Digital Event Horizon
BioCLIP 2: A Biology-Based Foundation Model Trained on the World's Biggest Dataset of Organisms
A revolutionary new AI model, BioCLIP 2, has been trained on a massive dataset of over 214 million images of organisms from 925,000 taxonomic classes, pushing the boundaries of what is possible in the field of computational biology.
BioCLIP 2 is a groundbreaking new AI model that has been trained on the largest dataset of organisms to date.The model can distinguish species' traits, determine inter-and intraspecies relationships, and even predict an organism's health based on training data.BioCLIP 2 was developed by Dr. Tanya Berger-Wolf and her team at The Ohio State University using a vast dataset of 214 million images of organisms.The model was trained on 32 NVIDIA H100 GPUs in just 10 days, displaying novel abilities such as distinguishing between adult and juvenile animals within species.BioCLIP 2 makes associations between related species without being explicitly taught these concepts, learning the hierarchy through associations.The researchers plan to develop a wildlife-based interactive digital twin to visualize and simulate ecological interactions between species.The technology has the potential to be deployed for public use, such as through interactive platforms at zoos, allowing people to explore and learn about the natural environment from new vantage points.
The world of computational biology has just taken a giant leap forward with the introduction of BioCLIP 2, a groundbreaking new AI model that has been trained on the largest dataset of organisms to date. This revolutionary breakthrough is the result of years of research and development by Dr. Tanya Berger-Wolf, director of the Translational Data Analytics Institute and a professor at The Ohio State University.
Berger-Wolf's first computational biology project started as a bet with a colleague: that she could build an AI model capable of identifying individual zebras faster than a zoologist. She won. Now, she is taking on the whole animal kingdom with BioCLIP 2, a biology-based foundation model that can distinguish species' traits and determine inter-and intraspecies relationships.
The model goes beyond extracting information from images. It can also make associations between related species and even determine the health of an organism based on training data. For example, it separated healthy apple or blueberry leaves from diseased leaves, as well as recognized differing types of diseases.
To curate this vast amount of data, Berger-Wolf's team at the Imageomics Institute collaborated with the Smithsonian Institution, experts from various universities and other field-related organizations. The dataset, TREEOFLIFE-200M, comprises 214 million images of organisms that span over 925,000 taxonomic classes - from monkeys to mealworms and magnolias.
After 10 days of training on 32 NVIDIA H100 GPUs, BioCLIP 2 displayed novel abilities, such as distinguishing between adult and juvenile as well as male and female animals within species without being explicitly taught these concepts. It also made associations between related species - like understanding how zebras relate to other equids.
"This model learns that at every level of taxonomy, all of these images of zebras have a particular genus label, and of these images of equids - including zebras, horses and donkeys - they have a particular family trait and so on," said Berger-Wolf. "It learns the hierarchy without ever being told it, just through these associations."
Berger-Wolf's team used a cluster of 64 NVIDIA Tensor Core GPUs to accelerate model training, plus individual Tensor Core GPUs for inference. "Foundation models like BioCLIP would not be possible without NVIDIA accelerated computing," said Berger-Wolf.
The researchers' next endeavor is to develop a wildlife-based interactive digital twin that can be used to visualize and simulate ecological interactions between species as well as their ways of engaging with the environment. The goal is to provide a safe, easy way to study organismal relationships that naturally occur in the wild, while minimizing impact and disturbance on ecosystems.
"The digital twin allows us to visualize species interactions and put them in context, as well as to play the what-if scenarios and test our models without destroying the actual environment - creating as light a footprint as possible," said Berger-Wolf. "The digital twin will give scientists the opportunity to explore the points of view of the species they're studying within the simulated environment, opening endless possibilities for more complex and accurate ecological research."
Eventually, versions of this technology could even be deployed for public use - such as through interactive platforms at zoos. People could explore, visualize and learn about the natural environment and its many species from entirely new vantage points.
"I'm getting goosebumps just imagining that scenario of a kid coming into the zoo and being like, wow - this is what you would see if you were another zebra part of that herd, or if you were the little spider sitting on that scratching post," Berger-Wolf said.
BioCLIP 2 is available under an open-source license on Hugging Face, where it was downloaded over 45,000 times last month. This paper builds on the first BioCLIP model, released over a year ago, which was also trained on NVIDIA GPUs and received the Best Student Paper award at the Computer Vision and Pattern Recognition (CVPR) conference.
The BioCLIP 2 paper will be presented at NeurIPS, taking place Nov. 30-Dec. 5 in Mexico City, and Dec. 2-7 in San Diego. Building on this breakthrough, Berger-Wolf's team is poised to revolutionize the field of computational biology and push the boundaries of what is possible with AI.
Related Information:
https://www.digitaleventhorizon.com/articles/Unlocking-the-Secrets-of-Life-The-Breakthroughs-and-Innovations-of-BioCLIP-2-deh.shtml
https://blogs.nvidia.com/blog/bioclip2-foundation-ai-model/
Published: Thu Nov 20 09:42:09 2025 by llama3.2 3B Q4_K_M
