Showing papers by "Soumik Sarkar published in 2020"

Computer vision and machine learning enabled soybean root phenotyping pipeline

Abstract: Root system architecture (RSA) traits are of interest for breeding selection; however, measurement of these traits is difficult, resource intensive, and results in large variability. The advent of computer vision and machine learning (ML) enabled trait extraction and measurement has renewed interest in utilizing RSA traits for genetic enhancement to develop more robust and resilient crop cultivars. We developed a mobile, low-cost, and high-resolution root phenotyping system composed of an imaging platform with computer vision and ML based segmentation approach to establish a seamless end-to-end pipeline - from obtaining large quantities of root samples through image based trait processing and analysis. This high throughput phenotyping system, which has the capacity to handle hundreds to thousands of plants, integrates time series image capture coupled with automated image processing that uses optical character recognition (OCR) to identify seedlings via barcode, followed by robust segmentation integrating convolutional auto-encoder (CAE) method prior to feature extraction. The pipeline includes an updated and customized version of the Automatic Root Imaging Analysis (ARIA) root phenotyping software. Using this system, we studied diverse soybean accessions from a wide geographical distribution and report genetic variability for RSA traits, including root shape, length, number, mass, and angle. This system provides a high-throughput, cost effective, non-destructive methodology that delivers biologically relevant time-series data on root growth and development for phenomics, genomics, and plant breeding applications. This phenotyping platform is designed to quantify root traits and rank genotypes in a common environment thereby serving as a selection tool for use in plant breeding. Root phenotyping platforms and image based phenotyping are essential to mirror the current focus on shoot phenotyping in breeding efforts.

Prognostics of Combustion Instabilities from Hi-speed Flame Video using A Deep Convolutional Selective Autoencoder

Abstract: The thermo-acoustic instabilities arising in combustion processes cause significant deterioration and safety issues in various human-engineered systems such as land and air based gas turbine engines. The phenomenon is described as selfsustaining and having large amplitude pressure oscillations with varying spatial scales of periodic coherent vortex shedding. Early detection and close monitoring of combustion instability are the keys to extending the remaining useful life (RUL) of any gas turbine engine. However, such impending instability to a stable combustion is extremely difficult to detect only from pressure data due to its sudden (bifurcationtype) nature. Toolchains that are able to detect early instability occurrence have transformative impacts on the safety and performance of modern engines. This paper proposes an endto- end deep convolutional selective autoencoder approach to capture the rich information in hi-speed flame video for instability prognostics. In this context, an autoencoder is trained to selectively mask stable flame and allow unstable flame image frames. Performance comparison is done with a wellknown image processing tool, conditional random field that is trained to be selective as well. In this context, an informationtheoretic threshold value is derived. The proposed framework is validated on a set of real data collected from a laboratory scale combustor over varied operating conditions where it is shown to effectively detect subtle instability features as a combustion process makes transition from stable to unstable region.

Spatiotemporally Constrained Action Space Attacks on Deep Reinforcement Learning Agents

Abstract: Robustness of Deep Reinforcement Learning (DRL) algorithms towards adversarial attacks in real world applications such as those deployed in cyber-physical systems (CPS) are of increasing concern. Numerous studies have investigated the mechanisms of attacks on the RL agent's state space. Nonetheless, attacks on the RL agent's action space (corresponding to actuators in engineering systems) are equally perverse, but such attacks are relatively less studied in the ML literature. In this work, we first frame the problem as an optimization problem of minimizing the cumulative reward of an RL agent with decoupled constraints as the budget of attack. We propose the white-box Myopic Action Space (MAS) attack algorithm that distributes the attacks across the action space dimensions. Next, we reformulate the optimization problem above with the same objective function, but with a temporally coupled constraint on the attack budget to take into account the approximated dynamics of the agent. This leads to the white-box Look-ahead Action Space (LAS) attack algorithm that distributes the attacks across the action and temporal dimensions. Our results showed that using the same amount of resources, the LAS attack deteriorates the agent's performance significantly more than the MAS attack. This reveals the possibility that with limited resource, an adversary can utilize the agent's dynamics to malevolently craft attacks that causes the agent to fail. Additionally, we leverage these attack strategies as a possible tool to gain insights on the potential vulnerabilities of DRL agents.

