Deep learning is often referred to as a &lsquo;black box&rsquo;, because, aside from selecting model and the model hyperparameters (like the learning rate, activation function parameters), users don&rsquo;t have access to what happens inside the model. We can know that the model works by measuring its accuracy on some testing data set, suggesting that the model must have &lsquo;learned&rsquo; something meaningful from the training data. But how can we peek inside and &lsquo;see&rsquo; what the model &lsquo;sees&rsquo;? 
Biologists have recently begun to explore how deep learning approaches could be used to address questions involving diverse biological datasets. For this community, producing an accurate model is important, but also insufficient, since the larger goal is to contribute to our understanding of complex biological systems and phenomena. Visualization could help biology researchers &lsquo;peak&rsquo; inside the black box and relate deep learning mechanics to existing biological knowledge. However, most visualization techniques for deep learning have been developed for images and text, so many of the encodings and layouts are designed for input data with fundamentally different perceptual properties than biological data sets, such as gene sequence data or gene expression data. 
This research explores current approaches in visualization for deep learning and how these approaches may (or may not) be applied to biological data. This work is conducted with researchers at Argonne National Labs studying deep learning for cancer research and genomics.

Argonne Leadership Computing Facility

Vis for Deep Learning for Biology

This collaborative research project led by the colleagues at the University of Delaware aims to develop and advance tools for using image-data appearing in scientific publications, in addition to text, in order to support beneficial, targeted access to the biomedical literature. 
UIC's principal investigator, Computer Science associate professor, G. Elisabeta Marai will be responsible for the image analysis algorithms to distinguish between gel and plate images, the image analysis algorithms to distinguish between Electronic Microscopy vs Fluorescent Microscopy vs Light Microscopy and X-Rays, and the design of the visual interface to present results to end users. She will lead the design, development and implementation of the feature extraction and selection tools and techniques related to this research project.

Incorporating Image-based Features into Biomedical Document Classification

Data &Time: Friday, May 4th, 2018, 11 am - 1pm 
Location: Room 3036 ERF 
Defense committee: 
Angus G. Forbes (Chair and Advisor) 
Andrew Johnson 
G. Elisabeta Marai 
Abstract: 
Simulating real-life light in a virtual world is computationally very expensive even in a limited capacity. The quality of images synthesized using computer graphics algorithms for offline rendering used in movies has increased significantly over the years. Real-time applications such as games still rely on non-physics based approximation methods for computing light to meet the time budget to render a frame. The demand for higher quality graphics for more realism in games is high which requires a better approximation technique. Acceleration data structures are used for speed up the rendering process by skipping the unwanted parts of the virtual scene for light calculations. Even with such structures rendering can take a very long time depending on the size of the 3D scene. Recently, deep learning approaches have shown great success in image processing problems. Although neural networks are extensively used for training AI agents and other parts of a game, only a handful of research exists in the 3D rendering space. 
In this work, I explore deep learning based approaches to 1) approximate indirect illumination and soft shadow from image space buffers, and 2) build acceleration structure for faster rendering. I also show evaluations of these networks to get insights about their performance and limitations.

MS Thesis Defense Announcement: &ldquo;Exploring Deep Learning Techniques for Real-time Graphics&rdquo;

We introduce a deep learning approach for the identification of shock locations in large-scale tensor field datasets. Such datasets are typically generated by turbulent combustion simulations. In this proof of concept approach, we use deep learning to learn mappings from strain tensors to Schlieren images which serve as labels. The use of neural networks allows for the Schlieren values to be approximated more efficiently than calculating the values from the density gradient. In addition, we show that this approach can be used to predict the Schlieren values for both two-dimensional and three-dimensional tensor fields, potentially allowing for anomaly detection in tensor flows. Results on two shock example datasets show that this approach can assist in the extraction of features from reacting flow tensor fields. 
Support: 
NSF CAREER IIS-1541277

Networks for 3D Sod dataset (left) and the 2D Blast dataset (right).

A Deep Learning Approach to Identifying Shock Locations in Turbulent Combustion Tensor Fields

Research Experience Undergraduate (REU)

Zhongyi Chen

Animation studios render 3D scenes using a technique called path tracing which enables them to create high quality photorealistic frames. Path tracing involves shooting 1000&rsquo;s of rays into a pixel randomly (Monte Carlo) which will then hit the objects in the scene and, based on the reflective property of the object, these rays reflect or refract or get absorbed. The colors returned by these rays are averaged to determine the color of the pixel. This process is repeated for all the pixels. Due to the computational complexity it might take 8-16 hours to render a single frame. We implemented a neural network-based solution to reduce the time it takes to render a frame to less than a second using a generative adversarial network (GAN), once the network is trained. The main idea behind this proposed method is to render the image using a much smaller number of samples per pixel than is normal for path tracing (e.g., 1, 4, or 8 samples instead of, say, 32,000 samples) and then pass the noisy, incompletely rendered image to our network, which is capable of generating a high-quality photorealistic image. 
Keywords: image denoising, rendering, deep learning

Gaussian noise image (left), denoised image (middle) and ground truth photorealistic image (right).

Image Denoising Using A Generative Adversarial Network

Maxine Brown (UIC PI), Andrew Johnson (UIC co-PI), Luc Renambot (UIC co-PI) 
Jason Leigh (University of Hawaii Manoa PI) Christopher North (Virginia Polytechnic PI) 
NSF Award Number: 2003800 
Award Period of Performance: 05/15/2020 - 04/30/2025 
Total Amount Obligated: $2,250,000 
UIC Computer Science Electronic Visualization Lab (EVL) Director Maxine Brown, PI, co-PIs UIC CS Professor Andy Johnson and UIC CS Research Associate Professor Luc Renambot, and Senior Personnel EVL Senior Research Programmer Lance Long are awarded a $2.25 million dollar National Science Foundation grant entitled &ldquo;Collaborative Research: CSSI Frameworks: SAGE3: Smart Amplified Group Environment for Harnessing the Data Revolution.&rdquo; The grant runs from May 2020 to April 2025. It&rsquo;s part of an overall $5 million project, that also includes a team from University of Hawaii Manoa led by Jason Leigh and a team from Virginia Polytechnic led by Christopher North. 
ABSTRACT 
The Big Data revolution necessitates the use of sophisticated tools such as Artificial Intelligence (AI) and Data Visualization to harness the sheer volume, velocity and variety of datasets that are becoming the norm. However, it is the research community that must make sense of the data being amassed, so cyberinfrastructure must extend to people. SAGE3 (Smart Amplified Group Environment) puts the &ldquo;human in the loop&rdquo; by providing scientists with an intuitive framework that integrates state-of-the-art AI technologies with applications, workflows, smart visualizations and collaboration services to help them access, share, explore and analyze their data, come to conclusions, and make decisions with greater speed, accuracy, comprehension and confidence. 
SAGE3 builds upon the NSF-funded first-generation SAGE (Scalable Adaptive Graphics Environment) and second-generation SAGE2 (Scalable Amplified Group Environment), which are today&rsquo;s de facto operating systems to manage Big Data on tiled display walls. Its development focuses on two fundamental components: AI-enhanced &ldquo;smart&rdquo; services and advanced computing resource orchestration to support reproducible work models for secure collaborative work. SAGE3 amplifies user productivity, providing them with commercially available and open-source AI solutions, which autonomously and transparently analyze data while continually learning and improving through user interactions. SAGE3 makes AI technologies broadly accessible, not just a privilege for the technically savvy. SAGE3 further democratizes AI by using Data Visualization to help interpret and explain AI models so users better understand how AI came to its decisions, which engenders user trust and can help identify potentially prejudiced or biased models. 
SAGE3 augments every step of the scientific discovery enterprise - from quickly summarizing large data, to finding trends and similarities or anomalies among one or more linked datasets, to communicating findings to scientists, public policy and government officials, and the general public, to educating the next-generation workforce. Ultimately, it is the scientists and future scientists who must Harness the Big Data revolution to solve the nation&rsquo;s grand challenge problems that will benefit society as a whole - from studying the diversity of life on Earth, to understanding the Earth and its systems from satellite imagery of its poles, to developing response scenarios for natural disasters such as landslides and pandemics that impact the citizens and economies of the world.

