Hello! I am a Ph.D. student at Carnegie Mellon University in the
Machine Learning Department. I conduct research on how machine
learning and generative AI can support personalization in
education. I am currently advised by
Dr. Zachary Lipton.
Previously, I have worked on AI safety and large language models
projects through internships at Microsoft Research (advised by
Dr. Adam Kalai
&
Dr. Rosa Arriaga), Indico Data (advised by
Madison May), and The MITRE Corporation (advised by
Dr. John Henderson).
Large language models trained at scale show emerging intelligent
behavior, such as coherent and grammatical structures, cultural
knowledge, and abstract reasoning capabilities.
@ Microsoft Research, I used GPT-3 and
Turning-NLP to simulate distributions of responses to psychology
experiments.
Model Robustness
Adversarial attacks may manipulate the behavior of AI systems to serve
a malicious end goal.
@ The MITRE Corporation, I prototyped a
docker containerized adversarial attack testing platform and
populated a public information resource.
@ MITRE NLP Lab, I supported research
into practical attacks on machine translation using paraphrase.
Cross-Source Information Extraction
Extracting information from documents requires the ability to link
events, entities and associated relations across multiple sources.
@ Indico Data R&D, I worked on deep
learning NLP and CV approaches to PDF information extraction.
@ Olin Satellite Lab, I consolidated
multiple possibly contradictory data sources when scraping the FCC's
international filings database.
Event Sequence Modeling
With language, voice, and time-series data, data items are dependent
on data before or after it.
@ Olin Microbiology Lab, I
characterized time-series data from perturbed and recovering
microbial communities using methods from compositional data
analysis.
@ Fidelity R&D, I analyzed
distributions of cryptocurrency technical trading indicators over
time.
Led research on analyzing composition shifts in time-series of
cultured, perturbed microbiomes.
Conducted literature review to find, apply, and analyze
limitations of Random Matrix Theory approach, Compositional Data
Analysis, and network analysis.
Presented poster at Northeastern Microbiologists: Physiology,
Ecology, and Taxonomy (NEMPET).
Led project on cleaning and interpreting 2D Fourier analysis to
isolate patterns in bacterial surface images to identify pattern
and shape of surface proteins.
Worked on undergraduate and industry research team to identify
factors driving value in 5G spectrum auctions.
Applied statistical analysis tests (ANOVA), and analyzed auction
context and mechanics
Presented session at Research Conference on Communication,
Information, and Internet Policy (TPRC).
Led project to automate web-scraping, PDF-text extraction, and
data cleaning of satellite filings from the FCC's International
Bureau Filing System, creating data sets to support two new
analysis projects and generate yearly review of satellite
filings.
Advised by Dr. Adam Kalai (Microsoft Research) and Professor Rosa
I. Arriaga (Georgia Institute of Technology)
Led research on using large language models (GPT-3, Turing-NLG)
to simulate demographically-aligned distributions of human
behavior on behavioral economics, psycholinguistics, and social
psychology experiments, resulting in paper (preprint on arXiv).
Designed zero-shot prompt methodology, compared predicted
distributions to literature on Ultimatum Game, Garden Path
sentence comprehension, and Milgram Shock Experiment.
Designed and ran novel alternatively worded prompts to mitigate
risk of models regurgitating training data.
Working with cognitive psychology and behavioral language model
researchers Professor Rosa I. Arriaga (Georgia Institute of
Technology) and Professor Micheal Kearns (University of
Pennsylvania) to evaluate limitations of method before
submitting to top-tier journal
As fullstack software engineer supporting team with four UXD and
product interns, protoyped and user-tested high-fidelity novel
React.js GUI for predicting and correcting groups of text
extractions.
In independent research project with R&D team, adapted object
detection Faster R-CNN model to classify handwriting on business
documents and incorporated methods for alternate pre-training,
multi-label tasks, and small object detection.
Both features were highlighted as two of the top five features
of Indico's fifth major release.
Software engineering intern working on Practical Attacks on
Machine Translation using Paraphrase, advised by Dr. John
Henderson (Principal Investigator)
Researched new methods to exploit vulnerabilities in natural
language machine learning systems.
Revived and adapted an academic lab's research code for
generating a paraphrase database using the bilingual pivoting
technique, debugged it in a new environment, incorporated newer
software packages, ran timing experiments to determine hardware
needs and Hadoop configuration for running resource-intensive
computation with 4x more data, modified code to run faster for
our specific use-case.
Augmented dataset with 900 million segments of parallel text,
performed word-alignment, and implemented methods for scrubbing
poorly aligned segments.
Used Logistic Regression for classification on an imbalanced
dataset, and improved performance with feature engineering
ablation studies.
High school researcher in the Software & Application Innovation
Lab, advised by Lucy Qin (PhD candidate), Kinan Dak Albab (PhD
candidate), and Dr. Andrei Lapets (Principal Investigator)
Evaluated secure multi-party computation JavaScript library
(JIFF) on sorting and set-intersection algorithms.
Wrote demos with different algorithm implementation for data
oblivious sorting algorithms and set-intersection tasks,
calculated throughput and latency, and implemented benchmark
tests.
Contributed to the development of new functionalities that
performed 92-95% faster than the existing JIFF functions on
common use-cases.
Presented poster at the Greater Boston Research Opportunities
for Women (GROW) research conference.
Peer-Reviewed Publications
Using large language models to simulate multiple humans and replicate
human subject studies G. Aher, R. I. Arriaga, and A. T. Kalai. ICML 2023, *Oral.
Evaluating the FCC's $10 Billion Gamble: Successfully Accelerating
Access to Spectrum in Auction 107 G. Aher, P. Post, P. Boyalakuntla, G. Miner, L. Heinrich, Y. Mao,
J. A. Musey, W. Lohmeyer. Journal of Information Policy (JIP 2023).
Analysis of Geostationary Federal Communication Commission Satellite
Applications from 2000 to 2022
P. Post, K. Fleming, K. Canavan, S. Cho, G. Aher, W. Lohmeyer.
Journal of Spacecraft and Rockets (2023).
Posters
What Factors Affect Microbial Community Composition?
Northeastern Microbiologists: Physiology, Ecology, and Taxonomy (NEMPET
2021)
SOARing with Drones in Education Massachusetts Computer Using Educators (MassCUE 2018)
Refining Private Set Intersection Under Secure Multi-Party
Computation Boston University, Greater Boston Research Opportunities for Women
(GROW 2018)
Artificial Intelligence, Chatbots, and Amazon Web Services International Society of Technology Educators (ISTE 2018)
Projects
400+ hours designing, building, integrating, and collecting gesture samples
