Staff

• 1: Geoffrey C. Fox
• 2: Gregor von Laszewski
• 3: Jacques Fleischer
• 4: Clara Grimmelbein
• 5: Junyang (Eric) He
• 6: Meryem Altin Karagoz
• 7: Nate Kimball
• 8: Varun Pavuloori
• 9: Arup Sarker
• 10: Mills W. Staylor
• 11: Christopher Zelaya
• 12: Tianle Zhong

1 - Geoffrey C. Fox

The Web Page is located at

https://biocomplexity.virginia.edu/person/geoffrey-fox

2 - Gregor von Laszewski

3 - Jacques Fleischer

Biography

Jacques is an undergraduate computer science major in the Computing for Global Challenges program of the Biocomplexity Institute at the University of Virginia. His interests include leveraging artificial intelligence to analyze timeseries; he also uses high performance computing to conduct geospatial AI predictions.

References

• von Laszewski, G., Fleischer, J. P., & Fox, G. C. (2022). Hybrid Reusable Computational Analytics Workflow Management with Cloudmesh. arXiv, 2210.16941.
• Fleischer, J. P., von Laszewski, G., Theran, C., & Parra Bautista, Y. J. (2022). Time Series Analysis of Cryptocurrency Prices Using Long Short-Term Memory. Algorithms, 15(7), 230. MDPI AG.
• Fleischer, J. P., von Laszewski, G., Theran, C., & Parra Bautista, Y. J. (2022). Time Series Analysis of Blockchain-Based Cryptocurrency Price Changes. arXiv, 2202.13874.

Project

The tevelop earthquake benchmark analyzes historical data and forecasts earthquakes with Normalized Nash-Sutcliffe Efficiency (NNSE) as a measure of accuracy. The benchmark is ran several times in parallel on several graphics cards and varying epochs to find the best hyperparameter combination.

The cloudmask benchmark reads satellite imagery and classifies portions of images as cloudy or clear sky. Validation loss and inference accuracy are used to identify which epoch number is ideal for optimal classification.

Resources

Compute Resources: 500 GPU-accelerated Slurm hours

Data Resources: at least 5 TB of space for data storage

4 - Clara Grimmelbein

---

References

• N/A

Project

Candle Uno utilizes deep learning architectures to predict tumor response to single and paired drugs. Candle combats the problems of cellular, molecular, and population problems within cancer. The project is trained with a subset of dose-response data from six data sources. Benchmark Pilot1 (P1) focuses on the molecular features of tumor cells from a singular data source, CCLE, to accelerate the use of effective drugs to cure cancerous cells.

Resources

Compute Resources:

https://github.com/ECP-CANDLE/Benchmarks/blob/develop/Pilot1/Uno/uno_baseline_keras2.py

(to initiate training)

Data Resources:

http://ftp.mcs.anl.gov/pub/candle/public/benchmarks/Pilot1/uno/top_21_auc_1fold.uno.h5

5 - Junyang (Eric) He

6 - Meryem Altin Karagoz

---

References

• KARAGOZ, M. A., AKAY, B., BASTURK, A., KARABOGA, D., & NALBANTOGLU, O. U., (2023). An unsupervised transfer learning model based on convolutional auto encoder for non-alcoholic steatohepatitis activity scoring and fibrosis staging of liver histopathological images. Neural Computing and Applications , vol.35, no.14, 10605-10619.

• KARAGOZ, M. A., KARABOGA, D., AKAY, B., BASTURK, A., & Nalbantoglu, O. U., (2022). Deep Learning Based Steatosis Quantification of Liver Histopathology Images Using Unsupervised Feature Extraction. 2nd International Conference on Computing and Machine Intelligence, ICMI 2022, İstanbul, Turkey

• KARAGOZ, M. A., & NALBANTOGLU, O. U., (2021). Taxonomic classification of metagenomic sequences from Relative Abundance Index profiles using deep learning. Biomedical Signal Processing and Control , vol.67.

• Karagoz, M. A., & Nalbantoglu, O. U., (2019). RAPID PATHOGEN DETECTION OF USING CONVOLUTIONAL NEURAL NETWORKS ON METAGENOMICS DATA. HIBIT 19- THE 12TH INTERNATIONAL SYMPOSIUM ON HEALTH INFORMATICS AND BIOINFORMATICS (pp.130-131). İzmir, Turkey

• Karagoz, M. A., & Nalbantoglu, O. U., (2018). TAXONOMIC ANALYSIS OF MICROBIOTA DATA BASED ON RELATIVE ABUNDANCE INDEX ALGORITHM. Uluslararası XXXVIII. TMC Kongresi (pp.381-382). Antalya, Turkey

Project

Title: Developing Self-Supervised Learning Models for Reducing the Dependency on Expert Supervision in Medical Image Analysis

Deep Learning methods are capable of learning highly complex mathematical functions using a large number of neural-net layers. Therefore, deep learning models require a large amount of data to learn large numbers of parameters. However, the two main problems encountered in current medical imaging studies with Deep Learning are: (i) lack of annotated datasets and (ii)imbalanced datasets between labels or classes. While limited labeled datasets cause overfitting, an imbalanced dataset between positive and negative samples causes an underfitting problem for deep networks. Thus, deep learning models are challenging to provide high performance in computer vision problems with insufficient data; medical imaging is one of these areas. The problem of generating largescale and public datasets for medical image analysis is difficult to solve in the short term due to practical difficulties such as cost, time, patient privacy, and lack of standard data publication procedures. On the other hand, the problem of an imbalanced dataset can be challenging for generating a large-scale annotated dataset. Therefore, it is an important field for medical image analysis studies that researchers develop Deep Learning strategies to overcome these data constraints by reducing the dependence on expert supervision. In recent years, self-supervised learning (SSL) has gained remarkable attention against lack of datasets owing to pretext task and downstream task mechanisms with prior studies. The pretext task is pre-designed as the former task for learning discriminative visual features on unlabeled data to be used in downstream tasks such as classification, detection, and segmentation. In addition to providing robust and performance-enhancing solutions in the lack of labeled data, self-supervised learning methods allow extraction of distinct signals by using labeled and unlabeled data all together as in an unsupervised manner during the pretext task step. Thus, SSL provides a solution to the problem of a large amount of unlabeled data, which is the most critical problem in medical image tasks. However, there is a noticeable lack of self-supervised learning models for medical image analysis tasks and needs to be improved with further studies. In conclusion, we focus on SSL models to deal with data limitations and enable automated diagnosis and detection systems at the expert level for clinical usage by reducing expert demand for medical image tasks.

Resources needs

Compute:

Data: at least 10 TB of space for data storage.

7 - Nate Kimball

---

8 - Varun Pavuloori

9 - Arup Sarker

---

10 - Mills W. Staylor

11 - Christopher Zelaya

---

References

• N/A

Project

State of the art scanning transmission electron microscopes (STEM) produce focused electron beams with atomic dimensions and allow to capture diffraction patterns arising from the interaction of incident electrons with nanoscale material volumes. Backing out the local atomic structure of said materials requires compute- and time-intensive analyses of these diffraction patterns (known as convergent beam electron diffraction, CBED). Traditional analyses of CBED requires iterative numerical solutions of partial differential equations and comparison with experimental data to refine the starting material configuration. This process is repeated anew for every newly acquired experimental CBED pattern and/or probed material.

Resources

Compute Resources: https://github.com/mlcommons/science/tree/main/benchmarks/stemdl

Data Resources: https://doi.ccs.ornl.gov/ui/doi/70

12 - Tianle Zhong