Encode Domain Knowledge Research Articles

Digital Twin Technologies for Autonomous Environmental Control and Life Support Systems

Environmental control and life support systems will require enhanced self-awareness and self-sufficiency as human spaceflights are designed to reach further destinations. These requirements have led to the development of autonomous technologies and systems to enable more Earth independence, while at the same time relying more heavily on the knowledge contained in their computational models (as opposed to the knowledge of ground control experts). For environmental control and life support systems, these consist of disparate models often tailored to specific subsystems and use cases, such as temperature control and CO2 removal. Therefore, there is a need for technologies supporting the integration of existing models. We propose to extend existing digital twin frameworks to integrate models and serve as a tool for answering onboard queries during operation. Toward this vision, we identify research directions for three types of technologies: i) streamlining the process of merging information models without redundancies by leveraging model-based systems engineering languages; ii) the calibration of simulation models that requires a better understanding of how to encode domain knowledge into a probabilistic representation of subsystem states and model parameters; and iii) automating the query understanding process and constructing a mapping between information and simulation models.

KEBLM: Knowledge-Enhanced Biomedical Language Models

Pretrained language models (PLMs) have demonstrated strong performance on many natural language processing (NLP) tasks. Despite their great success, these PLMs are typically pretrained only on unstructured free texts without leveraging existing structured knowledge bases that are readily available for many domains, especially scientific domains. As a result, these PLMs may not achieve satisfactory performance on knowledge-intensive tasks such as biomedical NLP. Comprehending a complex biomedical document without domain-specific knowledge is challenging, even for humans. Inspired by this observation, we propose a general framework for incorporating various types of domain knowledge from multiple sources into biomedical PLMs.We encode domain knowledge using lightweight adapter modules, bottleneck feed-forward networks that are inserted into different locations of a backbone PLM. For each knowledge source of interest, we pretrain an adapter module to capture the knowledge in a self-supervised way. We design a wide range of self-supervised objectives to accommodate diverse types of knowledge, ranging from entity relations to description sentences.Once a set of pretrained adapters is available, we employ fusion layers to combine the knowledge encoded within these adapters for downstream tasks. Each fusion layer is a parameterized mixer of the available trained adapters that can identify and activate the most useful adapters for a given input. Our method diverges from prior work by including a knowledge consolidation phase, during which we teach the fusion layers to effectively combine knowledge from both the original PLM and newly-acquired external knowledge using a large collection of unannotated texts. After the consolidation phase, the complete knowledge-enhanced model can be fine-tuned for any downstream task of interest to achieve optimal performance.Extensive experiments on many biomedical NLP datasets show that our proposed framework consistently improves the performance of the underlying PLMs on various downstream tasks such as natural language inference, question answering, and entity linking. These results demonstrate the benefits of using multiple sources of external knowledge to enhance PLMs and the effectiveness of the framework for incorporating knowledge into PLMs. While primarily focused on the biomedical domain in this work, our framework is highly adaptable and can be easily applied to other domains, such as the bioenergy sector.

Human-centered deep compositional model for handling occlusions

Despite their powerful discriminative abilities, Convolutional Neural Networks (CNNs) lack the properties of generative models. This leads to a decreased performance in environments where objects are poorly visible. Solving such a problem by adding more training samples can quickly lead to a combinatorial explosion, therefore the underlying architecture has to be changed instead. This work proposes a Human-Centered Deep Compositional model (HCDC) that combines low-level visual discrimination of a CNN and the high-level reasoning of a Hierarchical Compositional model (HCM). Defined as a transparent model, it can be optimized to real-world environments by adding compactly encoded domain knowledge from human studies and physical laws. The new FridgeNetv2 dataset and a mixture of publicly available datasets are used as a benchmark. The experimental results show the proposed model is explainable, has higher discriminative and generative power, and better handles the occlusion than the current state-of-the-art Mask-RCNN in instance segmentation tasks.

