Photo courtesy of the Colorado Tourism Office
As enthusiasm for and success in the Internet of Things (IoT), Cyber-Physical Systems (CPS), and Smart Buildings continues to grow, so too does the volume and variety of data generated by these systems. This raises important questions: How can we ensure high-quality data collection? And how can we maximize the utility of this data so that multiple projects can benefit from the time, cost, and effort invested in deployments?
With the rise of Foundational Models—particularly Large Language Models (LLMs)—we now have new tools that can potentially transform how we work with cyber-physical data. Yet, real-world data presents notable challenges, including diverse modalities, limited dataset sizes, and unstructured formats. Recent advances in large AI models, especially those based on transformer architectures, offer promise for improving how data is acquired, analyzed, manipulated, and consumed.
The DataFM: Data Acquisition & Analysis with Foundational Models workshop aims to look broadly at interesting data from interesting sensing systems and/or how such data can be adapted to Foundational Models. The workshop considers problems, solutions, and results from all across the real-world data pipeline. We solicit submissions on unexpected challenges and solutions in the collection of datasets, on new and novel datasets of interest to the community, on experiences and results, explicitly including negative results, in using prior datasets to develop new insights, and on discussions of impact and newfound opportunities with large AI foundational models.
Foundational Models could enhance data quality through sophisticated data cleaning, preprocessing, and augmentation techniques. They can also facilitate the analysis of data streams while identifying anomalies, inconsistencies, and potential biases. Generative AI can also create synthetic datasets that maintain the essential characteristics of real-world data while expanding the available training samples. This may be valuable when real data is challenging due to privacy concerns or logistical constraints. Transformer models can integrate multi-modal data, such as blending textual inputs from sensor logs with quantitative data from measurements. This new flavor of AI-driven analysis can factor in more contextual information, opening new areas of research in enhancing the predictive and diagnostic capabilities of data-driven AI systems deployed in smart environments.
Furthermore, new areas of future work may emerge from exploring the ethical implications of deploying Foundational Models within these domains, ensuring that the benefits of AI are equitably distributed while safeguarding user privacy. The workshop's focus on privacy challenges and solutions becomes increasingly relevant in the era of AI, where the capacity to analyze vast amounts of sensitive data poses significant risks.
The workshop aims to bring together a community of application researchers and algorithm researchers in the sensing systems and building domains to promote breakthroughs from the integration of the generators and users of datasets. The workshop will foster cross-domain understanding by enabling both the understanding of application needs and data collection limitations.
The workshop seeks contributions across two major thrusts, but is open to a broad view of interesting questions around the collection, dissemination, and use of data as well as interesting datasets:
To enable the longevity of submitted datasets, we plan on providing a central location where a repository for the data, and information about the data can be archived for at least 5 years.
Important: Each accepted submission is required to have at least one author attend the workshop and present to the workshop attendees.
Submissions may range from 2-5 pages in PDF format, excluding references, using the standard ACM conference template. Submissions are strongly encouraged to use only as much space as needed to clearly convey the ideas, contributions and the significance of the work.
Dataset submissions should prefix paper titles with "Dataset:" and must include a description of the dataset as well as a reasonable accompanying data sample. Once accepted, a fully described dataset must be shared to a public repository by the camera-ready deadline.
Issues on licenses will be resolved following procedures similar to CRAWDAD.
Datasets will be reviewed by an artifact evaluation committee. To support this, dataset submissions must include:
Full dataset access (not just samples)
Demonstrate potential insights from the data
Code samples, videos, or demonstrations
Submit your papers through the conference submission system:
Submission link will be available soon
Augusta University
UMass Amherst
Augusta University