UIC Computer Science/EVL awarded $2.25 Million from NSF for SAGE3

UIC PI: Andrew Johnson, co-PI: Luc Renambot 
University of Hawaii Manoa PI: Jason Leigh 
Virginia Polytechnic PI: Chirstopher North 
NSF Award Number: 2003800 
Award Period of Performance: 05/15/2020 - 04/30/2025 
Total Amount Obligated: $2,250,000 
The Big Data revolution necessitates the use of sophisticated tools such as Artificial Intelligence (AI) and Data Visualization to harness the sheer volume, velocity and variety of datasets that are becoming the norm. However, it is the research community that must make sense of the data being amassed, so cyberinfrastructure must extend to people. SAGE3 (Smart Amplified Group Environment) puts the &ldquo;human in the loop&rdquo; by providing scientists with an intuitive framework that integrates state-of-the-art AI technologies with applications, workflows, smart visualizations and collaboration services to help them access, share, explore and analyze their data, come to conclusions, and make decisions with greater speed, accuracy, comprehension and confidence. 
SAGE3 builds upon the NSF-funded first-generation SAGE (Scalable Adaptive Graphics Environment) and second-generation SAGE2 (Scalable Amplified Group Environment), which are today&rsquo;s de facto operating systems to manage Big Data on tiled display walls. Its development focuses on two fundamental components: AI-enhanced &ldquo;smart&rdquo; services and advanced computing resource orchestration to support reproducible work models for secure collaborative work. SAGE3 amplifies user productivity, providing them with commercially available and open-source AI solutions, which autonomously and transparently analyze data while continually learning and improving through user interactions. SAGE3 makes AI technologies broadly accessible, not just a privilege for the technically savvy. SAGE3 further democratizes AI by using Data Visualization to help interpret and explain AI models so users better understand how AI came to its decisions, which engenders user trust and can help identify potentially prejudiced or biased models. 
SAGE3 augments every step of the scientific discovery enterprise - from quickly summarizing large data, to finding trends and similarities or anomalies among one or more linked datasets, to communicating findings to scientists, public policy and government officials, and the general public, to educating the next-generation workforce. Ultimately, it is the scientists and future scientists who must Harness the Big Data revolution to solve the nation&rsquo;s grand challenge problems that will benefit society as a whole - from studying the diversity of life on Earth, to understanding the Earth and its systems from satellite imagery of its poles, to developing response scenarios for natural disasters such as landslides and pandemics that impact the citizens and economies of the world. 
<a href="https://youtu.be/39x-cOQJRVc">View a video of SAGE3 being demonstrated</a> on EVL&rsquo;s YouTube channel.

SAGE3: Smart Amplified Group Environment for Harnessing the Data Revolution

PhD Graduate Student

Mir Masood Ali

Graph convolutional networks (GCNs) have been applied to 3D human pose estimation (HPE) from 2D body joint detections and
have shown encouraging performance. One limitation of the vanilla graph convolution is that it models the relationships between neighboring nodes via a shared weight matrix. This is suboptimal for articulated body modeling as the relations between different body joints are different. The objective of this paper is to have a comprehensive and systematic study of weight sharing in GCNs for 3D HPE. We first show there are two different ways to interpret a GCN depending on whether feature transformation occurs before or after feature aggregation. These two interpretations lead to five different weight sharing methods and three more variants can be derived by decoupling the self-connections with other edges. We conduct extensive ablation study on these weight sharing methods under controlled settings and obtain new conclusions that will benefit the community.

Qualitative results of our approach on Human 3.6M

A Comprehensive Study of Weight Sharing in Graph Networks for 3D Human Pose Estimation

Graph convolutional networks (GCNs) have been applied to 3D human pose estimation (HPE) from 2D body joint detections and have demonstrated promising performance. However, since the vanilla graph convolution is performed on the one-hop neighbors of each node, it is unable to capture the long-range dependencies between body joints. They can help reduce the uncertainty caused by occlusion or depth ambiguity. To resolve this issue, we propose a high-order GCN for 3D HPE. Its core building block, termed a high-order graph convolution, aggregates features of nodes at various distances. As a result, the network can model a wide range of interactions among body joints. Furthermore, we investigate different methods to fuse those multi-order features and compare how they affect the performance. Experimental results demonstrate the effectiveness of the proposed approach.

Comparison between the vanilla graph convolution and the high-order graph convolution performed on a skeleton graph.

High-order Graph Convolutional Networks for 3D Human Pose Estimation

October 19, 2020 - - Fabio Miranda joins the EVL faculty as an assistant professor with the Department of Computer Science. Miranda interests include developing techniques that allow for the interactive visual analysis of large-scale data, combining methods from visualization, data management, machine learning, and computer graphics. Of particular concern is how visual data analytics can help address different problems cities face by integrating data on different resolutions and from different sources. 
Miranda&rsquo;s research in data science, human-computer interaction and visualization include: visual data analysis, data structures and indexing techniques for big data visualization, urban analytics, harnessing data sets of different types (images, land use information, 3D building geometry, etc.) to address concrete city problems, and machine learning. He has worked closely with domain experts from different fields, from urban planning to occupational therapy, and the outcome of these collaborations include not only research published in leading visualization, database, HCI, and AI venues, but also systems that were made available to experts in academia, industry, and government agencies. 
Miranda received his Ph.D. from New York University (2018).

Fabio Miranda Joins EVL Faculty

October 1 - - EVL Director of Research Andrew Johnson has been named Interim Director of EVL following Maxine Brown&rsquo;s retirement from UIC. Johnson now assumes the responsibility of heading EVL - overseeing EVL&rsquo;s overall mission in education, research, development, and technology transfer. 
For over 23 years, Johnson has been a valued EVL faculty member (Professor of Computer Science) and researcher who has focused upon interaction and collaboration using advanced visualization displays and the application of those displays to enhance discovery and learning. 
His areas of expertise include visualization, virtual and augmented reality, natural language processing, visual analytics, and high-performance networking. 
Over the years, Johnson has been instrumental in the development of EVL instrumentation and systems, including the Continuum smart environment, CAVE2&trade;, LambdaVision high-resolution tiled displays, LambdaTable tabletop touch system, GeoWall portable geoscience display environment, PARIS augmented reality and ImmersaDesk virtual reality systems. He has overseen the development of numerous visualization and virtual reality applications across diverse domain disciplines from science and engineering to the humanities and art. Johnson is a highly regarded mentor and educator to EVL&rsquo;s graduate students and undergraduates alike. 
Johnson retains his position as Director of Research while leading EVL&rsquo;s future initiatives and innovations as Interim Director.