Automated Configuration of Heterogeneous Graph Neural Networks With a Semantic Math Parser for IoT Systems

Efficient training of deep learning models from time series data for Internet of Things (IoT) systems requires a good understanding of the domain, particularly if the training is automated for large-scale applications. Heterogeneous graph neural networks (HGNNs) are a promising approach for incorporating domain knowledge into the modeling framework and consequently improving model performance. However, encoding domain knowledge into HGNNs is nontrivial for IoT systems and requires substantial manual effort. This complicates the adoption of HGNNs in practical settings. To overcome this drawback, we propose a framework for the automatic derivation of HGNN features by semantically parsing equations present in scientific and dedicated publications. We encode the derived features considering physical causation from these equations into an HGNN using an underlying Transformer for prediction and anomaly detection. We validate our approach using two IoT use cases, namely, the prediction of the remaining energy in the battery of an electric race car and the anomaly detection during pick and place operations in a robot workcell. We demonstrate that our approach significantly outperforms other competitive techniques.

Informative Bayesian Neural Network Priors for Weak Signals

Encoding domain knowledge into the prior over the high-dimensional weight space of a neural network is challenging but essential in applications with limited data and weak signals. Two types of domain knowledge are commonly available in scientific applications: 1. feature sparsity (fraction of features deemed relevant); 2. signal-to-noise ratio, quantified, for instance, as the proportion of variance explained. We show how to encode both types of domain knowledge into the widely used Gaussian scale mixture priors with Automatic Relevance Determination. Specifically, we propose a new joint prior over the local (i.e., feature-specific) scale parameters that encodes knowledge about feature sparsity, and a Stein gradient optimization to tune the hyperparameters in such a way that the distribution induced on the model’s proportion of variance explained matches the prior distribution. We show empirically that the new prior improves prediction accuracy compared to existing neural network priors on publicly available datasets and in a genetics application where signals are weak and sparse, often outperforming even computationally intensive cross-validation for hyperparameter tuning.

Unboxing the graph: Towards interpretable graph neural networks for transport prediction through neural relational inference

Predicting the supply and demand of transport systems is vital for efficient traffic management, control, optimization, and planning. For example, predicting where from/to and when people intend to travel by taxi can support fleet managers in distributing resources; better predictions of traffic speeds/congestion allows for pro-active control measures or for users to better choose their paths. Making spatio-temporal predictions is known to be a hard task, but recently Graph Neural Networks (GNNs) have been widely applied on non-Euclidean spatial data. However, most GNN models require a predefined graph, and so far, researchers rely on heuristics to generate this graph for the model to use. In this paper, we use Neural Relational Inference to learn the optimal graph for the model. Our approach has several advantages: 1) a Variational Auto Encoder structure allows for the graph to be dynamically determined by the data, potentially changing through time; 2) the encoder structure allows the use of external data in the generation of the graph; 3) it is possible to place Bayesian priors on the generated graphs to encode domain knowledge. We conduct experiments on two datasets, namely the NYC Yellow Taxi and the PEMS-BAY road traffic datasets. In both datasets, we outperform benchmarks and show performance comparable to state-of-the-art. Furthermore, we do an in-depth analysis of the learned graphs, providing insights on what kinds of connections GNNs use for spatio-temporal predictions in the transport domain and how these connections can help interpretability.

