Engineering topics

Coletti, Mark

Publications and source records attributed to Coletti, Mark.

33 records · Page 2

DOE OSTI2022

Training reinforcement learning models via an adversarial evolutionary algorithm

When training for control problems, more episodes used in training usually leads to better generalizability, but more episodes also requires significantly more training time. There are a variety of approaches for selecting the way that training episodes are chosen, including fixed episodes, uniform sampling, and stochastic sampling, but they can all leave gaps in the training landscape. In this work, we describe an approach that leverages an adversarial evolutionary algorithm to identify the worst performing states for a given model. We then use information about these states in the next cycle of training; this process can be repeated until the desired level of model performance is met. We demonstrate this approach with the OpenAI Gym cart-pole problem. With this problem, we show that the adversarial evolutionary algorithm did not reduce the number of episodes required in training needed to attain model generalizability when compared with stochastic sampling, and actually performed slightly worse.

Coletti, Mark↗

DOE OSTI2022

Proteome-scale Deployment of Protein Structure Prediction Workflows on the Summit Supercomputer

Deep learning has contributed to major advances in the prediction of protein structure from sequence, a fundamental problem in structural bioinformatics. With predictions now approaching the accuracy of crystallographic experiments, and with accelerators like GPUs and TPUs making inference using large models rapid, genome-level structure prediction becomes an obvious aim. Leadership-class computing resources can be used to perform genome-scale protein structure prediction using state-of-the-art deep learning models, providing a wealth of new data for systems biology applications. Here we describe our efforts to efficiently deploy the AlphaFold v.2 program, for full-proteome structure prediction, at scale on the Oak Ridge Leadership Computing Facility's resources, including the Summit supercomputer. We performed inference to produce the predicted structures for 40,526 protein sequences, corresponding to four prokaryotic proteomes and one plant proteome, using under 4,400 total Summit node hours, equivalent to using the majority of the supercomputer for a little over one hour. We also designed an optimized structure refinement that reduced the time for the relaxation stage of the AlphaFold pipeline by over 10X for longer sequences. We demonstrate the types of analyses that can be performed on proteome-scale collections of sequences, including a search for novel quaternary structures and implications for functional annotation.

Gao, Mu↗

DOE OSTI2022

Proteome-scale Structure Prediction Data - Pseudodesulfovibrio mercurii

The number of proteins predicted for Pseudodesulfovibrio mercurii is 3,446, each of which have five predicted structures from an AlphaFold run, as well as structural alignment results using the TMscore-based structural alignment method within the APoc program. Specifically, AlphaFold outputs the atoms and coordinates of the protein model in human-readable PDB files and quantitative prediction metrics in Python PICKLE files. The 5 models have been ranked based on the predicted TM-score (pTMS), a quantitative confidence metric output by AlphaFold that reports on protein model quality. The top ranked model has undergone an energy minimization calculation to relax and remove any potential clashes in the atomic coordinates. Structural alignment results are stored in two files for each protein; the top ranked model (as discussed above) is used for all alignment analyses. Both are compressed gzip files that, once unpacked, are human readable. The first file is the TMalign score results and contains the quantitative metrics for the top alignments between the predicted structure and experimental structures from the PDB70, a curated non-redundant database of about 80,000 experimental structures developed by the Soding lab. Each data point in this file is directly associated with one experimental structure; PDB ID and brief meta-data about the protein taken from the PDB70 file are reported alongside the quantitative metrics. The second results file contains the raw results associated with each alignment reported in the score results file. Specifically, the translation and rotation arrays for each alignment are provided so that the structural alignment can be recreated. Additionally, residue-level scores are reported to quantify the closeness of the aligned residues between the predicted and experimental models.

59 BASIC BIOLOGICAL SCIENCES↗

DOE OSTI2022

Gremlin

Gremlin is an evolutionary algorithm that discovers weaknesses in machine learning models.

Coletti, Mark↗

DOE OSTI2022

Smoky Mountain Data Challenge 2021: An Open Call to Solve Scientific Data Challenges Using Advanced Data Analytics and Edge Computing

The 2021 Smoky Mountains Computational Sciences and Engineering Conference enlists scientists from across Oak Ridge National Laboratory (ORNL) and industry to be data sponsors and help create data analytics and edge computing challenges for eminent datasets in a variety of scientific domains. This work describes the significance of each of the eight datasets and their associated challenge questions. The challenge questions for each dataset were required to cover multiple difficulty levels. An international call for participation was sent to students, asking them to form teams of up to six people and apply novel data analytics and edge computing methods to solve these challenges.

Devineni, Pravallika↗

DOE OSTI2021

Avoiding Excess Computation in Asynchronous Evolutionary Algorithms

Asynchronous evolutionary algorithms are becoming increasingly popular as a means of making full use of many processors while solving computationally expensive search and optimization problems. These algorithms excel at keeping large clusters fully utilized, but may sometimes inefficiently sample an excess of fast-evaluating solutions at the expense of higher-quality, slow-evaluating ones. We introduce a steady-state parent selection strategy, SWEET (“Selection whilE EvaluaTing”), that sometimes selects individuals that are still being evaluated and allows them to reproduce early. This gives slow-evaluating individuals that have higher fitnesses an increased ability to multiply in the population. We find that SWEET appears effective in simulated take-over time analysis, but that its benefit is confined mostly to early in the run, and our preliminary study on an autonomous vehicle controller problem that involves tuning a spiking neural network proves inconclusive.