Andrew Johnson Named Interim Director, EVL

For the Supercomputing (SC20) Conference this year, a research community partnership was formed to design an international networking testbed using existing facilities, with several international extensions. Demonstrations and experiments were conducted prior to the conference with results discussed during a SC20 virtual forum. &ldquo;The Forum For Experiments and Demonstrations Showcasing Innovations In Large-Scale, Data-Intensive-Science Networking&rdquo; was organized by Research Consortium for Data-Intensive Science Networking to communicate: a) the motivation behind these experiments and demonstrations, b) their design, c) technologies used, d) results and e) future research in these areas. 
EVL senior research programmer, Lance Long participated to discuss EVL&rsquo;s research in Composable Infrastructure, specifically MRI: Acquisition of a Composable Platform as a Service Instrument for Deep Learning & Visualization (COMPaaS DLV) funded under NSF Award #CNS-1828265. Items discussed included the benefits of composable infrastructure, implementation for deep learning in an academic research environment, operating systems, dynamic workflows, orchestration middleware, current use cases and future work. 
Other participants included: Ciena Research Lab, NASA Goddard Space Flight Center (GSFC), Naval Research Laboratory, iCAIR/Northwestern, StarLight, UCSD, FNAL, CalTech, SURFnet.

Composable Infrastructure for Data Science

Lance Long Presents EVL&rsquo;s Composable Infrastructure Research at SC20

Microscopy images are pervasive in biomedical research publications, where images obtained through various microscopy modalities (light, fluorescence, scanning, transmission) are often used to describe and summarize experiments and contributions. Hence, there is growing interest in automatically identifying these microscopy images’ modality and utilizing this knowledge in automated search tools. However, identifying microscopy images poses challenges due to a lack of extensive collections of labeled images. We describe and evaluate two alternative approaches to microscopy image classification. In the first approach, we progressively fine-tuned layers of ResNet models. The second approach uses shallow variants of ResNet networks, where we leverage the outputs from previous convolutional blocks. We compare these results against a Support Vector Machine (SVM)-based baseline. Our results show that fine-tuning specific layers yields better results than fine-tuning the whole model. Furthermore, shallower variants produce competitive results when compared to the entire fine-tuned model. 
Index Terms: microscopy, image classification, deep learning

Four microscopy images from the ImageCLEF dataset showcasing different modalities: (a) Electron microscopy (DMEL), (b) Transmission microscopy (DMTR), (c) Light microscopy (DMLI), and (d) Fluorescence microscopy (DMFL).

Modality-Classification of Microscopy Images Using Shallow Variants of Deep Networks

EVL faculty and Computer Science associate professor G. Elisabeta Marai (MPI) along with H. Shatkay (MPI) at the University of Delaware have been awarded supplemental funding for COVID-19 related research to their NIH research project entitled &ldquo;Uncovering Clinical Evidence in COVID-19 Publications: An Integrated Search via Text &amp; Images&rdquo;. This supplement expands the work done under current R01 funding &ldquo;Incorporating Image-based Features into Biomedical Document Classification&rdquo;, to support discovery within the CORD-19 collection -- a large, continually updated collection of more than 60,000 scientific articles pertaining to COVID-19. This supplement will create a comprehensive platform for retrieving and displaying well-focused, need-targeted medical information within COVID-19 biomedical publications. Our approach is unique compared to other text-mining methods because it leverages the image content in the dataset. 
Funding: NIH R01 NLM Covid-19 supplement to Parent Grant: 1R01LM012527 (&ldquo;Incorporating Image-based Features into Biomedical Document Classification&rdquo;)

EVL Faculty G.E. Marai Awarded NIH COVID-19 Supplement

MS Graduate, 2021

Elisa Tardini

EVL faculty and computer science (CS) associate professor Liz Marai in collaboration with co-investigator, CS assistant professor Xinhua Zhang are developing a computational artificial intelligence system to serve as a decision support tool for oncologists prescribing treatment for head and neck cancer. The research entitled &ldquo;Longitudinal Spatial-Nonspatial Decision Support for Competing Outcomes in Head and Neck Cancer Therapy&rdquo; is supported by a $2.8 million grant from the National Institutes of Health NIH grant #R01CA258827 with co-investigators at the University of Texas MD Anderson Cancer Center and the University of Iowa. The goal is to develop a scalable system based on real cancer patient data that can help guide physicians in how to treat specific patients. 
Marai and Zhang are using the computational resources at Electronic Visualization Laboratory to develop their cancer treatment AI. Guadalupe Canahuate of the University of Iowa and Dr. David Fuller of the University of Texas are co-investigators on the NIH grant. 
See the full press release on <a href="https://today.uic.edu/maximizing-cancer-survival-minimizing-treatment-side-effects-with-ai">UIC Today</a> by Jacqueline Carey.

Collaborators examine a large dataset in EVL&rsquo;s CAVE2

Maximizing cancer survival, minimizing treatment side effects with AI

Visualization, while an effective tool for identifying patterns and insights, requires expert knowledge due to challenges faced when translating user queries to visual encodings. Research has shown that using a natural language interface (NLI) is effective for these challenges because the user can simply talk to a computer capable of producing the graphs directly. In this paper, we discuss our intelligent assistant which processes speech and hand pointing gestures while also dealing with any number of visualizations on a large screen display. Evaluation of the system shows that it is capable of quickly producing visualizations. It also particularly effective
at responding to less ambiguous queries, while in certain cases can handle ambiguous or complex queries.

The System Pipeline

Intelligent Assistant for Exploring Data Visualizations

Our goal is to develop an intelligent assistant to support users explore data via visualizations. We have collected a new corpus of conversations, CHICAGO-CRIME-VIS, geared towards supporting data visualization exploration, and we have annotated it for a variety of features, including contextualized dialogue acts. In this paper, we describe our strategies and their evaluation for dialogue act classification. We highlight how thinking aloud affects interpretation of dialogue acts in our setting and how to best capture that information. A key component of our strategy is data augmentation as applied to the training data, since our corpus is inherently small. We ran experiments with the Balanced Bagging Classifier (BAGC), Condiontal Random Field (CRF), and several Long Short Term Memory (LSTM) networks, and found that all of them improved compared to the baseline (e.g., without the data augmentation pipeline). CRF outperformed the other classification algorithms, with the LSTM networks showing modest improvement, even after obtaining a performance boost from domain-trained word embeddings. This result is of note because training a CRF is far less resource-intensive than training deep learning models, hence given a similar if not better performance, traditional methods may still be preferable in order to lower resource consumption. 
Keywords: Dialogue, Corpus, Statistical and Machine Learning Methods

Contextualized actionable requests segment relevant user utterances in each turn of the conversation.