MultiplexNet: Towards Fully Satisfied Logical Constraints in Neural Networks

We propose a novel way to incorporate expert knowledge into the training of deep neural networks. Many approaches encode domain constraints directly into the network architecture, requiring non-trivial or domain-specific engineering. In contrast, our approach, called MultiplexNet, represents domain knowledge as a quantifier-free logical formula in disjunctive normal form (DNF) which is easy to encode and to elicit from human experts. It introduces a latent Categorical variable that learns to choose which constraint term optimizes the error function of the network and it compiles the constraints directly into the output of existing learning algorithms. We demonstrate the efficacy of this approach empirically on several classical deep learning tasks, such as density estimation and classification in both supervised and unsupervised settings where prior knowledge about the domains was expressed as logical constraints. Our results show that the MultiplexNet approach learned to approximate unknown distributions well, often requiring fewer data samples than the alternative approaches. In some cases, MultiplexNet finds better solutions than the baselines; or solutions that could not be achieved with the alternative approaches. Our contribution is in encoding domain knowledge in a way that facilitates inference. We specifically focus on quantifier-free logical formulae that are specified over the output domain of a network. We show that this approach is both efficient and general; and critically, our approach guarantees 100% constraint satisfaction in a network's output.

A novel tool that allows interactive screening of PubMed citations showed promise for the semi-automation of identification of Biomedical Literature

Background and ObjectivesSystematic reviews form the basis of evidence-based medicine, but are expensive and time-consuming to produce. To address this burden, we have developed a literature identification system (Pythia) that combines the query formulation and citation screening steps. MethodsPythia incorporates a set of natural-language questions with machine-learning algorithms to rank all PubMed citations based on relevance, returning the 100 top-ranked citations for human screening. The tagged citations are iteratively exploited by Pythia to refine the search and re-rank the citations. ResultsAcross seven systematic reviews, the ability of Pythia to identify the relevant citations (sensitivity) ranged from 0.09 to 0.58. The number of abstracts reviewed per relevant abstract number needed to read (NNR) was lower than in the manually screened project in four reviews, higher in two, and had mixed results in one. The reviews that had greater overall sensitivity retrieved more relevant citations in early batches, but retrieval was generally unaffected by other aspects, such as study design, study size, and specific key question. ConclusionDue to its low sensitivity, Pythia is not ready for widespread use. Future research should explore ways to encode domain knowledge in query formulation to better enrich the questions used in the search.

Simulation-assisted machine learning for operational digital twins

In the environmental sciences, there are ongoing efforts to combine multiple models to assist the analysis of complex systems. Combining process-based models, which have encoded domain knowledge, with machine learning models, which can flexibly adapt to input data, can improve modeling capabilities. However, both types of models have input data limitations. We propose a methodology to overcome these issues by using a process-based model to generate data, aggregating them to a lower resolution to mimic real situations, and developing machine learning models using a fraction of the process-based model inputs. We showcase this method with a case study of pasture nitrogen response rate prediction. We train models of different scales and test them in sampled and unsampled location experiments to assess their practicality in terms of accuracy and generalization. The resulting models provide accurate predictions and generalize well, showing the usefulness of the proposed method for tactical decision support.

Encoding Human Domain Knowledge to Warm Start Reinforcement Learning

Deep reinforcement learning has been successful in a variety of tasks, such as game playing and robotic manipulation. However, attempting to learn tabula rasa disregards the logical structure of many domains as well as the wealth of readily available knowledge from domain experts that could help "warm start" the learning process. We present a novel reinforcement learning technique that allows for intelligent initialization of a neural network weights and architecture. Our approach permits the encoding domain knowledge directly into a neural decision tree, and improves upon that knowledge with policy gradient updates. We empirically validate our approach on two OpenAI Gym tasks and two modified StarCraft 2 tasks, showing that our novel architecture outperforms multilayer-perceptron and recurrent architectures. Our knowledge-based framework finds superior policies compared to imitation learning-based and prior knowledge-based approaches. Importantly, we demonstrate that our approach can be used by untrained humans to initially provide >80% increase in expected reward relative to baselines prior to training (p < 0.001), which results in a >60% increase in expected reward after policy optimization (p = 0.011).