Robustifying Reinforcement Learning Agents via Action Space Adversarial Training

Abstract: Adoption of machine learning (ML)-enabled cyber-physical systems (CPS) are becoming prevalent in various sectors of modern society such as transportation, industrial, and power grids. Recent studies in deep reinforcement learning (DRL) have demonstrated its benefits in a large variety of data-driven decisions and control applications. As reliance on ML-enabled systems grows, it is imperative to study the performance of these systems under malicious state and actuator attacks. Traditional control systems employ resilient/fault-tolerant controllers that counter these attacks by correcting the system via error observations. However, in some applications, a resilient controller may not be sufficient to avoid a catastrophic failure. Ideally, a robust approach is more useful in these scenarios where a system is inherently robust (by design) to adversarial attacks. While robust control has a long history of development, robust ML is an emerging research area that has already demonstrated its relevance and urgency. However, the majority of robust ML research has focused on perception tasks and not on decision and control tasks, although the ML (specifically RL) models used for control applications are equally vulnerable to adversarial attacks. In this paper, we show that a well-performing DRL agent that is initially susceptible to action space perturbations (e.g. actuator attacks) can be robustified against similar perturbations through adversarial training.

Soybean Root System Architecture Trait Study through Genotypic, Phenotypic, and Shape-Based Clusters

Abstract: We report a root system architecture (RSA) traits examination of a larger scale soybean accession set to study trait genetic diversity. Suffering from the limitation of scale, scope, and susceptibility to measurement variation, RSA traits are tedious to phenotype. Combining 35,448 SNPs with an imaging phenotyping platform, 292 accessions (replications = 14) were studied for RSA traits to decipher the genetic diversity. Based on literature search for root shape and morphology parameters, we used an ideotype-based approach to develop informative root (iRoot) categories using root traits. The RSA traits displayed genetic variability for root shape, length, number, mass, and angle. Soybean accessions clustered into eight genotype- and phenotype-based clusters and displayed similarity. Genotype-based clusters correlated with geographical origins. SNP profiles indicated that much of US origin genotypes lack genetic diversity for RSA traits, while diverse accession could infuse useful genetic variation for these traits. Shape-based clusters were created by integrating convolution neural net and Fourier transformation methods, enabling trait cataloging for breeding and research applications. The combination of genetic and phenotypic analyses in conjunction with machine learning and mathematical models provides opportunities for targeted root trait breeding efforts to maximize the beneficial genetic diversity for future genetic gains.

Predicting county-scale maize yields with publicly available data.

Abstract: Maize (corn) is the dominant grain grown in the world. Total maize production in 2018 equaled 1.12 billion tons. Maize is used primarily as an animal feed in the production of eggs, dairy, pork and chicken. The US produces 32% of the world's maize followed by China at 22% and Brazil at 9% ( https://apps.fas.usda.gov/psdonline/app/index.html#/app/home ). Accurate national-scale corn yield prediction critically impacts mercantile markets through providing essential information about expected production prior to harvest. Publicly available high-quality corn yield prediction can help address emergent information asymmetry problems and in doing so improve price efficiency in futures markets. We build a deep learning model to predict corn yields, specifically focusing on county-level prediction across 10 states of the Corn-Belt in the United States, and pre-harvest prediction with monthly updates from August. The results show promising predictive power relative to existing survey-based methods and set the foundation for a publicly available county yield prediction effort that complements existing public forecasts.

