MATCH Engagements

I am new to ACCESS. I have a little bit of past experience running code on NCSA's Blue Waters. As a self-taught programmer, it would be interesting to learn from an experienced mentor. 

Here's an overview of my project:

Anxiety detection is topic that is actively studied but struggles to generalize and perform outside of controlled lab environments. I propose to critically analyze state of the art detection methods to quantitatively quantify failure modes of existing applied machine learning models and introduce methods to robustify real-world challenges. The aim is to start the study by performing sensitivity analysis of existing best-performing models, then testing existing hypothesis of real-world failure of these models. We predict that this will lead us to understand more deeply why models fail and use explainability to design better in-lab experimental protocols and machine learning models that can perform better in real-world scenarios. Findings will dictate future directions that may include improving personalized health detection, careful design of experimental protocols that empower transfer learning to expand on existing reach of anxiety detection models, use explainability techniques to inform better sensing methods and hardware, and other interesting future directions.

Sea levels are rising (3.7 mm/yr and increasing!)! The primary contributor to rising sea levels is enhanced polar ice discharge due to climate change. However, their dynamic response to climate change remains a fundamental uncertainty in future projections. Computational cost limits the simulation time on which models can run to narrow the uncertainty in future sea level rise predictions.  The project's overarching goal is to leverage GPU hardware capabilities to significantly alleviate the computational cost and narrow the uncertainty in future sea level rise predictions. 

[Plan A] The PI is investigating numerical techniques to predict ice-sheet flow for large-to-continental glaciers on GPUs. Suppose a successful technique has been identified, implemented in MATLAB, and verified prior to the start of this project. In that case, the objective will be to port the code to CUDA C and investigate techniques to justify the GPU implementation in the price and power consumption to performance metric compared to a "standard" CPU implementation.

[Plan B] The PI developed a preliminary GPU implementation to predict ice-sheet flow for regional-scale glaciers on GPUs. The GPU ice velocity predictions agreed (~ 1% discrepancy) with a "standard" CPU implementation for the chosen glacier model configurations and input data. The GPU implementation was justified in the price and power consumption to performance metrics. However, the profiling results from the preliminary GPU implementation indicated non-optimal global memory access patterns reported in the LITEX and L2 cache. Furthermore, we observed significant drops in effective memory throughput with increased spatial resolution or degrees of freedom (DoFs). This project will aim to investigate techniques to reduce mesh non-localities and drop in effective memory throughput with an increase in spatial resolution. 

My ongoing project is focused on using species trait value (as data matrices) and its corresponding phylogenetic relationship (as a distance matrix) to reconstruct the evolutionary history of the smoke-induced seed germination trait. The results of this project are expected to increase the predictability of which untested species could benefit from smoke treatment, which could promote germination success of native species in ecological restoration. This computational resources allocated for this project pull from the high-memory partition of our Ivy cluster of HPCC (Centos 8, Slurm 20.11, 1.5 TB memory/node, 20 core /node, 4 node). However, given that I have over 1300 species to analyze, using the maximum amount of resources to speed up the data analysis is a challenge for two reasons: (1) the ancestral state reconstruction (the evolutionary history of plant traits) needs to use the Markov Chain Monte Carlo (MCMC) in Bayesian statistics, which runs more than 10 million steps and, according to experienced evolutionary biologists, could take a traditional single core simulation up 6 months to run; and (2) my data contain over 1300 native species, with about 500 polymorphic points (phylogenetic uncertainty), which would need a large scale of random simulation to give statistical strength. For instance, if I use 100 simulations for each 500 uncertainty points, I would have 50,000 simulated trees. Based on my previous experience with simulations, I could design codes to parallel analyze 50,000 simulated trees but even with this parallelization the long run MCMC will still require 50000 cores to run for up to 6 months. Given this computational and evolutionary research challenge, my current work is focused on discovering a suitable parallelization methods for the MCMC steps. I hope to have some computational experts to discuss my project.