Augmenting Small Data to Classify Contextualized Dialogue Acts for Exploratory Visualization

As the human body skeleton can be represented as a sparse graph, it is natural to exploit graph convolutional networks (GCNs) to model the articulated body structure for 3D human pose estimation (HPE). However, a vanilla graph convolutional layer, the building block of a GCN, only models the local relationships between each body joint and their neighbors on the skeleton graph. Some global attributes, e.g., the action of the person, can be critical to 3D HPE, especially in the case of occlusion or depth ambiguity. To address this issue, this paper introduces a new 3D HPE framework by learning global pose features in GCNs. Specifically, we add a global node to the graph and connect it to all the body joint nodes. On one hand, global features are updated by aggregating all body joint features to model the global attributes. On the other hand, the feature update of each body joint depends on not only their neighbors but also the global node. Furthermore, we propose a heterogeneous multi-task learning framework to learn the local and global features. While each local node regresses the 3D coordinate of the corresponding body joint, we force the global node to classify an action category or learn a low-dimensional pose embedding. Experimental results demonstrate the effectiveness of the proposed approach. 
Funding: The COMPaaS DLV project (NSF award CNS-1828265)

Illustration of a graph convolutional network (GCN) with a global nodeufor 3D human pose estimation

Learning Global Pose Features in Graph Convolutional Networks for 3D Human Pose Estimation

Dependency relations among visual entities are ubiquitous because both objects and scenes are highly structured. They provide prior knowledge about the real world that can help improve the generalization ability of deep learning approaches. Different from contextual reasoning which focuses on feature aggregation in the spatial domain, visual dependency reasoning explicitly models the dependency relations among visual entities. In this paper, we introduce a novel network architecture, termed the dependency network or DependencyNet, for semantic segmentation. It unifies dependency reasoning at three semantic levels. Intra-class reasoning decouples the representations of different object categories and updates them separately based on the internal object structures. Inter-class reasoning then performs spatial and semantic reasoning based on the dependency relations among different object categories. We will have an in-depth investigation on how to discover the dependency graph from the training annotations. Global dependency reasoning further refines the representations of each object category based on the global scene information. Extensive ablative studies with a controlled model size and the same network depth show that each individual dependency reasoning component benefits semantic segmentation and they together significantly improve the base network. Experimental results on two benchmark datasets show the DependencyNet achieves comparable performance to the recent states of the art. 
Funding: The COMPaaS DLV project (NSF award CNS-1828265)

Visual dependence relations are ubiquity since both ob-jects and scenes are highly structured. They provide prior knowl-edge about the real world that can be used to improve the gener-alization ability of a learning model.

Exploit Visual Dependency Relations for Semantic Segmentation

PhD Student

Kazi Shahrukh Omar

Construction activities for the new $117.8 million, 135,000-square-foot Computer Design Research and Learning Center to be located at 900 W. Taylor St. is underway. The CDRLC will support the Department of Computer Science&rsquo;s growing student enrollment with 16,000 square feet of classroom space for collaborative teaching and learning environments, as well as 35 labs for faculty research activities in the computer science and engineering disciplines. 
The CDRLC will be the new home for the Electronic Visualization Laboratory featuring the latest technological and computing innovations to support research in advanced visualization, networking and collaborative environments. 
Read more about the CDRLC groundbreaking in the <a href="https://today.uic.edu/uic-breaks-ground-on-new-computer-design-research-and-learning-center">July 15th issue of UIC Today.</a> 
See a live webcam of the building&rsquo;s progress via the <a href="https://webcampub.multivista.com/index.cfm?fuseaction=aPublicWebcam.embed&WebcamPublicEmbedUID=5D4DCF05-AB2B-4770-A090-D79BE467F958">Multivista Construction Documentation webcam.</a>

New Computer Design Research and Learning Center Breaks Ground

The latest developments of machine learning (ML) and deep learning (DL) algorithms have paved the way to effectively analyze the atomic structure of chemically-complex materials. In this work, we present a DL model built upon a fully convolutional neural network (FCN) to resolve the random elements distribution of complex PtNiPdCoFe high-entropy alloys (HEAs) represented in the scanning transmission electron microscopy (STEM) images at atomic resolution. The objective of the proposed neural network is to learn through semantic seg-mentation the non-linear correlations between the pixels&rsquo; intensities of STEM images and the number of atoms of different constituent elements in the atomic columns (i.e., column heights) in the HEA&rsquo;s structure. We demonstrate that our DL model is capable of correctly estimating the column heights or with an error up to 1 atom for the majority of the columns in the HEA structures represented in the simulated STEM images used to train and test the network. This establishes a sufficiently high level of confidence in the estimation of the element distribution in experimental images. The predicted distributions in different STEM images of nanoparticles reveal inhomogeneous fluctuations with local aggregations in the elemental atomic fractions within the columns. The most pronounced aggregation is displayed by Pt, which is the largest and most electronegative element in the synthesized HEA material. The proposed DL method is beneficial for an in-depth characterization of the structural properties of HEAs and multielement 3D materials in general. 
Keywords: Deep learning, High entropy alloys, STEM images, Column heights, Elements distribution

Example of atomic model of HEA characterized by a nearly equiatomic composition of Pt, Ni, Pd, Co and Fe.

Deep learning for mapping element distribution of high-entropy alloys in scanning transmission electron microscopy images

For the last several years EVL faculty member and CS associate professor Liz Marai has been collaborating with the University of Texas M.D. Anderson Cancer Center and the University of Iowa to develop tools to help oncologists make better assessments for treating cancer patients. More recently, Liz Marai and co-investigator Xinhua Zhang also from UIC have been developing a scalable artificial intelligence system based on real patient data to serve as a decision support tool for physicians treating head and neck cancer patients. This research is funded by the National Institutes of Health, National Cancer Institute Award #3R01CA258827-01S1, entitled &ldquo;Longitudinal Spatial-Nonspatial Decision Support for Competing Outcomes in Head and Neck Cancer Therapy&rdquo; with principal investigators from UIC, the University of Texas, and the University of Iowa. 
See the full article in the <a href="https://engineering.uic.edu/about/coe-news/rise/fall-2021/better-ai-for-better-decisions/">College of Engineering&rsquo;s CEO News.</a>

EVL Artificial Intelligence Research Featured in COE News

The National Institutes of Health: National Cancer Institute has awarded Liz Marai and collaborators at the University of Texas M.D. Anderson Cancer Center a Research Supplement to Promote Diversity in Health-Related Research Programs, for federal award R01CA258827, &ldquo;Longitudinal Spatial-Nonspatial Decision Support for Competing Outcomes in Head and Neck Cancer Therapy&rdquo; to run 2021 -2024. The award will support the interdisciplinary training of Dr. Rachel Glenn, an individual from a disadvantaged background (as defined by PA-21-071), under the direction of PIs Marai (UIC) and Fuller (U of Texas).

G. Elisabeta Marai Awarded NIH Diversity Supplement

In recent years deep meta-reinforcement learning has extended the applicability of reinforcement learning (RL) algorithms: by integrating recurrent networks, trained models have the ability to quickly adapt to new unseen environments without the need for further backpropagation. These models, however, cannot adapt without having information on past rewards, and are therefore not directly applicable to a sequential decision-making setting in which multiple steps are required before observing the final reward. 
One of the main applications affected by this limitation are dynamic treatment regimes, i.e. the problem of selecting the optimal medical treatment sequence for a patient at each step, keeping into account the complete past treatment history. By expanding deep meta-reinforcement learning to handle sequential decisions, a model would be able to prescribe the optimal treatment for each patient even if the patient&rsquo;s (or physician&rsquo;s) preferences on the outcome were never encountered by the model in training. 
We propose a recursive deep meta-reinforcement learning approach which enables the model of each decision of the sequential process to learn from and adapt to unseen circumstances by recursively integrating the feedback of the models of other decisions in the process. We evaluate our approach on synthetic two-step processes with fixed transition probabilities but varying reward functions, to test the models' ability to propagate environment information from the final reward to intermediate steps. Finally, we train our model on a dataset of three-step chemo-radiotherapeutic and surgical treatment of oropharyngeal squamous cell carcinoma patients, proving our approach&rsquo;s ability to optimally handle previously unseen patient's preferences on survival and toxicity outcomes.

