About

Christopher M. Jermaine is the chair of the Rice University Department of Computer Science in the George R. Brown School of Engineering and holds the title of Victor E. Cameron Professor of Computer Science. His academic background includes a BA from the Mathematics Department at UCSD, an MSc from the Computer Science and Engineering Department at Ohio State University, where his advisor was Renee Miller, and a PhD from the College of Computing at Georgia Tech under the supervision of Ed Omiecinski. Jermaine has been a faculty member at Rice University since January 2009 and previously served on the faculty of the computer science department at the University of Florida from 2002 through August 2010. His research focuses on data analytics, specifically on how to analyze, store, retrieve, and manipulate large and heterogeneous data sets. His work emphasizes the systems-oriented problems that arise when building software to manage large and diverse data, as well as the challenges encountered when applying statistical methods to such data. Jermaine has received numerous awards, including a 2008 Alfred P. Sloan Foundation Research Fellowship, a National Science Foundation CAREER award, and several best paper awards from ACM SIGMOD, ACM SIGKDD, and ICDE. His teaching areas include applications of discrete structures, information and database systems, indexing, large databases, database system implementation, design and analysis of algorithms, and tools and models for data science.

Research topics

• Computer Science
• Data Mining
• Mathematics
• Machine Learning
• Database
• Statistics
• Engineering
• Algorithm
• Management science
• Data science
• Theoretical computer science

Selected publications

• MHA-RAG: Improving Efficiency, Accuracy, and Consistency by Encoding Exemplars as Soft Prompts

  ArXiv.org · 2025-10-06

  preprintOpen accessSenior author

  Adapting Foundation Models to new domains with limited training data is challenging and computationally expensive. While prior work has demonstrated the effectiveness of using domain-specific exemplars as in-context demonstrations, we investigate whether representing exemplars purely as text is the most efficient, effective, and stable approach. We explore an alternative: representing exemplars as soft prompts with an exemplar order invariant model architecture. To this end, we introduce Multi-Head Attention Retrieval-Augmented Generation (MHA-RAG), a framework with the number of attention heads serving as a simple hyperparameter to control soft prompt-generation across different tasks. Across multiple question-answering benchmarks and model scales, MHA-RAG achieves a 20-point performance gain over standard RAG, while cutting inference costs by a factor of 10X GFLOPs-delivering both higher accuracy and greater efficiency, invariant to exemplar order.

  PublisherOA PDFDOI

• Dataset for ADAPT: Lightweight, Long-Range Machine Learning Force Fields Without Graphs

  Zenodo (CERN European Organization for Nuclear Research) · 2025-10-11

  datasetOpen access

  This is the data used in the *ADAPT: Lightweight, Long-Range Machine Learning Force Fields Without Graphs* paper found at: https://arxiv.org/abs/2509.24115. It is a set of DFT calculations for point defects in silicon. The majority of the set is complex defects featuring two defect atoms, although vacancy and simple substitional defects are also present. Supplementary materials Section A in the original paper gives a description of how the data was generated and further details. Further details about this research can be found at: https://evandramko.github.io/ADAPT_webpage/ We release all the raw pickle data files. We also release the data combined into xyz files for convenience. We suggest using all files except 11.pickle as training, and using 11.pickle as the testing file. We also release a selection of 100 frames that were used to make the diagrams in the associated ADAPT paper. Following updates to the ADAPT library at https://github.com/EvanDramko/ADAPT_Released we have updated the representation of atoms from a (n,12) setup (using all available atomic descriptors in the raw data) to a more standard (n,4) representation of (x, y, z, Z) per atom. The original (n,12) data can still be extracted from the pickle files if you would prefer to use that.

  PublisherDOI

• TURNIP: A "Nondeterministic" GPU Runtime with CPU RAM Offload

  arXiv (Cornell University) · 2024-05-25

  preprintOpen access

  An obvious way to alleviate memory difficulties in GPU-based AI computing is via CPU offload, where data are moved between GPU and CPU RAM, so inexpensive CPU RAM is used to increase the amount of storage available. While CPU offload is an obvious idea, it can greatly slow down a computation, due to the relatively slow transfer rate between CPU RAM and GPU RAM. Thus, any system for CPU offload needs to ensure that when such a transfer needs to happen, no computation is blocked waiting for the transfer to finish. One of the key challenges when using CPU offload is that memory transfers introduce nondeterminacy into the system: it is not possible to know before runtime when the transfers will finish, and hence what is the best order of operations to run to ensure there is no blocking. In this paper, we describe TURNIP, which is a system for running AI computations using CPU offload. The key innovation in TURNIP is the compilation of the AI computation into a dependency graph that gives the TURNIP runtime freedom to run operations such as GPU kernel calls in many different orders; at runtime, TURNIP chooses the best order in response to real-time events.