In this project, we plan to apply machine learning methods such as the Transformer to company fundamentals to predict their future stock returns or future earnings. 

Our data is monthly from 1965 to 2022 and the size of the data is about 750M.

Our dependent variables are stock return or earnings in the next periods. Our independent variables are fundamental. We will implement rolling methods. For example, at the end of year t, we use the past five years of data, with 4 years for training and 1 year for validation. With the selected model, we test return or earnings predictions using data in year t+1. Then, we repeat this process for years t+1, t+2, and so on. We will need computing powers.

I am still in the process of figuring out how to apply transformer in such a setting.

The project involves the improvements of the alphafold program. The projects includes two parts, improvements of the program and possible applications of it in collaboration with different researchers. The effort will be in coordination of research groups interested in the application of alphafold. 

Specifically the performance improvements involves a hierarchical/diverse aspects of the code, (1)I/O, (2) databases query, (3) multithread/multi process in python, as well as GPU acceleration, and more.   

We have requested educational allocation (EVE220001) to support a new course developed by the Department of Bioengineering at UIUC, BIOE486 Applied Deep Learning for Biomedical Imaging. This course as part of our new Master of Science in Biomedical Image Computing (MS-BIC) program,  covers basic concepts, methodology and algorithms in deep learning and their applications to solve modern biomedical imaging challenges. To achieve the educational goals, this course requires assess to high-performance computing for the students, especially GPUs, to study, train and improve various deep neural network based models. The purpose of this request is to seek opportunities to build a consistent and more accessible computing and data management platform for our course and even for the MS-BIC program in general.

While students who used or are using this resource have provided good feedback, they have found that dealing with the process of getting access to the GPU computing resource, setting up the necessary environment, and installing all the needed software can be taking a lot of efforts and distracting them from the core computational and deep learning problems. Moving forward, we hope to work with the MATCH program to create some support for our students. 

Specifically, if possible, we would like to offer the amazing computing resources in ACCESS, e.g., the Expanse GPU cluster to all of our MS-BIC students. To that end, we would like to have a consistent way of setting up the environment(s) for them to develop deep learning models and algorithms in Tensorflow or Pytorch. A training module on how to efficiently work with interactive programming mode for writing codes to build and train deep neural networks, e.g., using Jupyter notebook or VScode to write TF or Pytorch codes on the cluster would be highly appreciated. Lastly, the availability o student support through the MATCH program to help students resolve issues when needed will be very helpful.

Ideally, we would like to have some technical support for the remaining of the semester (particularly on using Tensorflow/Pytorch on the Expanse cluster) as the students start to use the Expanse cluster more frequently with the HW and project getting more complicated. In terms of consistent environment set up and training materials, we do not have an urgent deadline. 

GeoACT (GEOspatial Agent-based model for Covid Transmission) is a designed to simulate a range of intervention scenarios to help schools evaluate their COVID-19 plans to prevent super-spreader events and outbreaks. It consists of several modules, which compute infection risks in classrooms and on school buses, given specific classroom layouts, student population, and school activities. The first version of the model was deployed on the Expanse (and earlier, COMET) resource at SDSC and accessed via the Apache Airavata portal (geoact.org). The second version is a rewrite of the model which makes it easier to adjust to new strains, vaccines and boosters, and include detailed user-defined school schedules, school floor plans, and local community transmission rates. This version is nearing completion. We’ll use Expanse to run additional scenarios using the enhanced model and the newly added meta-analysis module. The current goal is to make the model more general so that it can be used for other health emergencies. GeoACT has been in the news, e.g. https://ucsdnews.ucsd.edu/feature/uc-san-diego-data-science-undergrads-help-keep-k-12-students-covid-safe, and https://www.hpcwire.com/2022/01/13/sdsc-supercomputers-helped-enable-safer-school-reopenings/  (HPCWire 2022 Editors' Choice Award)