term active over-labeling. In experiments in NER and document classification tasks, we show that active over-labeling substantially improves area under the precision-recall curve when compared with standard passive or active learning. Finally, because finer-grained labels may be more expensive to obtain, we also present a cost-sensitive active learner that uses a multi-armed bandit approach to dynamically choose the label granularity to target, and show that the bandit-based learner is robust to differences in label cost and labeling budget.

Minimality, a highly desirable consistency property of Constraint Satisfaction Problems (CSPs), is in general too expensive to enforce. Previous work has shown the practical benefits of restricting minimality to the clusters of a tree decomposition, allowing us to solve many difficult problems in a backtrack-free manner. We explore two alternative algorithms for enforcing minimality whose performance widely vary from one instance to another. We advocate a fine-grain portfolio approach to dynamically choose, during lookahead, the most appropriate algorithm for a cluster. Our strategy operates by selecting among two algorithms for enforcing minimality and an algorithm that enforces the lowest-level of consistency, which, in our setting, is Generalized Arc Consistency. Empirical evaluation on benchmark problems shows a significant improvement both in terms of the number of instances solved and CPU time.

We study the problem of identifying reactive compound(s) in a solution as efficiently as possible, with the goal of minimizing the number of chemical tests required to make exact identification. The area of group testing is appropriate for this problem, except that most group testing approaches assume that arbitrary tests can be performed, which is not the case in our application. To address this, we introduce a new model called mask-based constrained group testing, develop a randomized algorithm for it, and prove that under the right conditions, the algorithm is guaranteed w.h.p. to efficiently identify the active compounds of a solution with a small number of tests. We also show that our algorithm performs very well empirically on synthetic and real data.

BACKGROUND:Our environment is composed of biological components of varying magnitude. The relationships between the different biological elements can be represented as a biological network. The process of mating in S. cerevisiae is initiated by secretion of pheromone by one of the cells. Our interest lies in one particular question: how does a cell dynamically adapt the pathway to continue mating under severe environmental changes or under mutation (which might result in the loss of functionality of some proteins known to participate in the pheromone pathway). Our work attempts to answer this question. To achieve this, we first propose a model to simulate the pheromone pathway using Petri nets. Petri nets are directed graphs that can be used for describing and modeling systems characterized as concurrent, asynchronous, distributed, parallel, non-deterministic, and/or stochastic. We then analyze our Petri net-based model of the pathway to investigate the following: 1) Given the model of the pheromone response pathway, under what conditions does the cell respond positively, i.e., mate? 2) What kinds of perturbations in the cell would result in changing a negative response to a positive one?METHOD:In our model, we classify proteins into two categories: core component proteins (set ψ) and additional proteins (set λ). We randomly generate our model's parameters in repeated simulations. To simulate the pathway, we carry out three different experiments. In the experiments, we simply change the concentration of the additional proteins (λ) available to the cell. The concentration of proteins in ψ is varied consistently from 300 to 400. In Experiment 1, the range of values for λ is set to be 100 to 150. In Experiment 2, it is set to be 151 to 200. In Experiment 3, the set λ is further split into σ and ς, with the idea that proteins in σ are more important than those in ς. The range of values for σ is set to be between 151 to 200 while that of ς is 100 to 150. Decision trees were derived from each of the first two experiments to allow us to more easily analyze the conditions under which the pheromone is expressed.CONCLUSION:The simulation results reveal that a cell can overcome the detrimental effects of the conditions by using more concentration of additional proteins in λ. The first two experiments provide evidence that employing more concentration of proteins might be one of the ways that the cell uses to adapt itself in inhibiting conditions to facilitate mating. The results of the third experiment reveal that in some case the protein set σ is sufficient in regulating the response of the cell. Results of Experiments 4 and 5 reveal that there are certain conditions (parameters) in the model that are more important in determining whether a cell will respond positively or not.

Real-world datasets often contain large numbers of unlabeled data points, because there is additional cost for obtaining the labels. Semi-supervised learning (SSL) algorithms use both labeled and unlabeled data points for training that can result in higher classification accuracy on these datasets. Generally, traditional SSLs tentatively label the unlabeled data points on the basis of the smoothness assumption that neighboring points should have the same label. When this assumption is violated, unlabeled points are mislabeled injecting noise into the final classifier. An alternative SSL approach is cluster-then-label (CTL), which partitions all the data points (labeled and unlabeled) into clusters and creates a classifier by using those clusters. CTL is based on the less restrictive cluster assumption that data points in the same cluster should have the same label. As shown, this allows CTLs to achieve higher classification accuracy on many datasets where the cluster assumption holds for the CTLs, but smoothness does not hold for the traditional SSLs. However, cluster configuration problems (e.g., irrelevant features, insufficient clusters, and incorrectly shaped clusters) could violate the cluster assumption. We propose a new framework for CTLs by using a genetic algorithm (GA) to evolve classifiers without the cluster configuration problems (e.g., the GA removes irrelevant attributes, updates number of clusters, and changes the shape of the clusters). We demonstrate that a CTL based on this framework achieves comparable or higher accuracy with both traditional SSLs and CTLs on 12 University of California, Irvine machine learning datasets.

We explore the problem of budgeted machine learning, in which the learning algorithm has free access to the training examples' labels but has to pay for each attribute that is specified. This learning model is appropriate in many areas, including medical applications. We present new algorithms for choosing which attributes to purchase of which examples in the budgeted learning model based on algorithms for the multi-armed bandit problem. All of our approaches outperformed the current state of the art. Furthermore, we present a new means for selecting an example to purchase after the attribute is selected, instead of selecting an example uniformly at random, which is typically done. Our new example selection method improved performance of all the algorithms we tested, both ours and those in the literature.