Spatiotemporal Attention for Multivariate Time Series Prediction and Interpretation

Abstract: Multivariate time series modeling and prediction problems are abundant in many machine learning application domains. Accurate interpretation of such prediction outcomes from a machine learning model that explicitly captures temporal correlations can significantly benefit the domain experts. In this context, temporal attention has been successfully applied to isolate the important time steps for the input time series. However, in multivariate time series problems, spatial interpretation is also critical to understand the contributions of different variables on the model outputs. We propose a novel deep learning architecture, called spatiotemporal attention mechanism (STAM) for simultaneous learning of the most important time steps and variables. STAM is a causal (i.e., only depends on past inputs and does not use future inputs) and scalable (i.e., scales well with an increase in the number of variables) approach that is comparable to the state-of-the-art models in terms of computational tractability. We demonstrate our models' performance on two popular public datasets and a domain-specific dataset. When compared with the baseline models, the results show that STAM maintains state-of-the-art prediction accuracy while offering the benefit of accurate spatiotemporal interpretability. The learned attention weights are validated from a domain knowledge perspective for these real-world datasets.

Automated detection of part quality during two-photon lithography via deep learning

Abstract: Two-photon lithography (TPL) is an additive manufacturing technique for fabricating three-dimensional objects with nanoscale features. A main challenge of TPL is the routine and labor-intensive task of finding suitable light dosage parameters, i.e. writing speed and laser intensity that induce photo-polymerization within a wide variety of candidate photo-curing polymers. Another challenge is the monitoring required during fabrication. In this work, we apply machine learning (ML) models to accelerate the process of identifying optimal light dosage parameters and automate the detection of part quality. We curate TPL videos of different parts fabricated under a range of light dosage parameters using different resins and train spatial-temporal ML models on this data. Our results show that ML models can detect TPL part quality with a 95.1% accuracy in milliseconds. We also evaluate classification failures and identify two operating modes: parameter optimization and part quality detection. Last but not least, we publicly release this labelled dataset so that it may serve as a useful benchmark to the community. Our approach to process optimization and part quality detection addresses important aspects of TPL industrialization, is applicable beyond TPL and should benefit other additive manufacturing techniques with similar barriers to operating at industrial scale.

Automated trichome counting in soybean using advanced image‐processing techniques

Abstract: Premise Trichomes are hair-like appendages extending from the plant epidermis. They serve many important biotic roles, including interference with herbivore movement. Characterizing the number, density, and distribution of trichomes can provide valuable insights on plant response to insect infestation and define the extent of plant defense capability. Automated trichome counting would speed up this research but poses several challenges, primarily because of the variability in coloration and the high occlusion of the trichomes. Methods and Results We developed a simplified method for image processing for automated and semi-automated trichome counting. We illustrate this process using 30 leaves from 10 genotypes of soybean (Glycine max) differing in trichome abundance. We explored various heuristic image-processing methods including thresholding and graph-based algorithms to facilitate trichome counting. Of the two automated and two semi-automated methods for trichome counting tested and with the help of regression analysis, the semi-automated manually annotated trichome intersection curve method performed best, with an accuracy of close to 90% compared with the manually counted data. Conclusions We address trichome counting challenges including occlusion by combining image processing with human intervention to propose a semi-automated method for trichome quantification. This provides new opportunities for the rapid and automated identification and quantification of trichomes, which has applications in a wide variety of disciplines.

Deep Learning Algorithms for Detecting Combustion Instabilities

Abstract: Combustion instabilities are prevalent in a variety of systems including gas turbine engines. In this regard, the introduction of active control opens the potential for new paradigms in combustor design and optimization. However, the limited ability to detect the onset of instabilities can lead to difficulty in implementing active control approaches. Machine learning—specifically deep learning tools—may be employed to detect instabilities from various measurement and sensor data related to the combustion process. Deep learning models have recently shown remarkable potential for extraction of meaningful features from data without the need to hand-craft. As one of the early studies of deep learning for combustion instability detection, we extract sequential image frames from high-speed images of a premixed, bluff-body stabilized flame which exhibits varying levels of combustion instability. Using an efficient detection framework (based on 2-D convolutional neural networks) to detect the growth of an unstable mode can lead to effective control schemes. In addition, we apply a second deep learning framework to capture the temporal correlations in the data with corresponding learned spatial features.

