UNIVERSITY PARK, Pa. — A team of researchers at Penn State are following a “selective learning” approach — where scientists only collect data needed for a specific problem, instead of collecting all available data and sorting through to find what applies to a problem — to make artificial intelligence (AI)-powered systems like unmanned autonomous vehicles (UAVs) more efficient.
“The approach is akin to the ‘selective attention’ mechanism of the human brain. The brain only gathers the necessary information to conserve energy, memory and processing resources, like selective learning,” said Soundar Kumara, Allen E. and Allen M. Pearce Professor of Industrial Engineering at Penn State, who led a team that recently published their approach in [Scientific Reports](https://www.nature.com/articles/s41598-024-83706-8). “Because sensors operate in the real world with compute, power, storage, transmission and latency constraints, it is important to build efficient AI models that can operate in low size, weight, power and cost scenarios. This research leverages insights from biology to design novel data and compute-efficient algorithms.”
By combining a streamlined training approach, the team’s method reduces the amount of data, computing power and energy the AI-powered systems need to maintain desired performance accuracies. The proposed approach results in up to a 435-fold reduction in computing resources for applications in UAVs, wearable electronics and more.
The team includes Kumara; Ankur Verma, who earned a doctorate from Penn State in 2024; Ayush Goyal, who graduated from Penn State in 2024 with a master’s degree in computer science; and Sanjay Sarma, Fred Fort Flowers and Daniel Fort Flowers Professor in Mechanical Engineering at the Massachusetts Institute of Technology.
In the Q&A below, Kumara, who also directs the Center for Applications of Artificial Intelligence and Machine Learning to Industry at Penn State, discussed AI-powered vehicles and the work making them more efficient.
**Q: What are AI-powered vehicles? Where do they appear in everyday life?**
**Kumara:** UAVs are aerial, on-land or underwater vehicles with autonomous navigation capabilities, meaning that these vehicles use data collected by sensors to drive themselves rather than relying on a human pilot. In everyday life, these vehicles could be used for search and rescue or resupply missions, infrastructure inspection, lunar and Martian exploration and navigating hazardous environments. Basically, we can put them in environments that may be dangerous for or inaccessible to humans to help achieve our goals.
UAVs are a specific example of an asset that use different sensors like cameras that can see colors humans can detect, multi-spectral cameras, gyroscopes and accelerometers, that can benefit from our paper’s proposed approach. This approach has been tested by our team on sensor data from a variety of assets like industrial motors, pumps, gearboxes, gas turbines and more. Our approach exploits the structure or patterns of real-world sensor data, which is a universal property across different data modalities, such as images, sounds, or vibrations.
**Q: How much data and energy do AI-powered unmanned autonomous vehicles consume?**
**Kumara:** Data consumption by different modalities differs depending on the UAV or wearable. The amount of sensor data to be collected for various sensing modalities and tasks is governed by the Shannon-Nyquist sampling theorem, which states that the sampling rate needs to be at least twice the highest frequency present in the signal to avoid information loss. For machines with fast underlying dynamics such as UAVs, sampling frequencies need to be in kilohertz or megahertz range — 10,000 to 1,000,000 data points per second. That’s massive amounts of data in a very short period of time. Another reason for the high data volumes are the channels of data. An aircraft may have up to 300 different sensors for measuring different physical properties which generate several gigabytes of data per flight.
Based on the results from my team’s research, we can expect a 7- to 435-fold reduction in the energy required for processing the data from different sensor modalities that may be present in various UAVs. Current UAVs require charging and are typically battery powered. Using our new approach, we could significantly extend the battery life as the amount of computing power required to analyze the sensor data would go down significantly. This could also help in reducing the weight of the processors that need to fly, giving a weight reduction advantage as well.
**Q: How is your lab working to make AI-powered UAVs more energy- and data-efficient?**
**Kumara:** Ankur Verma, the first author on this paper who earned a doctorate from Penn State in 2024 under my supervision, worked on developing novel neural network architectures that can jointly sample and train models on much smaller amounts of data, resulting in fast training models and inference. Instead of a two-step process of collecting and discarding raw data, our approach combines the two by only collecting the useful data needed in the first place. The combined sampling, or joint sampling, and fast model training reduces the amount of data required by 10-fold and compute — the power necessary to complete and perform computational tasks — by 435-fold. This also results in smaller models that are lighter and faster with less energy consumption during inference. These models can also be trained much faster than larger models.
Our paper in Scientific Reports showed a 435-fold reduction in compute power in comparison to typical neural networks, which require considerable training data, energy and compute to learn representative properties about the data. We exploit the structure in the data to reduce all three — data, compute and energy — while maintaining performance accuracies. Energy reduction is very important for deploying AI in the physical world, as these scenarios often include compute, power, storage, transmission and latency-restricted environments. Bringing AI on satellites, drones, UAVs, Mars rovers, etc., requires AI to be very resource efficient as we cannot increase the form factor of these devices due to operational and economic constraints.
**Q: What’s next?**
**Kumara:** We have already applied our research to a few different use cases, including industrial asset health monitoring of gas turbines, blowers, motors and pumps to avoid unplanned downtime and human activity recognition for wearables on a very small hardware form factor — physical specifications like size, shape and layout of the wearable computer hardware components — identifying activities like walking, running and swimming. We believe that our methodology will change the way next generation edge-computing will transform. We are working on expanding our research through Lightscline, the company which Verma, Goyal and I founded. Through Lightscline, we hope to commercialize this technology in various industries like space and defense.
_The initial market study for this work is funded by the U.S. National Science Foundation I-Corps program._