Main features, outcomes, and decisions of the OPC treatment sequence

Recursive meta-Reinforcement Learning for Personalized Sequential Dynamic Treatment Policies

Former EVL student and graduate research assistant Elisa Tardini is recognized as a recipient of the Outstanding Thesis/Dissertation Award from the Graduate College at the University of Illinois at Chicago for her thesis entitled &ldquo;Recursive deep meta-reinforcement learning for personalized sequential dynamic treatment policies.&rdquo; This research focuses on a recursive deep meta-reinforcement learning approach to develop a model to prescribe an optimal treatment for patients.

2021 Outstanding Thesis/Dissertation Award to Elisa Tardini

Pose guided person image generation means to generate a photo-realistic person image conditioned on an input person image and a desired pose. This task requires spatial manipulation of the source image according to the target pose. However, convolutional neural networks (CNNs) are inherently limited to geometric transformations due to the fixed geometric structures in their building modules, i.e., convolution, pooling and unpooling, which cannot handle large motion and occlusions caused by large pose transform. This paper introduces a novel two-stream context-aware appearance transfer network to address these challenges. It is a three-stage architecture consisting of a source stream and a target stream. Each stage features an appearance transfer module, a multi-scale context module and two-stream feature fusion modules. The appearance transfer module handles large motion by finding the dense correspondence between the two-stream feature maps and then transferring the appearance information from the source stream to the target stream. The multi-scale context module handles occlusion via contextual modeling, which is achieved by atrous convolutions of different sampling rates. Both quantitative and qualitative results indicate the proposed network can effectively handle challenging cases of large pose transform while retaining the appearance details. Compared with state-of-the-art approaches, it achieves comparable or superior performance using much fewer parameters while being significantly faster. 
Keywords: Person image generation, Appearance transfer, Multi-scale context, Efficient image generation 
Relevant Funding: The COMPaaS DLV project (NSF award CNS-1828265) 
<a href="https://doi.org/10.1016/j.patcog.2021.108451">https://doi.org/10.1016/j.patcog.2021.108451</a>

Exploiting appearance transfer and multi-scale context for efficient person image generation

Figures within biomedical articles present essential evidence of the relevance of a publication in a curation workflow. In particular, visual cues of the image modality or experimental methods can help expert curators identify relevant papers from an increasing number of publications. Automating the identification of these content-bearing images can thus be helpful in computer-assisted curation. However, the paucity of labeled datasets and the specialized training required to label such images hinder the development of such tools. To address this problem, we present the design of ANIMO, a labeling system that integrates extraction and segmentation tools to ease the annotation burden. We first introduce two taxonomies of image modalities and experimental methods, derived in collaboration with curators. On the back-end of the system, we process batches of documents and create a labeling task per document. At the front-end, expert curators can access these tasks through a web interface and access the article of interest. We describe the evaluation of this system by a group of biocurators, and the human factor lessons learned from this interdisciplinary experience. 
doi: 10.1109/BIBM52615.2021.9669898

Front-end interface for labeling a document showing a taxonomy adapted for the CORD-19 dataset.

ANIMO: Annotation of Biomed Image Modalities

While designing sustainable and resilient urban built environment is increasingly promoted around the world, significant data gaps have made research on pressing sustainability issues challenging to carry out. Pavements are known to have strong economic and environmental impacts; however, most cities lack a spatial catalog of their surfaces due to the cost-prohibitive and time-consuming nature of data collection. Recent advancements in computer vision, together with the availability of street-level images, provide new opportunities for cities to extract large-scale built environment data with lower implementation costs and higher accuracy. In this paper, we propose CitySurfaces, an active learning-based framework that leverages computer vision techniques for classifying sidewalk materials using widely available street-level images. We trained the framework on images from New York City and Boston and the evaluation results show a 90.5% mIoU score. Furthermore, we evaluated the framework using images from six different cities, demonstrating that it can be applied to regions with distinct urban fabrics, even outside the domain of the training data. CitySurfaces can provide researchers and city agencies with a low-cost, accurate, and extensible method to collect sidewalk material data which plays a critical role in addressing major sustainability issues, including climate change and surface water management. 
Keywords: Sustainable built environment; Surface materials; Urban heat island; Semantic segmentation; Sidewalk assessment; Urban analytics; Computer vision 
ISSN 2210-6707 
<a href="https://www.sciencedirect.com/science/article/pii/S2210670721008933">https://www.sciencedirect.com/science/article/pii/S2210670721008933</a>

Predictions of the model on the held-out test set. Fine details and boundaries of objects like poles, plants, wooden sticks, and fire hydrants are very precisely predicted.

CitySurfaces: City-scale semantic segmentation of sidewalk materials

Motivation: Biomedical research findings are typically disseminated through publications. To simplify access to domain-specific knowledge while supporting the research community, several biomedical databases devote significant effort to manual curation of the literature - a labor intensive process. The first step toward biocuration requires identifying articles relevant to the specific area on which the database focuses. Thus, automatically identifying publications relevant to a specific topic within a large volume of publications is an important task toward expediting the biocuration process and, in turn, biomedical research. Current methods focus on textual contents, typically extracted from the title-and-abstract. Notably, images and captions are often used in publications to convey pivotal evidence about processes, experiments and results. 
Results: We present a new document classification scheme, using both image and caption information, in addition to titles-and-abstracts. To use the image information, we introduce a new image representation, namely Figure-word, based on class labels of subfigures. We use word embeddings for representing captions and titles-and-abstracts. To utilize all three types of information, we introduce two information integration methods. The first combines Figure-words and textual features obtained from captions and titles-and-abstracts into a single larger vector for document representation; the second employs a meta-classification scheme. Our experiments and results demonstrate the usefulness of the newly proposed Figure-words for representing images. Moreover, the results showcase the value of Figure-words, captions and titles-and-abstracts in providing complementary information for document classification; these three sources of information when combined, lead to an overall improved classification performance. 
Availability and implementation: Source code and the list of PMIDs of the publications in our datasets are available upon request. 
<a href="
https://doi.org/10.1093/bioinformatics/btab331">https://doi.org/10.1093/bioinformatics/btab331</a>

The process for converting figures into their corresponding Figure-words. The set of figures extracted from the biomedical document (left). The corresponding Figure-word with their vector representations indicating the types of comprising panel (right).