Optimising Automatic Calibration of Electric Muscle Stimulation

Electrical Muscle Stimulation (EMS) has become a popular interaction technology in Human-Computer Interaction; allowing the computer to take direct control of the user's body. To date, however, the explorations have been limited to coarse, toy examples, due to the low resolution of achievable control. To increase this resolution, the EMS needs to increase significantly in complexity - using large numbers of electrodes in complex patterns. The calibration of such a system remains an unsolved challenge. We present a new SAT-based black-box calibration method, which requires no spatial information about muscular or electrode positioning. The method encodes domain knowledge and observations in a constraint model, and uses these to prune the space of feasible control signals. In a simulated environment we find this method can scale reliably to large arrays while requiring only a modest number of trials, and preliminary tests on real hardware show we can effectively calibrate an electrode array in a few minutes.

SMAT: An attention-based deep learning solution to the automation of schema matching.

Schema matching aims to identify the correspondences among attributes of database schemas. It is frequently considered as the most challenging and decisive stage existing in many contemporary web semantics and database systems. Low-quality algorithmic matchers fail to provide improvement while manually annotation consumes extensive human efforts. Further complications arise from data privacy in certain domains such as healthcare, where only schema-level matching should be used to prevent data leakage. For this problem, we propose SMAT, a new deep learning model based on state-of-the-art natural language processing techniques to obtain semantic mappings between source and target schemas using only the attribute name and description. SMAT avoids directly encoding domain knowledge about the source and target systems, which allows it to be more easily deployed across different sites. We also introduce a new benchmark dataset, OMAP, based on real-world schema-level mappings from the healthcare domain. Our extensive evaluation of various benchmark datasets demonstrates the potential of SMAT to help automate schema-level matching tasks.

Gnu-RL: A Practical and Scalable Reinforcement Learning Solution for Building HVAC Control Using a Differentiable MPC Policy

We propose Gnu-RL: a novel approach that enables real-world deployment of reinforcement learning (RL) for building control and requires no prior information other than historical data from existing Heating Ventilation and Air Conditioning (HVAC) controllers. In contrast with existing RL agents, which are opaque to expert interrogation and need ample training to achieve reasonable control performance, Gnu-RL is much more attractive for real-world applications. Furthermore, Gnu-RL avoids the need to develop and calibrate simulation environments for each building or system under control, thus making it highly scalable. To achieve this, we bootstrap the Gnu-RL agent with domain knowledge and expert demonstration. Specifically, Gnu-RL adopts a recently-developed Differentiable Model Predictive Control (MPC) policy, which encodes domain knowledge on planning and system dynamics, making it both sample-efficient and interpretable. Prior to any interaction with the environment, a Gnu-RL agent is pre-trained on historical data using imitation learning, enabling it to match the behavior of the existing controller. Once it is put in charge of controlling the environment, the agent continues to improve its policy end-to-end, using a policy gradient algorithm. We evaluate Gnu-RL in both simulation studies and a real-world testbed. Firstly, we validated the Gnu-RL agent is indeed practical and scalable by applying it to three commercial reference buildings in simulation, and demonstrated the superiority of the Differentiable MPC policy compared to a generic neural network. In another simulation experiment, our approach saved 6.6% of the total energy compared to the best published RL result for the same environment, while maintaining a higher level of occupant comfort. Finally, Gnu-RL was deployed to control the airflow to a real-world conference room with a Variable-Air-Volume (VAV) system during a three-week period. Our results show that Gnu-RL reduced cooling demand by 16.7% compared to the existing controller and tracked the temperature set-point better.

Multi-turn intent determination and slot filling with neural networks and regular expressions