  PublisherOA PDFDOI

• Online Cascade Learning for Efficient Inference over Streams

  arXiv (Cornell University) · 2024-02-07

  preprintOpen access

  Large Language Models (LLMs) have a natural role in answering complex queries about data streams, but the high computational cost of LLM inference makes them infeasible in many such tasks. We propose online cascade learning, the first approach to address this challenge. The objective here is to learn a "cascade" of models, starting with lower-capacity models (such as logistic regression) and ending with a powerful LLM, along with a deferral policy that determines the model to be used on a given input. We formulate the task of learning cascades online as an imitation-learning problem, where smaller models are updated over time imitating the collected LLM demonstrations, and give a no-regret algorithm for the problem. Experimental results across four benchmarks show that our method parallels LLMs in accuracy while cutting down inference costs by as much as 90% with strong robustness against input distribution shifts, underscoring its efficacy and adaptability in stream processing.

  PublisherOA PDFDOI

• Studying Early Decision Making with Progressive Bar Charts

  IEEE Transactions on Visualization and Computer Graphics · 2022 · 11 citations

  • Computer Science
  • Computer Science
  • Data Mining

  We conduct a user study to quantify and compare user performance for a value comparison task using four bar chart designs, where the bars show the mean values of data loaded progressively and updated every second (progressive bar charts). Progressive visualization divides different stages of the visualization pipeline-data loading, processing, and visualization-into iterative animated steps to limit the latency when loading large amounts of data. An animated visualization appearing quickly, unfolding, and getting more accurate with time, enables users to make early decisions. However, intermediate mean estimates are computed only on partial data and may not have time to converge to the true means, potentially misleading users and resulting in incorrect decisions. To address this issue, we propose two new designs visualizing the history of values in progressive bar charts, in addition to the use of confidence intervals. We comparatively study four progressive bar chart designs: with/without confidence intervals, and using near-history representation with/without confidence intervals, on three realistic data distributions. We evaluate user performance based on the percentage of correct answers (accuracy), response time, and user confidence. Our results show that, overall, users can make early and accurate decisions with 92% accuracy using only 18% of the data, regardless of the design. We find that our proposed bar chart design with only near-history is comparable to bar charts with only confidence intervals in performance, and the qualitative feedback we received indicates a preference for designs with history.

  PublisherOA PDFDOI

• ResIST: Layer-Wise Decomposition of ResNets for Distributed Training

  arXiv (Cornell University) · 2021-07-02 · 1 citations

  preprintOpen access

  We propose ResIST, a novel distributed training protocol for Residual Networks (ResNets). ResIST randomly decomposes a global ResNet into several shallow sub-ResNets that are trained independently in a distributed manner for several local iterations, before having their updates synchronized and aggregated into the global model. In the next round, new sub-ResNets are randomly generated and the process repeats until convergence. By construction, per iteration, ResIST communicates only a small portion of network parameters to each machine and never uses the full model during training. Thus, ResIST reduces the per-iteration communication, memory, and time requirements of ResNet training to only a fraction of the requirements of full-model training. In comparison to common protocols, like data-parallel training and data-parallel training with local SGD, ResIST yields a decrease in communication and compute requirements, while being competitive with respect to model performance.

  PublisherOA PDFDOI

• The Cumulative Perioperative Model: Predicting 30-Day Mortality in Abdominal Surgery Cancer Patients

  Journal of Surgical Oncology · 2020-03-10 · 1 citations

  articleOpen access

  OBJECTIVES: 1) To develop a cumulative perioperative model (CPM) using the hospital clinical course of abdominal surgery cancer patients that predicts 30 and 90-day mortality risk; 2) To compare the predictive ability of this model to ten existing other models. MATERIALS AND METHODS: We constructed a multivariate logistic regression model of 30 (90)-day mortality, which occurred in 106 (290) of the cases, using 13,877 major abdominal surgical cases performed at the University of Texas MD Anderson Cancer Center from January 2007 to March 2014. The model includes race, starting location (home, inpatient ward, intensive care unit or emergency center), Charlson Comorbidity Index, emergency status, ASA-PS classification, procedure, surgical Apgar score, destination after surgery (hospital ward location) and delayed intensive care unit admit within six days. We computed and compared the model mortality prediction ability (C-statistic) as we accumulated features over time. RESULTS: We were able to predict 30 (90)-day mortality with C-statistics from 0.70 (0.71) initially to 0.87 (0.84) within six days postoperatively. CONCLUSION: We achieved a high level of model discrimination. The CPM enables a continuous cumulative assessment of the patient's mortality risk, which could then be used as a decision support aid regarding patient care and treatment, potentially resulting in improved outcomes, decreased costs and more informed decisions.