Robust Deep Reinforcement Learning for Traffic Signal Control

Abstract: A traffic signal is a fundamental part of the traffic control system to reduce congestion and enhance safety. Since the inception of motorized vehicles, traffic signal controllers are put in place to coordinate and maintain traffic flow. With the number of vehicles on the road increasing exponentially, it is imperative to innovate new traffic control frameworks to cope with the high-density traffic demand. In this regard, recent advances in machine/deep learning have enabled significant progress towards reducing congestion using reinforcement learning for traffic signal control. However, most of these works are still not ready for deployment due to assumptions of perfect knowledge of the traffic environment. In reality, congestion detection or prediction systems are at best able to approximate the traffic state with significant noise. In this work, we propose a robust training framework for reinforcement learning agents that can handle such noisy approximation of the traffic states. Specifically, we show that by carefully adding synthetic perturbations to the state space, such as the queue length during training, the reinforcement learning agents can be robustified. Conceptually, our approach is similar to adversarial training schemes and can lead to successful deployment of reinforcement learning agent-based traffic signal controllers.

Query-based Targeted Action-Space Adversarial Policies on Deep Reinforcement Learning Agents

Abstract: Advances in computing resources have resulted in the increasing complexity of cyber-physical systems (CPS). As the complexity of CPS evolved, the focus has shifted from traditional control methods to deep reinforcement learning-based (DRL) methods for control of these systems. This is due to the difficulty of obtaining accurate models of complex CPS for traditional control. However, to securely deploy DRL in production, it is essential to examine the weaknesses of DRL-based controllers (policies) towards malicious attacks from all angles. In this work, we investigate targeted attacks in the action-space domain, also commonly known as actuation attacks in CPS literature, which perturbs the outputs of a controller. We show that a query-based black-box attack model that generates optimal perturbations with respect to an adversarial goal can be formulated as another reinforcement learning problem. Thus, such an adversarial policy can be trained using conventional DRL methods. Experimental results showed that adversarial policies that only observe the nominal policy's output generate stronger attacks than adversarial policies that observe the nominal policy's input and output. Further analysis reveals that nominal policies whose outputs are frequently at the boundaries of the action space are naturally more robust towards adversarial policies. Lastly, we propose the use of adversarial training with transfer learning to induce robust behaviors into the nominal policy, which decreases the rate of successful targeted attacks by 50%.

Using Machine Learning To Develop A Fully Automated Soybean Nodule Acquisition Pipeline (SNAP)

Abstract: Nodules form on plant roots through the symbiotic relationship between soybean (Glycine max L. Merr.) roots and bacteria (Bradyrhizobium japonicum), and are an important structure where atmospheric nitrogen (N2) is fixed into bio-available ammonia (NH3) for plant growth and developmental. Nodule quantification on soybean roots is a laborious and tedious task; therefore, assessment is done on a less informative qualitative scale. We report the Soybean Nodule Acquisition Pipeline (SNAP) for nodule quantification that combines RetinaNet and UNet deep learning architectures for object (i.e., nodule) detection and segmentation. SNAP was built using data from 691 unique roots from diverse soybean genotypes, vegetative growth stages, and field locations; and has a prediction accuracy of 99%. SNAP reduces the human labor and inconsistencies of counting nodules, while acquiring quantifiable traits related to nodule growth, location and distribution on roots. The ability of SNAP to phenotype nodules on soybean roots at a higher throughput enables researchers to assess the genetic and environmental factors, and their interactions on nodulation from an early development stage. The application of SNAP in research and breeding pipelines may lead to more nitrogen use efficient soybean and other legume species cultivars, as well as enhanced insight into the plant-Bradyrhizobium relationship.