Intent determination and slot filling are two prominent research areas related to natural language understanding (NLU). In a multi-turn NLU system, contextual information from dialogue history is exploited to mitigate the ambiguity of user utterance. State-of-the-art models employ memory networks to encode dialogue context, which is used by neural networks for determining user intent and associated slots. However, these methods rely on a large amount of labelled data, whereas we often have limited labelled data. To address this problem, we propose a multi-task learning model based on neural networks and regular expressions (REs), to jointly perform intent determination and slot filling tasks. The proposed model integrates neural networks with REs to encode domain knowledge and handle cases with a limited amount of labelled data in an end-to-end trainable manner. More specifically, the model employs a pre-trained BERT model to obtain contextual word representations of user utterances. These representations are utilized by a memory network to encode multi-turn information which is shared by the tasks. Furthermore, the convolutional neural network (CNN) and the recurrent neural network (RNN) are applied to contextual word representations and dialogue context for intent determination and slot filling tasks, respectively. These neural networks are then combined with REs which encode domain knowledge about a particular intent or slot value. Finally, the two neural networks are trained simultaneously by minimizing the joint loss. Extensive experiments on Key-Value Retrieval and Frames datasets show that the proposed model outperforms baseline methods in both tasks while requiring modest human effort.

Learning to rank developers for bug report assignment

Bug assignment is a burden for projects receiving many bug reports. To automate the process of assigning bug reports to the appropriate developers, several studies have relied on combining natural language processing and information retrieval techniques to extract two categories of features. One of these categories targets developers who have fixed similar bugs before, and the other determines developers working on source files similar to the description of the bug. Commit messages represent another rich source for profiling developer expertise as the language used in commit messages is closer to that used in bug reports.In this work, we propose a more enhanced profiling of developers through their commits, which are captured in a new set of features that we combine with features used in previous studies. More precisely, we propose an adaptive ranking approach that takes as input a given bug report and ranks the top developers who are most suitable to fix it. This approach learns from the history of previously fixed bugs to profile developers in terms of their expertise. With respect to a given bug report, the ranking score of each developer is computed as a weighted combination of an array of features encoding domain knowledge, where the weights are trained automatically on previously solved bug reports using a learning-to-rank technique. Our model was evaluated using around 22,000 bug reports, exported from four large scale open-source Java projects. Results show that our model significantly outperformed two recent state-of-the-art methods in recommending the suitable developer to handle a certain bug report. Specifically, the percentage of recommending a developer within the top 5 ranked developers correctly was over 80% for both the Eclipse UI Platform and Birt projects.

Granular space, knowledge-encoded deep learning architecture and remote sensing image classification

Hand-crafted features of remotely sensed (RS) image require the involvement of expensive human experts for classification. This factor motivates for designing the classification model with representative feature learning-based deep architecture to automate the feature extraction process and improve the generalization capability of the model. With this reasoning, we propose a deep auto-encoder neural network (NN) architecture with knowledge-encoded granular space for the classification of RS images. The network works with wavelet-rough granulated spaces and its architecture is designed with the encoded domain knowledge that strategically initializes the network parameters. Mostly, the learning time and performance of deep auto-encoders are persuaded by randomly selected weights and thus, we aim here to minimize these efforts with the domain knowledge. Neighborhood rough sets (NRS) are used to encode the domain knowledge and explore the contextual information for improved decision. We perform the knowledge-encoding operation for all stages of the auto-encoder. The proposed model thus exploits the mutual merits of deep network, wavelet-rough granular space and knowledge-encoding method. Comparative experimental results with multispectral and hyperspectral RS images demonstrate the superiority of our model to the related advanced methods.

Machine Learning of Discriminative Gate Locations for Clinical Diagnosis.

High‐throughput single‐cell cytometry technologies have significantly improved our understanding of cellular phenotypes to support translational research and the clinical diagnosis of hematological and immunological diseases. However, subjective and ad hoc manual gating analysis does not adequately handle the increasing volume and heterogeneity of cytometry data for optimal diagnosis. Prior work has shown that machine learning can be applied to classify cytometry samples effectively. However, many of the machine learning classification results are either difficult to interpret without using characteristics of cell populations to make the classification, or suboptimal due to the use of inaccurate cell population characteristics derived from gating boundaries. To date, little has been done to optimize both the gating boundaries and the diagnostic accuracy simultaneously. In this work, we describe a fully discriminative machine learning approach that can simultaneously learn feature representations (e.g., combinations of coordinates of gating boundaries) and classifier parameters for optimizing clinical diagnosis from cytometry measurements. The approach starts from an initial gating position and then refines the position of the gating boundaries by gradient descent until a set of globally‐optimized gates across different samples are achieved. The learning procedure is constrained by regularization terms encoding domain knowledge that encourage the algorithm to seek interpretable results. We evaluate the proposed approach using both simulated and real data, producing classification results on par with those generated via human expertise, in terms of both the positions of the gating boundaries and the diagnostic accuracy. © 2019 The Authors. Cytometry Part A published by Wiley Periodicals, Inc. on behalf of International Society for Advancement of Cytometry.