Utilizing Image and Caption Information for Biomedical Document Classification

Each year, over 60,000 people in the United States are diagnosed with head and neck cancers. Treatment strategies are dynamic and complex and can involve multiple courses of therapy with different potential outcomes. With head and neck cancer numbers increasing every year, partly due to viral infections, clinicians aim to develop customized treatment strategies tailored to each individual patient, under a healthcare model termed &ldquo;precision medicine&rdquo;. Precision medicine proposes the customization of healthcare, with medical decisions, practices, and products being tailored to the individual patient. The customization is based on individual factors collected from cohorts of patients who are similar to the patient under consideration. These factors may include disease markers, treatment options, demographics, and genetic, lifestyle, environmental, laboratory, or quality of life data. These data are heterogeneous, large-scale, and locally sparse. Making similarity-based precise recommendations for a specific patient requires novel approaches which blend biomedicine with complex quantitative methodology. We describe a first application of deep-Q-learning (DQL) and simulation as a digital twin dyad to optimize multi-step treatment of head and neck cancer patients, and to predict patient survival and toxicity outcomes. The treatment decision DQL digital twin and the patient’s digital twin were created, trained and evaluated using Google Colab for the HPC solution on a dataset of 536 oropharyngeal squamous cell carcinoma (OPC) patients with the goal of, respectively, determining the optimal treatment decisions with respect to survival and toxicity metrics, and predicting the outcomes of the optimal treatment on the patient. The models were trained on a subset of 402 patients (split randomly) and evaluated on a separate set of 134 patients. The main outcomes and measures include accuracy in predicting the final outcomes of a treatment sequence, and improvement in predicted outcomes rates of the DQL when compared to the observed physician outcomes. On the validation set, we obtained 87.09% mean and 90.85%
median accuracy in treatment outcome prediction. Given the prediction accuracy and predicted improvement on medically relevant outcomes yielded by this approach, this digital twin dyad of the patient-physician dynamic treatment problem has the potential of aiding physicians in determining the optimal course of treatment and in assessing its outcomes. 
Keywords: digital twin dyad, reinforcement learning, head and neck cancer

Model performance for the combined outcome (OS+DP) models without (left) and with radiomics (right).

A digital twin dyad with deep Q-learning in head and neck cancer treatment

Support vector regression (SVR) is particularly beneficial when the outcome and predictors are nonlinearly related. However, when many covariates are available, the method&rsquo;s flexibility can lead to overfitting and an overall loss in predictive accuracy. To overcome this drawback, we develop a feature selection method for SVR based on a genetic algorithm that iteratively searches across potential subsets of covariates to find those that yield the best performance according to a user-defined fitness function. We evaluate the performance of our feature selection method for SVR, comparing it to alternate methods including LASSO and random forest, in a simulation study. We find that our method yields higher predictive accuracy than SVR without feature selection. Our method outperforms LASSO when the relationship between covariates and outcome is nonlinear. Random forest performs equivalently to our method in some scenarios, but more poorly when covariates are correlated. We apply our method to predict donor kidney function 1 year after transplant using data from the United Network for Organ Sharing national registry.

Results of simulation 1 setting with 300 total covariates X, 15 of them truly associated with the outcome Y.

Feature Selection for Support Vector Regression Using a Genetic Algorithm

Candidate: Juan Trelles Trabucco 
Date: Friday, April 29, 2022 
Time: 11 am CT 
Committee Members: Dr. G. Elisabeta Marai, Dr. Andy Johnson, Dr. Wei Tang, Dr. Steven Drucker, Dr. Cecilia Arighi 
Abstract: 
Biomedical researchers need to search over an increasing number of publications to find knowledge. Among these researchers, biocurators specialize in identifying relevant publications and extracting information to populate structured databases. Efforts from biocuration groups benefit a much larger community who access this curated information daily. To effectively deal with the extensive collection of documents, the biocuration workflow uses text-mining approaches to automate tasks, such as predicting the relevance of a publication. While exploiting patterns in text data has yielded significant breakthroughs for biocuration, images in publications contain important cues to determine the document's relevance. Still, integrating these image content with textual features is underexplored due to difficulties collecting images, defining classification schemes, and having access to domain experts for data labeling. 
Rooted in a years-long collaboration with text-mining researchers and biocurators, I propose strategies for harvesting labels from biomedical publications and modeling image classifiers to support biocuration. First, I present two approaches for training image classifiers with deep learning for small and unbalanced datasets. Then, I characterize the labeling task in biocuration and introduce a labeling system tailored for domain experts. Next, I build on lessons from this labeling system to develop a labeling approach for model builders by combining visual analytics and machine learning techniques. These works provide the building blocks for the future design of search interfaces that leverage text- and image-based data for biomedical experts. In addition, it supports the development of novel multi-modal approaches for detecting publication's relevance and interpreting these predictions.

Biomedical Image Classification Application User-Interface

PhD Prelim Announcement: Integrating Image Features to Support the Biocuration Workflow

The COMPaaS DLV: Composable Platform as a Service Instrument for Deep Learning & Visualization user community consists of faculty from Physics and four College of Engineering departments (Computer Science, Civil, Materials & Environmental Engineering, Mechanical and Industrial Engineering, and Electrical and Computer Engineering) - primarily run applications that are GPU-centric for compute, with significant variability in storage and networking around their data requirements. 
Computer Science applications primarily focus on security, data science, computer vision and Machine Learning (ML). Security projects explore the complexity of modern web applications and the intricacies of security mechanisms that often result in flaws that expose users to significant security and privacy threats. These projects try to develop methods and tools that enable users to understand and more effectively manage retrospective privacy in the context of modern, long-lived, online archives. Composable resources were used to develop Natural Language Processing (NLP)-based domain-specific classifiers that identified data practices stated in privacy policies. Adherence of corresponding applications were then adjusted based on this ground truth. 
Data-science applications have an intuitive framework that integrates state-of-the-art AI technologies with applications, workflows, smart visualizations and collaboration services to help users access, share, explore and analyze their data, whether local or remote, come to conclusions, and make decisions with greater speed, accuracy, comprehensiveness and confidence. One such project is developing and advancing tools that identify image data in biomedical literature to locate beneficial, targeted publications. This work involves training image classifiers, integrating classifiers into labeling pipelines, designing retrieval user interfaces, and identifying related visual representations. 
Computer vision projects include semantic segregation and 3D human pose estimation. Researchers are developing a novel network architecture, termed DependencyNet (dependency network), for semantic segmentation. They also achieved experimental results that demonstrate an effective approach for 3D human pose estimation. Over the past two years, they found COMPaaS to be consistently stable and efficient, with result output as expected, and the group&rsquo;s models achieved state-of-the-art performance on their respective benchmarks. 
ML applications include frameworks for many complex real-world reinforcement learning problems, such as the coordination of autonomous vehicles, network packet delivery, and distributed logistics. 
Civil, Materials & Environmental Engineering applications focus on simulation and modeling. Researchers run data-driven models on high-performance computers to develop an accurate and general neural network ML model that uses crystallographic data to study patterns of synthesizability. They also perform simulations of mass transport in alloys and ceramics. COMPaaS has been performing 1.5-10 times faster than comparable infrastructures they are familiar with. Additionally, they found our use of Jupyter Notebooks to be a significant asset. 
Mechanical and Industrial Engineering researchers use COMPaaS for three research projects: feature extraction in fluid flow using a Convolutional Neural Network (CNN); column height detection in metallic nanoparticles using a CNN; and, electric vehicle battery state-of-charge estimation using different ML methods. 
Electrical and Computer Engineering researchers recently started running mathematical models of ML algorithms and training language models using long-short term memory (LSTMs). 
In 2020, we began implementing a GPUoE (GPU over Ethernet) prototype. Using a GPU expansion chassis connected to compute nodes over Ethernet, we were able to compose remote GPUs into our existing composable infrastructure. Our APIs developed for composable infrastructure have been extended to support these remote GPUs. 
In 2021, we enhanced COMPaaS with a public-facing JupyterHub server providing secure web-based access to experiments, a modern PCIe fabric and a supporting expansion chassis, thereby making composable co-location available to our users. As familiarity with composable infrastructure grows, the idea of using composable co-location instead of traditional servers co-located in racks is possible. By composable co-location, we mean the integration of emerging hardware needs without changing the existing core infrastructure. Researchers can now co-locate their own accelerators or compute nodes and leverage existing COMPaaS resources natively.