  PublisherOA PDFDOI

• Workshop on Quantification, Communication, and Interpretation of Uncertainty in Simulation and Data Science

  arXiv (Cornell University) · 2020 · 4 citations

  • Computer Science
  • Data Mining
  • Computer Science

  Modern science, technology, and politics are all permeated by data that comes from people, measurements, or computational processes. While this data is often incomplete, corrupt, or lacking in sufficient accuracy and precision, explicit consideration of uncertainty is rarely part of the computational and decision making pipeline. The CCC Workshop on Quantification, Communication, and Interpretation of Uncertainty in Simulation and Data Science explored this problem, identifying significant shortcomings in the ways we currently process, present, and interpret uncertain data. Specific recommendations on a research agenda for the future were made in four areas: uncertainty quantification in large-scale computational simulations, uncertainty quantification in data science, software support for uncertainty computation, and better integration of uncertainty quantification and communication to stakeholders.

  PublisherOA PDFDOI

• Scalable linear algebra on a relational database system

  Communications of the ACM · 2020 · 11 citations

  Senior authorCorresponding
  • Computer Science
  • Computer Science
  • Theoretical computer science

  As data analytics has become an important application for modern data management systems, a new category of data management system has appeared recently: the scalable linear algebra system. We argue that a parallel or distributed database system is actually an excellent platform upon which to build such functionality. Most relational systems already have support for cost-based optimization---which is vital to scaling linear algebra computations---and it is well known how to make relational systems scalable. We show that by making just a few changes to a parallel/distributed relational database system, such a system can become a competitive platform for scalable linear algebra. Taken together, our results should at least raise the possibility that brand new systems designed from the ground up to support scalable linear algebra are not absolutely necessary, and that such systems could instead be built on top of existing relational technology.

  PublisherOA PDFDOI

• Declarative Parameterizations of User-Defined Functions for Large-Scale Machine Learning and Optimization

  IEEE Transactions on Knowledge and Data Engineering · 2018-10-01 · 11 citations

  articleSenior author

  Large-scale optimization has become an important application for data management systems, particularly in the context of statistical machine learning. In this paper, we consider how one might implement the join-and-co-group pattern in the context of a fully declarative data processing system. The join-and-co-group pattern is ubiquitous in iterative, large-scale optimization. In the join-and-co-group pattern, a user-defined function g is parameterized with a data object x as well as the subset of the statistical model Θ <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x</sub> that applies to that object, so that g(x|Θ <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">x</sub> ) can be used to compute a partial update of the model. This is repeated for every x in the full data set X. All partial updates are then aggregated and used to perform a complete update of the model. The join-and-co-group pattern has several implementation challenges, including the potential for a massive blow-up in the size of a fully parameterized model. Thus, unless the correct physical execution plan be chosen for implementing the join-and-co-group pattern, it is easily possible to have an execution that takes a very long time or even fails to complete. In this paper, we carefully consider the alternatives for implementing the join-and-co-group pattern on top of a declarative system, as well as how the best alternative can be selected automatically. Our focus is on the SimSQL database system, which is an SQL-based system with special facilities for large-scale, iterative optimization. Since it is an SQL-based system with a query optimizer, those choices can be made automatically.

  PublisherDOI

Recent grants

• SEI: Data Mining for Multiple Antibiotic Resistance

  NSF · $595k · 2006–2010

• ABI Innovation: Algorithms and Models for Distributed Computation of Bayesian Phylogenetics

  NSF · $1.2M · 2014–2019

• III: Medium: Collaborative Research: Data Mining and Cleaning for Medical Data Warehouses

  NSF · $600k · 2010–2015

• CAREER: New Technologies for Online Aggregation

  NSF · $440k · 2004–2010

• III: Medium: SimSQL: A Database System Supporting Implementation and Execution of Distributed Machine Learning Codes

  NSF · $1.2M · 2014–2020

Frequent coauthors

• Luis L. Perez

  University of Florida

  12 shared

• Abhijit Pol

  University of Florida

  8 shared

• Alin Dobra

  University of Florida

  7 shared

• Risa B. Myers

  Rice University

  7 shared

• Zekai J. Gao

  6 shared

• Shantanu Joshi

  6 shared

• Subramanian Arumugam

  6 shared

• Mingxi Wu

  Bozhou People's Hospital

  5 shared

Awards & honors

• 2008 Alfred P. Sloan Foundation Research Fellowship
• 2007 ACM SIGMOD Best Paper Award
• 2009 ACM SIGKDD Best Paper runner up
• 2017 ICDE Best Paper Award