How Does ACCRE Support Generative AI Initiatives?
ACCRE is a key player for collaborations, support and expertise when it comes to shared campus resources that can be leveraged for generative AI research and exploration.
Ways We Support AI and Machine Learning
- Software Engineering Services: Our in-house team of engineers is available to provide customized software solutions tailored to your research needs.
- Specialized Hardware Support: We offer dedicated hardware resources, including NVIDIA DGX systems, optimized for machine learning tasks.
- Embedded Research Scientists: Our co-hired research scientists work directly within ACCRE, providing a seamless bridge between academic research and computational resources.
- GPU Resource Allocation: We provide GPUs for computationally-intensive projects, offering a range of options to fit diverse research requirements.
- Exploratory Research through ACCRE X Program: Our ACCRE X program provides a platform for proof-of-concept studies, giving researchers the opportunity to explore innovative ideas with the support of our advanced computing infrastructure.
- Documentation & User Support: Comprehensive guides and tutorials are available to help researchers effectively utilize our resources for specialized use-cases.
- Interdisciplinary Collaborations: ACCRE actively engages in partnerships across various departments and institutes, including the Data Science Institute and specialized research centers focused on AI applications in fields like protein dynamics.
Example Projects
Project: SUD Phenotyping, PI: Alvin Jeffery, School of Nursing
Acronym Detection and disambiguation depends largely on domain specificities. For instance, MS can disambiguate to Multiple Sclerosis or Milliseconds based on whether it is used in the context of a doctor’s note or a physics paper. For this project, a general acronym detection model was fine tuned to work with custom data to identify acronyms in a text. Furthermore, a pipeline to disambiguate the acronym was created using clinicalBERT. The pipeline uses pre-existing knowledge of an acronym as well as customized user-inputs to disambiguate it.
Answering certain questions from a patient’s notes can provide a lot of insights on whether that patient is susceptible to substance use disorder. An LLM based approach to answer these questions from the patients’ notes was created for this project.
Project: Brilliance, PI: Kathrine Aboud, Peabody
Human-in-the-loop is generally a bottleneck in data pipelines. This is because of the fact that human review helps to clean data and ensure optimal input to the processing pipeline. A pipeline was created to tackle this problem. Specifically, a combination of Whisper and GPT4 based solution was created. The pipeline creates transcripts of recorded audio sessions and further process those transcripts with a GPT4 based approach to obtain the required data point from input. With the pipeline in place, the reviewers now only review the validity of the data rather than having to manually create data points. This largely improves the use of reviewer’s time and, in turn, works to decrease the human-in-the loop bottleneck.