COMPaaS DLV User Community

In today&rsquo;s Big Data era, data scientists require modern workflows to quickly analyze large-scale datasets using complex codes to maintain the rate of scientific progress. These scientists often rely on available campus resources or off-the-shelf computational systems for their applications. Unified infrastructure or over-provisioned servers can quickly become
bottlenecks for specific tasks, wasting time and resources. Composable infrastructure helps solve these problems by providing users with new ways to increase resource utilization. Composable infrastructure disaggregates a computer&rsquo;s components – CPU, GPU (accelerators), storage and networking - into fluid pools of resources, but typically relies upon infrastructure engineers to
architect individual machines. Infrastructure is either managed with specialized command-line utilities, user interfaces or specification files. These management models are cumbersome and difficult to incorporate into data-science workflows. We
developed a high-level software API, Composastructure, which, when integrated into modern workflows, can be used by infrastructure engineers as well as data scientists to reorganize composable resources on demand. Composastructure enables infrastructures to be programmable, secure, persistent and reproducible. Our API composes machines, frees resources, supports multi-rack operations, and includes a Python module for Jupyter Notebooks. 
Keywords - distributed systems, testbed implementation and deployment, composable infrastructure, deep learning, visualization, infrastructure as code

Composastructure System Architecture

Moving from Composable to Programmable

Composable infrastructure holds the promise of accelerating the pace of academic research and discovery by enabling researchers to tailor the resources of a machine (e.g., GPUs, storage, NICs), on-demand, to address application needs. We were first introduced to composable infrastructure in 2018, and at the same time, there was growing demand among our College of Engineering faculty for GPU systems for data science, artificial intelligence / machine learning / deep learning, and visualization. Many purchased their own individual desktop or deskside systems, a few pursued more costly cloud and HPC solutions, and others looked to the College or campus computer center for GPU resources which, at the time, were scarce. After surveying the diverse needs of our faculty and studying product offerings by a few nascent startups in the composable infrastructure sector, we applied for and received a grant from the National Science Foundation in November 2019 to purchase a mid-scale system, configured to our specifications, for use by faculty and students for research and research training. 
This paper describes our composable infrastructure solution and implementation for our academic community. Given how modern workflows are progressively moving to containers and cloud frameworks (using Kubernetes) and to programming notebooks (primarily Jupyter), both for ease of use and for ensuring reproducible experiments, we initially adapted these tools for our system. We have since made it simpler to use our system, and now provide our users with a public facing JupyterHub server. We also added an expansion chassis to our system to enable composable co-location, which is a shared central architecture in which our researchers can insert and integrate specialized resources (GPUs, accelerators, networking cards, etc.) needed for their research. 
In February 2020, installation of our system was finalized and made operational and we began providing access to faculty in the College of Engineering. Now, two years later, it is used by over 40 faculty and students plus some external collaborators for research and research training. Their use cases and experiences are briefly described in this paper. Composable infrastructure has proven to be a useful computational system for workload variability, uneven applications, and modern workflows in academic environments. 
Keywords - composable infrastructure, deep learning, visualization, resource management, workload management, user workflow, composable co-location, infrastructure as code

COMPaaS Hardware Racks

Composable Infrastructures for an Academic Research Environment: Lessons Learned

EVL is a technology Innovator, building industry and cross-institutional connections and partnerships. In 2022, EVL deployed two GPU FIONAs in partnership with the Pacific Research Platform (PRP) and the University of California San Diego. The PRP is a partnership of more than 50 institutions that establishes a science-driven high-capacity data-centric &lsquo;freeway system&rsquo;. PRP has support from the National Science Foundation, and includes the US Department of Energy&rsquo;s Energy Science Network (ESnet), as well as scores of research universities in the nation and around the world. PRP leverages Nautilus, a HyperCluster for running containerized Big Data Applications using Kubernetes, and FIONA, flash input/output network applications. A GPU FIONA is a rack-mounted PCs designed by the PRP team that is optimized for 100Gbps large data storage with Mellanox ConnectX-6 SmartNICs, a 24 core Epic Rome CPU, 256GB of memory and 8 NVIDIA RTX-3090 GPUs. Nautilus provides over 7500 CPU cores, over 500 GPUs and 4 PBs of storage. The addition of two Nautilus GPU FIONAs at EVL builds upon our 2018 DTN (Data Transfer Node) that has provided over 300GB of Ceph Storage to Nautilus and our CAVE2 GPU Nodes that are contributing 5 NVIDIA GPUS. EVL provides Nautilus with 300GB of Ceph Storage, 128 CPU cores, and 21 NVIDIA GPUs. The platform is utilized for a wide variety of applications, including applying advanced astronomy AI to microscopy workflow, simulations for natural disasters, climate modeling, virtual reality and even accelerating NSF&rsquo;s IceCube Photon Propagation. UIC is enabled to participate in these cross-institutional connections using UIC&rsquo;s HPRNet, a network environment tailored to the needs of high-performance science applications, built on a Science DMZ architecture. 
<a href="https://pacificresearchplatform.org/media/">Read more about PRP.</a> 
<a href="https://pacificresearchplatform.org/media/sc21-larry-smarr-on-the-rise-of-supernetwork-data-intensive-computing/">See media coverage from SC21 on PRP and Nautilus.</a>

FIONA

Pacific Research Platform (PRP) - Nautilus

In today&rsquo;s Big Data era, data scientists require modern workflows to quickly analyze large-scale datasets using complex codes to maintain the rate of scientific progress. These scientists often rely on available campus resources or off-the-shelf
computational systems for their applications. Unified infrastructure or over-provisioned servers can quickly become bottlenecks for specific tasks, wasting time and resources. Composable infrastructure helps solve these problems by providing users with new ways to increase resource utilization. Composable infrastructure disaggregates a computer&rsquo;s components - CPU, GPU (accelerators), storage and networking - into fluid pools of resources, but typically relies upon infrastructure engineers to architect individual machines. Infrastructure is either managed with specialized command-line utilities, user interfaces or specification files. These management models are cumbersome and difficult to incorporate into data-science workflows. We developed a high-level software API, Composastructure, which, when integrated into modern workflows, can be used by infrastructure engineers as well as data scientists to reorganize composable resources on demand. Composastructure enables infrastructures to be programmable, secure, persistent and reproducible. Our API composes machines, frees resources, supports multi-rack operations, and includes a Python module for Jupyter Notebooks. 
Keywords: distributed systems, testbed implementation and deployment, composable infrastructure, deep learning, visualization, infrastructure as code