Decentralized Deep Learning using Momentum-Accelerated Consensus

Abstract: We consider the problem of decentralized deep learning where multiple agents collaborate to learn from a distributed dataset. While there exist several decentralized deep learning approaches, the majority consider a central parameter-server topology for aggregating the model parameters from the agents. However, such a topology may be inapplicable in networked systems such as ad-hoc mobile networks, field robotics, and power network systems where direct communication with the central parameter server may be inefficient. In this context, we propose and analyze a novel decentralized deep learning algorithm where the agents interact over a fixed communication topology (without a central server). Our algorithm is based on the heavy-ball acceleration method used in gradient-based optimization. We propose a novel consensus protocol where each agent shares with its neighbors its model parameters as well as gradient-momentum values during the optimization process. We consider both strongly convex and non-convex objective functions and theoretically analyze our algorithm's performance. We present several empirical comparisons with competing decentralized learning methods to demonstrate the efficacy of our approach under different communication topologies.

Traffic System Anomaly Detection using Spatiotemporal Pattern Networks

Abstract: Traffic dynamics in the urban interstate system are critical in terms of highway safety and mobility. This paper proposes a systematic data mining technique to detect traffic system-level anomalies in a batch-processing fashion. Built on the concepts of symbolic dynamics, a spatiotemporal pattern network (STPN) architecture is developed to capture the system characteristics. This novel spatiotemporal graphical modeling approach is shown to be able to extract salient time series features and discover spatial and temporal patterns for a traffic system. An information-theoretic metric is used to quantify the causal relationships between sub-systems. By comparing the structural similarity of the information-theoretic metrics of the STPNs learnt from each day, a day with anomalous system characteristics can be identified. A case study is conducted on an urban interstate in Iowa, USA, with 11 roadside radar sensors collecting 20-second resolution speed and volume data. After applying the proposed methods on one-month data (Feb. 2017), several system-level anomalies are detected. The potential causes that include inclement weather condition and non-recurring congestion are also verified to demonstrate the efficacies of the proposed technique. Compared to the traditional predefined performance measures for the traffic systems, the proposed framework has advantages in capturing spatiotemporal features in a fast and scalable manner.

InvNet: Encoding Geometric and Statistical Invariances in Deep Generative Models

Abstract: Generative Adversarial Networks (GANs), while widely successful in modeling complex data distributions, have not yet been sufficiently leveraged in scientific computing and design. Reasons for this include the lack of flexibility of GANs to represent discrete-valued image data, as well as the lack of control over physical properties of generated samples. We propose a new conditional generative modeling approach (InvNet) that efficiently enables modeling discrete-valued images, while allowing control over their parameterized geometric and statistical properties. We evaluate our approach on several synthetic and real world problems: navigating manifolds of geometric shapes with desired sizes; generation of binary two-phase materials; and the (challenging) problem of generating multi-orientation polycrystalline microstructures.

Deep Generative Models that Solve PDEs: Distributed Computing for Training Large Data-Free Models.

Abstract: Recent progress in scientific machine learning (SciML) has opened up the possibility of training novel neural network architectures that solve complex partial differential equations (PDEs). Several (nearly data free) approaches have been recently reported that successfully solve PDEs, with examples including deep feed forward networks, generative networks, and deep encoder-decoder networks. However, practical adoption of these approaches is limited by the difficulty in training these models, especially to make predictions at large output resolutions ($\geq 1024 \times 1024$). Here we report on a software framework for data parallel distributed deep learning that resolves the twin challenges of training these large SciML models - training in reasonable time as well as distributing the storage requirements. Our framework provides several out of the box functionality including (a) loss integrity independent of number of processes, (b) synchronized batch normalization, and (c) distributed higher-order optimization methods. We show excellent scalability of this framework on both cloud as well as HPC clusters, and report on the interplay between bandwidth, network topology and bare metal vs cloud. We deploy this approach to train generative models of sizes hitherto not possible, showing that neural PDE solvers can be viably trained for practical applications. We also demonstrate that distributed higher-order optimization methods are $2-3\times$ faster than stochastic gradient-based methods and provide minimal convergence drift with higher batch-size.

