ABE’s Chaterji receives $550,000 NSF CAREER grant for streaming analytics and IoT project

Somali Chaterji, assistant professor, was awarded a $550,000 CAREER grant from National Science Foundation’s CISE (Computer and Information Science and Engineering) Directorate.

Somali Chaterji, assistant professor in the Department of Agricultural and Biological Engineering (ABE), has been awarded a $550,000 CAREER grant from National Science Foundation’s CISE (Computer and Information Science and Engineering) Directorate for her project Sirius. The title of her work is “Robust and Adaptive Streaming Analytics for Sensorized Farms: Internet-of-Small-Things to the Rescue.”

Sirius makes machine learning (ML) feasible for the Internet of Small Things (IoST) world by creating a compute fabric that is adaptive to cyber and physical conditions, providing prompt actuation while being resilient to failures.

Somali Chaterji
Somali Chaterji

“We live in an increasingly sensorized world generating a deluge of streaming data, both low- and high-bandwidth,” said Chaterji, who also serves as director of the Innovatory for Cells and Neural Machines (ICAN) and on the leadership team for WHIN (Wabash Heartland Innovation Network) Digital Ag. “An example of high-bandwidth data is video from drone surveillance. Leveraging data effectively from this continuum of sensors, static and mobile, we can convert data to decisions. This can go a long way in solving the food and ecological problems faced by the planet today.”

Johnny Park, CEO of WHIN, is a Sirius collaborator and proponent of the work.

“Somali's ideas for building a living testbed of IoT devices, static and mobile, ground-based and aerial, are superb. She has shared her plans of bringing realistic applications to this testbed, which will prove out the validity of her algorithms," Park said.

Chaterji says two trends in computation have catalyzed the IoST in relation to digital and sustainable agriculture.

First is the growing availability of inexpensive sensors that are robust to the rigors of agriculture. Second is the development of approximation algorithms for on-device computation of heavyweight data analytics algorithms.

“In parallel, some demanding algorithms can be opportunistically offloaded to edge devices or to the cloud when the deadlines to make decisions are not that tight. There is an increasing trend to leverage the data from these IoST nodes to actuate dependable, prompt, and resilient actions,” Chaterji explained.

Dependable means the algorithms need to deal with missing or corrupted data, network disruption, and node failures. Prompt refers to low-latency decisions, which enable timely decisions for farmers and digital agriculture providers.

An external collaborator, Ranveer Chandra, managing director at Microsoft Research, praised the project, saying, “This work is innovative as well as timely. Somali is in a unique position to bring together her insight in IoT and federated data analytics to the application domains of drone surveillance and sustainable agriculture.”

Sirius is structured around three overarching thrusts:

• First, the project will approximate computation, adapting to the capability of heterogeneous devices and to the network condition, workload characteristics, and contending applications that are using the same compute fabric. On-device computation reduces data transfer costs, can work in disconnected mode, and makes the entire solution scalable.

• Second, Sirius will leverage decentralized learning that is resilient to the uncertainty and heterogeneity of the devices — think of devices ranging from tiny microcontrollers to more powerful mobile GPUs.

• Finally, Sirius will perform partitioning of the computation between the device, edge, and cloud to reduce the end-to-end latency while reducing the cloud computing costs of execution.

“My project will, for the first time, deliver technologies that are considered today to be outside the reach of IoST — decentralized learning, in-network analytics on continuously streaming data, and serverless computing — all on heterogeneous and failure-prone small devices.”