Composable infrastructure holds the promise of accelerating the pace of academic research and discovery by enabling researchers to tailor the resources of a machine (e.g., GPUs, storage, NICs), on-demand, to address application needs. We were first introduced to composable infrastructure in 2018, and at the same time, there was growing demand among our College of Engineering faculty for GPU systems for data science, artificial intelligence / machine learning / deep learning, and visualization. Many purchased their own individual desktop or deskside systems, a few pursued more costly cloud and HPC solutions, and others looked to the College or campus computer center for GPU resources which, at the time, were scarce. After surveying the diverse needs of our faculty and studying product offerings by a few nascent startups in the composable infrastructure sector, we applied for and received a grant from the National Science Foundation in November 2019 to purchase a mid-scale system, configured to our specifications, for use by faculty and students for research and research training. 
This paper describes our composable infrastructure solution and implementation for our academic community. Given how modern workflows are progressively moving to containers and cloud frameworks (using Kubernetes) and to programming notebooks (primarily Jupyter), both for ease of use and for ensuring reproducible experiments, we initially adapted these tools for our system. We have since made it simpler to use our system, and now provide our users with a public facing JupyterHub server. We also added an expansion chassis to our system to enable composable co-location, which is a shared central architecture in which our researchers can insert and integrate specialized resources (GPUs, accelerators, networking cards, etc.) needed for their research. 
In February 2020, installation of our system was finalized and made operational and we began providing access to faculty in the College of Engineering. Now, two years later, it is used by over 40 faculty and students plus some external collaborators for research and research training. Their use cases and experiences are briefly described in this paper. Composable infrastructure has proven to be a useful computational system for workload variability, uneven applications, and modern workflows in academic environments. 
Keywords: composable infrastructure, deep learning, visualization, resource management, workload management, user workflow, composable co-location, infrastructure as code

This academic year 2023 EVL alum Mike Papka has joined UIC&rsquo;s Department of Computer Science and EVL faculty as a professor, bringing years of expertise in high performance computing to campus. For nearly three decades, Papka has held the position of Research Scientist at Argonne National Laboratory contributing to research in high performance computing, scientific visualization of large and complex datasets, information visualization related to high-performance computing, virtual reality for science and the integration of edge services. 
Since 2006, he’s served as Deputy Associate Laboratory Director, Computing, Environment and Life Sciences (CELS) Directorate, providing strategic planning and management of the Laboratory Directorate comprising Biosciences, Computational Science, Data Science and Learning, Environmental Sciences, Mathematics and Computer Science, a joint institute with Northwestern University, and the Argonne Leadership Computing Facility. In his role, Papka aids in executing a research program that aims to couple computing-related activities with several science domains whose future is tied to progress in computing.
 For the last 12 years, Papka has served and continues to serve as Division Director at Argonne Leadership Computing Facility (ALCF). He is responsible for planning and executing facility upgrades and integrated data strategies, risk assessment and management, operations reporting and reviews, staff promotion and retention, and safety. The ALCF is home to some of the world&rsquo;s fastest supercomputers: Theta; a Cray XC40 supercomputer based on Intel’s Xeon Phi KNL processor and NVIDIA DGX-3, Polaris; a HPE Apollo 6500 Gen10+ based on AMD Milan processor and NVIDIA A100s, and Aurora; a Intel CPU/GPU system currently being installed that support scientific and engineering open science research. 
We welcome Mike&rsquo;s return to UIC and EVL! He undoubtedly will be a valued asset to our research and educational enterprise.

Alum Michael E. Papka Joins EVL Faculty

The SAGE3 Team has been quiet far too long, but we&rsquo;ve been very busy! Here&rsquo;s what to look forward to in the coming months: 
* SAGE3 beta version software with documentation will be released the week of the IEEE/ACM Supercomputing (SC22) conference, to be held in Dallas, Texas, USA, November 13-18, 2022. More information about the release will be forthcoming. For information on SC22, see: <a href="https://sc22.supercomputing.org/">https://sc22.supercomputing.org/</a> 
* A 1-hour SAGE3 Birds of a Feather (BoF) session and ongoing demonstrations and presentations will take place during SC22. At the BoF (Tuesday, Nov 15 from 12:15-13:15 in Room D168), the SAGE3 team will introduce the latest software release. The team will also share a booth in the research exhibition (StarLight booth #2847 as well as partner booths) and will conduct demos and presentations. More information will be forthcoming. 
* Educational SAGE3 webinars will be held, starting next year. Team members will present SAGE3&rsquo;s features and capabilities, provide use cases, answer questions, and solicit user feedback and use cases. More information will be forthcoming. 
Have you tried the SAGE3 alpha version for Linux, Windows and MacOS, which was released in December 2021? If not, go to the SAGE3 website, download the software and display client, and watch the &ldquo;How To&rdquo; video. 
<a href="https://sage3.sagecommons.org/">https://sage3.sagecommons.org/</a> 
REMINDER: SAGE3 / SAGE2 COMMUNITY USER FORUMS are now hosted on Slack. To join, go to the SAGE3 website and click on the Support menu: 
<a href="https://sage3.sagecommons.org/?page_id=60">https://sage3.sagecommons.org/?page_id=60</a> 
For SAGE3, subscribe to the sage3-support channel. 
For SAGE2, subscribe to the sage2-support channel. 
Thank you for your longstanding support of SAGE3 and we look forward to continued collaborations. 
Maxine Brown 
SAGE team member 
<a href="mailto:maxine@uic.edu">maxine@uic.edu</a> 
SAGE3 is being developed by the Laboratory for Advanced Visualization and Applications (LAVA) and the Hawaii Data Science Institute (HI-DSI) at University of Hawai&rsquo;i at Manoa (UHM), the Electronic Visualization Laboratory (EVL) at University of Illinois Chicago (UIC), and the InfoVis Lab at Virginia Tech (VT). SAGE3 receives major funding from the U.S. National Science Foundation (NSF), awards #2004014 (UHM), #2003800 (UIC) and #2003387 (VT). 
SAGE2 received major NSF support, award #ACI-1441963, to UHM and UIC. 
SAGE, SAGE2 and SAGE3 are trademarks of the University of Illinois Board of Trustees.

Announcing SAGE3 Upcoming Events !!

The Electronic Visualization Laboratory at the University of Illinois Chicago acquired a 24 compute node, 64 GPU composable infrastructure compute cluster, COMPaaS DLV - Composable Platform as a Service: Instrument for Deep Learning & Visualization in 2019.  Since then 40+ users have designed and run computations on COMPaaS.  To track cluster utilization, evaluate hardware
performance, and simply analyze application resource utilization of this experimental composable infrastructure system we set out to
collect metrics on the hardware and software infrastructure.  Users deploy their applications on the cluster using Kubernetes, a container orchestration platform that enables users to deploy applications on the cluster without the need to have complete knowledge of the underlying hardware and with minimal assistance from system administrators.  Following this, we utilized tooling already adapted for Kubernetes-based clusters to collect hardware and software metrics.  The metrics provided by this tooling is useful for system administrators in evaluating cluster performance and utilization, as well as useful for users in determining their application&rsquo;s performance and discovering bottlenecks.

Deep Learning

Bibliography

pubs (22)

Related Topics

people (4)

events (14)

research (5)

Contact