Usefulness of interpretability methods to explain deep learning based plant stress phenotyping

Abstract: Deep learning techniques have been successfully deployed for automating plant stress identification and quantification. In recent years, there is a growing push towards training models that are interpretable -i.e. that justify their classification decisions by visually highlighting image features that were crucial for classification decisions. The expectation is that trained network models utilize image features that mimic visual cues used by plant pathologists. In this work, we compare some of the most popular interpretability methods: Saliency Maps, SmoothGrad, Guided Backpropogation, Deep Taylor Decomposition, Integrated Gradients, Layer-wise Relevance Propagation and Gradient times Input, for interpreting the deep learning model. We train a DenseNet-121 network for the classification of eight different soybean stresses (biotic and abiotic). Using a dataset consisting of 16,573 RGB images of healthy and stressed soybean leaflets captured under controlled conditions, we obtained an overall classification accuracy of 95.05 \%. For a diverse subset of the test data, we compared the important features with those identified by a human expert. We observed that most interpretability methods identify the infected regions of the leaf as important features for some -- but not all -- of the correctly classified images. For some images, the output of the interpretability methods indicated that spurious feature correlations may have been used to correctly classify them. Although the output explanation maps of these interpretability methods may be different from each other for a given image, we advocate the use of these interpretability methods as `hypothesis generation' mechanisms that can drive scientific insight.

Traffic sensor health monitoring using spatiotemporal graphical modeling

Abstract: Accurate traffic sensor data is essential for traffic operation management systems and acquisition of real-time traffic surveillance data depends heavily on the reliability of the traffic sensors (e.g., wide range detector, automatic traffic recorder). Therefore, detecting the health status of the sensors in a traffic sensor network is critical for the departments of transportation as well as other public and private entities, especially in the circumstances where real-time decision is required. With the purpose of efficiently determining the sensor health status and identifying the failed sensor(s) in a timely manner, this paper proposes a graphical modeling approach called spatiotemporal pattern network (STPN). Traffic speed and volume measurement sensors are used in this paper to formulate and analyze the proposed sensor health monitoring system and historical time-series data from a network of traffic sensors on the Interstate 35 (I-35) within the state of Iowa is used for validation. Based on the validation results, we demonstrate that the proposed approach can: (i) extract spatiotemporal dependencies among the different sensors which leads to an efficient graphical representation of the sensor network in the information space, and (ii) distinguish and quantify a sensor issue by leveraging the extracted spatiotemporal relationship of the candidate sensor(s) to the other sensors in the network.

Supervisory Control and Distributed Optimization of Building Energy Systems

Abstract: Energy consumption in commercial buildings is significantly affected by the performance of heating, ventilation, and air-conditioning (HVAC) systems, which are traditionally operated using centralized controllers. HVAC control requires adjusting multiple setpoints such as chilled water temperatures and supply air temperature (SAT). Supervisory control framework in a distributed setting enables optimal HVAC operation and provides scalable solutions for optimizing energy across several scales from homes to regional areas. This paper proposes a distributed optimization framework for achieving energy efficiency in large-scale building energy systems. It is highly desirable to have building management systems that are scalable, robust, flexible, and are low cost. For addressing the scalability and flexibility, a modular problem formulation is established that decouples the distributed optimization level from local thermal zone modeling level. We leverage a recently developed generalized gossip algorithm for robust distributed optimization. The supervisory controller aims at minimizing the energy input considering occupant comfort. For validating the proposed scheme, a numerical case study based on a physical testbed in the Iowa Energy Center is presented. We show that the distributed optimization methodology outperforms the typical baseline strategy, which is a rule-based controller to set a constant supply air temperature. This paper also incorporates a software architecture based on the volttron platform, developed by the Pacific Northwest National Laboratory (PNNL), for practical implementation of the proposed framework via the BACnet system. The experimental results show that the supervisory control framework proposed in this paper can save energy by approximately 11%.

Effect of integrated nutrient management on growth and fruit yield of cucumber ( Cucumis sativus L.)