Scott, Eric↗

DOE OSTI2021

Quantitative Evaluation of Autonomous Driving in CARLA

There has been a great deal of recent advancements in end-to-end imitation and reinforcement learning for self-driving vehicles. Despite this, there is a severe lack of standardized metrics for evaluating the performance of autonomous self-driving agents. Existing metrics are generally lacking in their ability to capture a wide range of driving behaviors and compare the severity of different failure cases. In this work, we introduce the Quantitative Evaluation for Driving metric, or QED, which assigns a quantitative score from 0-100 that captures the quality of driving for any driving agent. Our QED metric assesses different aspects of driving behavior including the ability to stay in the center of the lane, avoid weaving and erratic behavior, follow the speed limit, and avoid collisions, and it can be used under a wide range of driving scenarios. To show the effectiveness of our QED metric, we compare the scores generated by QED against scores assigned by human evaluators on a total of 30 different drivers and 6 different towns in the CARLA driving simulator. In ``easy'' evaluation scenarios, where it is relatively straightforward to distinguish better drivers from worse drivers, QED attains an average Pearson correlation of 0.96 and average Spearman correlation of 0.97 when compared against human evaluators. In ``hard'' evaluation scenarios, where it is far more ambiguous how to rank/score different types of bad driving behavior, QED attains an average Pearson correlation of 0.82 and average Spearman correlation of 0.75 when compared against human evaluators, which are both slighter higher than when we compare human evaluators against each other. While QED may not capture every characteristic that defines good driving, we consider it an important foundation for reproducibility and standardization in the community.

Gao, Shang↗

DOE OSTI2021

Diagnosing autonomous vehicle driving criteria with an adversarial evolutionary algorithm

We repurposed an adversarial evolutionary algorithm, Gremlin, from finding driving scenarios where a model of an autonomous vehicle drove poorly to troubleshooting driving quality evaluation criteria. We evaluated the driving performance of a "perfect driver" robot in a virtual town environment using the same fitness criteria intended for a deep learner (DL) trained driver. We found that the fitness evaluation criteria poorly handled turns, and used Gremlin to iteratively improve that criteria. We were confident that the same criteria could then be applied to the DL-based models as originally intended, and that this approach could be used as a general means of troubleshooting autonomous vehicle driving criteria.

Coletti, Mark↗

DOE OSTI2021

Multi-Objective Hyperparameter Optimization for Spiking Neural Network Neuroevolution

Neuroevolution has had significant success over recent years, but there has been relatively little work applying neuroevolution approaches to spiking neural networks (SNNs). SNNs are a type of neural network that includes temporal processing component, are not easily trained using other methods, and can be deployed into energy-efficient neuromorphic hardware. In this work, we investigate two evolutionary approaches for training SNNs. We explore the impact of the hyperparameters of the evolutionary approaches, including tournament size, population size, and representation type, on the performance of the algorithms. We present a multi-objective Bayesian-based hyperparameter optimization approach to tune the hyperparameters to produce the most accurate and smallest SNNs. We show that the hyperparameters can significantly affect the performance of these algorithms. We also perform sensitivity analysis and demonstrate that every hyperparameter value has the potential to perform well, assuming other hyperparameter values are set correctly.

Parsa, Maryam↗

DOE OSTI2021

2021 Smoky Mountains Conference Data Challenge Synthetic-to-Real Domain Adaptation for Autonomous Driving Dataset

The dataset is comprised of both real and synthetic images from a vehicle's forward-facing camera. Each camera image is accompanied by a corresponding pixel-level semantic segmentation image (all files are .png files). In total, the dataset contains 5600 images in the training/validation set and 1400 images in the testing set. The training dataset contains mostly synthetic RGB images collected with a wide range of weather and lighting conditions using the CARLA simulator [1]. In addition, the training data also includes a small pre-selected subset of data from the Cityscapes training dataset â€“ which is comprised of RGB-segmentation image pairs from driving scenarios in various European cities [2]. The testing data is split into three sets. The first set contains synthetic CARLA images with weather/lighting conditions that were not present in the training set. The second set is a subset of the Cityscapes testing dataset. Finally, the third set is an unknown testing set which will not be revealed to the participants until after the submission deadline. [1] Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., and Koltun, V. (2017, October). CARLA: An open urban driving simulator. In Conference on robot learning (pp. 1-16). PMLR. [2] Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., ... and Schiele, B. (2016). The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 3213-3223).