Inverse reinforcement learning for intelligent mechanical ventilation and sedative dosing in intensive care units

BackgroundReinforcement learning (RL) provides a promising technique to solve complex sequential decision making problems in health care domains. To ensure such applications, an explicit reward function encoding domain knowledge should be specified beforehand to indicate the goal of tasks. However, there is usually no explicit information regarding the reward function in medical records. It is then necessary to consider an approach whereby the reward function can be learned from a set of presumably optimal treatment trajectories using retrospective real medical data. This paper applies inverse RL in inferring the reward functions that clinicians have in mind during their decisions on weaning of mechanical ventilation and sedative dosing in Intensive Care Units (ICUs).MethodsWe model the decision making problem as a Markov Decision Process, and use a batch RL method, Fitted Q Iterations with Gradient Boosting Decision Tree, to learn a suitable ventilator weaning policy from real trajectories in retrospective ICU data. A Bayesian inverse RL method is then applied to infer the latent reward functions in terms of weights in trading off various aspects of evaluation criterion. We then evaluate how the policy learned using the Bayesian inverse RL method matches the policy given by clinicians, as compared to other policies learned with fixed reward functions.ResultsResults show that the inverse RL method is capable of extracting meaningful indicators for recommending extubation readiness and sedative dosage, indicating that clinicians pay more attention to patients’ physiological stability (e.g., heart rate and respiration rate), rather than oxygenation criteria (FiO2, PEEP and SpO2) which is supported by previous RL methods. Moreover, by discovering the optimal weights, new effective treatment protocols can be suggested.ConclusionsInverse RL is an effective approach to discovering clinicians’ underlying reward functions for designing better treatment protocols in the ventilation weaning and sedative dosing in future ICUs.

A versatile approach for ranking and modelling of non-functional requirements

To effectively encode domain knowledge of customers and implementation strategies of developers in critical computer based system (CBS), requirement engineering plays a significant role. For development of quality software, it is indeed necessary to specify both functional and non-functional requirements (NFRs). However, unlike functional requirements, NFRs are not given much importance and are often generically captured which results in missing out on quality parameters during CBS development. This lacuna can be taken care with inclusion of details regarding NFRs at initial stages of CBS development. NFRs can be addressed by use of interpretive structural modelling (ISM) and analytical hierarchy process (AHP) methods for identification of critical NFRs. Further, new artifacts have been introduced in use case diagram and reference model to document and validate the identified critical NFRs. Each proposed artefact in software requirement specification (SRS) document was initially analysed individually and then compared with other noticeable methodologies.

Formal model-driven engineering of critical information systems

Model-driven engineering is the generation of software artefacts from abstract models. This is achieved through transformations that encode domain knowledge and implementation strategies. The same transformations can be used to produce quite different systems, or to produce successive versions of the same system. A model-driven approach can thus reduce the cost of development. It can also reduce the cost of verification: if the transformations are shown or assumed to be correct, each new system or version can be verified in terms of its model, rather than its implementation. This paper introduces an approach to model-driven engineering that is particularly suited to the development of critical information systems. The language of the models, and the language of the transformations, are amenable to formal analysis. The transformation strategy, and the associated development methodology, are designed to preserve systems integrity and availability.