Abstract: Field experiment was conducted at AICRP on Vegetable Crops of OUAT, Bhubaneswar, Odisha during summer, 2018 to study the efficacy of different sources of nutrients on growth and yield of cucumber. The experiment was laid out in RBD replicated thrice having twelve INM modules including absolute control.The results revealed significant variations among different treatments for growth and yield in cucumber. Invariably, integrated application of 50% of RDF + FYM @ 10 t ha-1 + vermicompost @ 2 t ha-1 + biofertilizer recorded not only significantly highest vine length (1313.00 cm), primary branches vine-1 (3.00), fruit length (19.79 cm), fruit girth (15.13 cm), average fruit weight (194.13 g) and fruit vine-1 (11.07) but also fruit yield (214.05 q ha-1) than rest of the treatments. Similarly, the module also showed significantly lowest sex ratio (3.43) and maximum extended period of fruit harvesting (41.00 to 68.00 days). On the other hand, significantly lowest vegetative growth (i.e., vine length of 550.67 cm and primary branches vine-1 of 1.93), delayed in 1st fruit harvest (44.33 days), lowest yield attributing parameters (i.e., fruit length : 12.76 cm, fruit girth : 10.57 cm, average fruit weight : 92.00 g, fruits vine-1 : 4.20) and total fruit yield (53.70 q ha-1) in plots without any fertilizer and biofertilizer. The next better treatment was integrated application of 100% RDF + FYM @ 10 t ha-1 + biofertilizer for growth and yield in cucumber.Thus, it may be concluded that integrated application of 50% RDF + FYM @ 10 tha-1 + Vermicompost @ 2 t ha-1 + biofertilizer not only increases vegetative growth but also fruit yield in cucumber.

Physics-consistent deep learning for structural topology optimization.

Abstract: Topology optimization has emerged as a popular approach to refine a component's design and increasing its performance. However, current state-of-the-art topology optimization frameworks are compute-intensive, mainly due to multiple finite element analysis iterations required to evaluate the component's performance during the optimization process. Recently, machine learning-based topology optimization methods have been explored by researchers to alleviate this issue. However, previous approaches have mainly been demonstrated on simple two-dimensional applications with low-resolution geometry. Further, current approaches are based on a single machine learning model for end-to-end prediction, which requires a large dataset for training. These challenges make it non-trivial to extend the current approaches to higher resolutions. In this paper, we explore a deep learning-based framework for performing topology optimization for three-dimensional geometries with a reasonably fine (high) resolution. We are able to achieve this by training multiple networks, each trying to learn a different aspect of the overall topology optimization methodology. We demonstrate the application of our framework on both 2D and 3D geometries. The results show that our approach predicts the final optimized design better than current ML-based topology optimization methods.

How useful is Active Learning for Image-based Plant Phenotyping?

Abstract: Deep learning models have been successfully deployed for a diverse array of image-based plant phenotyping applications including disease detection and classification. However, successful deployment of supervised deep learning models requires large amount of labeled data, which is a significant challenge in plant science (and most biological) domains due to the inherent complexity. Specifically, data annotation is costly, laborious, time consuming and needs domain expertise for phenotyping tasks, especially for diseases. To overcome this challenge, active learning algorithms have been proposed that reduce the amount of labeling needed by deep learning models to achieve good predictive performance. Active learning methods adaptively select samples to annotate using an acquisition function to achieve maximum (classification) performance under a fixed labeling budget. We report the performance of four different active learning methods, (1) Deep Bayesian Active Learning (DBAL), (2) Entropy, (3) Least Confidence, and (4) Coreset, with conventional random sampling-based annotation for two different image-based classification datasets. The first image dataset consists of soybean [Glycine max L. (Merr.)] leaves belonging to eight different soybean stresses and a healthy class, and the second consists of nine different weed species from the field. For a fixed labeling budget, we observed that the classification performance of deep learning models with active learning-based acquisition strategies is better than random sampling-based acquisition for both datasets. The integration of active learning strategies for data annotation can help mitigate labelling challenges in the plant sciences applications particularly where deep domain knowledge is required.