99 GENERAL AND MISCELLANEOUS↗

DOE OSTI2020

Library for Evolutionary Algorithms in Python (LEAP)

There are generally three types of scientific software users: users that solve problems using existing science software tools, researchers that explore new approaches by extending existing code, and educators that teach students scientific concepts. Python is a general-purpose programming language that is accessible to beginners, such as students, but also as a language that has a rich scientific programming ecosystem that facilitates writing research software. Additionally, as high-performance computing (HPC) resources become more readily available, software support for parallel processing becomes more relevant to scientific software.There currently are no Python-based evolutionary computation frameworks that support all three types of scientific software users. Moreover, some support synchronous concurrent fitness evaluation that do not efficiently use HPC resources. We pose here a new Python-based EC framework that uses an established generalized unified approach to EA concepts to provide an easy to use toolkit for users wishing to use an EA to solve a problem, for researchers to implement novel approaches, and for providing a low-bar to entry to EA concepts for students. Additionally, this toolkit provides a scalable asynchronous fitness evaluation implementation friendly to HPC that has been vetted on hardware ranging from laptops to the world’s fastest supercomputer, Summit.

Coletti, Mark↗

DOE OSTI2019

Evolving Larger Convolutional Layer Kernel Sizes for a Settlement Detection Deep-Learner on Summit

Deep-learner hyper-parameters, such as kernel sizes, batch sizes, and learning rates, can significantly influence the quality of trained models. The state of the art for finding optimal hyper-parameters generally uses a brute force, grid search approach, random search, or Bayesian-based optimization among other techniques. We applied an evolutionary algorithm to optimize kernel sizes for a convolutional neural network used to detect settlements in satellite imagery. Usually convolutional layer kernel sizes are small - typically one, three, or five - but we found that the system converged at, or near, kernel sizes of nine for the last convolutional layer, and that this occurred for multiple runs using two different datasets. Moreover, the larger kernel sizes had fewer false positives than the 3x3 kernel sizes found as optimal via a brute force uniform grid search. This suggests that this large kernel size may be leveraging patterns found in larger areal features in the source imagery, and that this may be generalized as possible guidance for similar remote sensing deep-learning tasks.

Coletti, Mark↗

DOE OSTI2019

Performance analysis and optimization for scalable deployment of deep learning models for country-scale settlement mapping on Titan supercomputer

Here, we present a scalable object detection workflow for detecting objects, such as settlements, from remotely sensed (RS) imagery. We have successfully deployed this workflow on Titan supercomputer and utilized it for the task of mapping human settlement at a country scale. The performance of various stages in the workflow was analyzed before making it operational. The workflow implemented various strategies to address issues such as suboptimal resource utilization and long-tail effects due to unbalanced image workload, data loss due to runtime failures, and maximum wall-time constraints imposed by Titan's job scheduling policy. A mean shift clustering–based static load balancing strategy was implemented, which partitions the image load such that each partition contained similar-sized images. Furthermore, a checkpoint-restart strategy was added in the workflow as a fault-tolerance mechanism to prevent the data losses due to unforeseen runtime failures. The performance of the above-mentioned strategies was observed in various scenarios, such as node failure, exceeding wall time, and successful completion. Using this workflow, we have examined an RS data set that has a spatial resolution of 0.31 m and is comprised of 685 675 km 2 of area of the Republic of Zambia in under six hours using 5426 nodes of the Titan supercomputer.

97 MATHEMATICS AND COMPUTING↗

DOE OSTI2018

Ramifications of Evolving Misbehaving Convolutional Neural Network Kernel and Batch Sizes

Deep-learners have many hyper-parameters including learning rate, batch size, kernel size - all playing a significant role toward estimating high quality models. Discovering useful hyper-parameter guidelines is an active area of research, though the state of the art generally uses a brute force, uniform grid approach or random search for finding ideal settings. We share the preliminary results of using an alternative approach to deep learner hyper-parameter tuning that uses an evolutionary algorithm to improve the accuracy of a deep-learner models used in satellite imagery building footprint detection. We found that the kernel and batch size hyper-parameters surprisingly differed from sizes arrived at via a brute force uniform grid approach. These differences suggest a novel role for evolutionary algorithms in determining the number of convolution layers, as well as smaller batch sizes in improving deep-learner models.

Coletti, Mark↗

DOE OSTI2017

Validating Safecast data by comparisons to a U. S. Department of Energy Fukushima Prefecture aerial survey

Safecast is a volunteered geographic information (VGI) project where the lay public collects radiation measurements via hand-held sensors that are uploaded to a central server and then made freely available. However, Safecast data fidelity is uncertain given that the these sensors are hand assembled with various levels of technical proficiency, and that the sensors may not be properly deployed. We validated Safecast data gathered in the Fukushima Prefecture shortly after the Daiichi nuclear power plant catastrophe by comparing it with authoritative data collected by the Department of Energy (DOE) and the National Nuclear Security Administration (NNSA) during the same period. We found that the two data sets were highly correlated, though the DOE/NNSA observations were generally higher than the Safecast measurements. We concluded that this high correlation alone makes Safecast a viable data source for detecting and monitoring fallout during radiological emergencies.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