Artificial intelligence (AI) has become an indispensable tool in the analysis of microscopic data. However, as AI models become more complex, the associated computing power and energy consumption also increase. In light of this, researchers at the Leibniz-Institut für Analytische Wissenschaften (ISAS) and Peking University have developed EfficientBioAI, an open-source compression software that enables scientists to run existing bioimaging AI models faster and with lower energy consumption.
EfficientBioAI was introduced in an article published in Nature Methods by the researchers from ISAS. The software aims to address the challenges of high-latency image analysis and restricted computing power by utilizing model compression techniques. These techniques reduce the computational requirements of AI models while maintaining the accuracy of predictions.
Dr. Jianxu Chen, head of the AMBIOM research group at ISAS, explained that high network latency and large images lead to increased computing power and energy consumption. By employing model compression techniques, such as pruning excess nodes in the neural network, memory consumption and model inference latency can be reduced.
Yu Zhou, a Ph.D. student at AMBIOM and the first author of the paper, emphasized the importance of making these techniques accessible to the bioimaging community. EfficientBioAI offers a user-friendly and ready-to-use solution that can be seamlessly integrated into existing PyTorch libraries, without requiring code modifications for widely used models like Cellpose. The toolbox also provides demos and tutorials to support specific customization requests.
To evaluate the effectiveness of EfficientBioAI, the researchers conducted tests on various real-life applications. The results demonstrated significant reductions in latency and energy consumption, ranging from 12.5% to 80.6%, depending on the hardware and bioimaging analysis tasks involved.
Dr. Chen provided an example to illustrate the potential energy savings. If a thousand users were to compress the commonly used Cellpose model with EfficientBioAI and apply it to a dataset of approximately one million microscope images, the energy saved would be equivalent to the emissions from a car journey of around 7,300 miles.
EfficientBioAI not only offers adjustable compression levels but also allows for effortless switching between the central processing unit (CPU) and graphics processing unit (GPU), providing further flexibility for users.
The researchers are committed to continuously improving the toolbox and are currently working on making it compatible with MacOS, in addition to Linux and Windows systems. Although the current focus is on improving inference efficiency, the team aims to explore ways to increase training phase efficiency in the future.
In summary, EfficientBioAI is a free compression software developed by researchers at ISAS and Peking University, enabling scientists to run bioimaging AI models faster and with lower energy consumption. By reducing latency and cutting down on computing power requirements, this toolbox presents an efficient and eco-friendly solution for analyzing large-scale microscopic data sets. Through its user-friendly interface and integration capabilities, EfficientBioAI aims to facilitate the widespread adoption of model compression techniques in the biomedical research community.