Few Shot Clustering For Indoor Occupancy Detection With Extremely Low Quality Images From Battery Free Cameras

Abstract: Reliable detection of human occupancy in indoor environments is critical for various energy efficiency, security, and safety applications. We consider this challenge of occupancy detection using extremely low-quality, privacy-preserving images from low power image sensors. We propose a combined few shot learning and clustering algorithm to address this challenge that has very low commissioning and maintenance cost. While the few shot learning concept enables us to commission our system with a few labeled examples, the clustering step serves the purpose of online adaptation to changing imaging environment over time. Apart from validating and comparing our algorithm on benchmark datasets, we also demonstrate performance of our algorithm on streaming images collected from real homes using our novel battery free camera hardware.

Interpretable Deep Attention Model for Multivariate Time Series Prediction in Building Energy Systems

Abstract: Multivariate time series prediction has important applications in the domain of energy-efficient building technology. With the buildings consuming large amounts of electrical energy, it is critical to reducing energy consumption and economic costs while ensuring a better quality of urban living standards. As sensor-actuator rich, smart buildings are becoming complex dynamic data-driven applications systems (DDDAS), accurate and interpretable data-driven decision-making tools can have immense value. In this context, we develop a novel deep learning model that can explicitly capture the temporal correlations through LSTM layers. The model can isolate the important timesteps in the input time series for prediction. Also, it is critical to identify the contributions of different variables in the multivariate input. Our proposed model based on attention mechanisms can simultaneously learn important timesteps and variables. We demonstrate the results using a public multivariate time series dataset collected from an air handling unit in a building heating, ventilation, and air-conditioning (HVAC) system. The model with enhanced interpretability does not compromise with the prediction accuracy. The interpretations are validated from a domain knowledge perspective.

Robustifying Reinforcement Learning Agents via Action Space Adversarial Training

Abstract: Adoption of machine learning (ML)-enabled cyber-physical systems (CPS) are becoming prevalent in various sectors of modern society such as transportation, industrial, and power grids. Recent studies in deep reinforcement learning (DRL) have demonstrated its benefits in a large variety of data-driven decisions and control applications. As reliance on ML-enabled systems grows, it is imperative to study the performance of these systems under malicious state and actuator attacks. Traditional control systems employ resilient/fault-tolerant controllers that counter these attacks by correcting the system via error observations. However, in some applications, a resilient controller may not be sufficient to avoid a catastrophic failure. Ideally, a robust approach is more useful in these scenarios where a system is inherently robust (by design) to adversarial attacks. While robust control has a long history of development, robust ML is an emerging research area that has already demonstrated its relevance and urgency. However, the majority of robust ML research has focused on perception tasks and not on decision and control tasks, although the ML (specifically RL) models used for control applications are equally vulnerable to adversarial attacks. In this paper, we show that a well-performing DRL agent that is initially susceptible to action space perturbations (e.g. actuator attacks) can be robustified against similar perturbations through adversarial training.

Orthogonal Distance Fields Representation for Machine-Learning Based Manufacturability Analysis

Abstract: Computer-aided Design for Manufacturing (DFM) systems play an essential role in reducing the time taken for product development by providing manufacturability feedback to the designer before the manufacturing phase. Traditionally, DFM rules are hand-crafted and used to accelerate the engineering product design process by integrating manufacturability analysis during design. Recently, the feasibility of using a machine learning-based DFM tool in intelligently applying the DFM rules have been studied. These tools use a voxelized representation of the design and then use a 3D-Convolutional Neural Network (3D-CNN), to provide manufacturability feedback. Although these frameworks work effectively, there are some limitations to the voxelized representation of the design. In this paper, we introduce a new representation of the computer-aided design (CAD) model using orthogonal distance fields (ODF). We provide a GPU-accelerated algorithm to convert standard boundary representation (B-rep) CAD models into ODF representation. Using the ODF representation, we build a machine learning framework, similar to earlier approaches, to create a machine learning-based DFM system to provide manufacturability feedback. As proof of concept, we apply this framework to assess the manufacturability of drilled holes. The framework has an accuracy of more than 84% correctly classifying the manufacturable and non-manufacturable models using the new representation.