ViTAEv2 Breaks ImageNet Real Record with 91.2% Accuracy – How a 600M‑Parameter Model Redefines Few‑Shot Learning

JD Research Institute, in collaboration with the University of Sydney, announced the super deep learning model ViTAEv2, designed for larger scale, better performance, and broader adaptability across visual tasks.

ViTAEv2, with 600 million parameters and no reliance on external data, achieved a world‑leading 91.2% top‑1 accuracy on the ImageNet Real classification benchmark, setting a new record in image‑classification technology.

ImageNet, the largest public image‑classification dataset, has long served as a key metric for evaluating computer‑vision progress, attracting top tech companies and leading universities.

Computer vision, a core AI technology, aims to give machines the ability to observe, perceive, and understand images; image classification is its fundamental task. ViTAEv2 adopts a pre‑training‑fine‑tuning paradigm, leveraging inductive bias in large models and matching pre‑training and transfer‑learning algorithms to achieve its results.

The researchers also examined ViTAEv2’s few‑shot learning capability by fine‑tuning the model with 1 %, 10 %, and 100 % of the data. Even with only 10 % of the data, the large model outperformed smaller models trained on the full dataset, confirming strong representation, learning, and sample‑efficiency properties.

This demonstrates ViTAEv2’s potential to tackle low‑resource or zero‑resource tasks, lower data‑annotation costs, accelerate algorithm development cycles, simplify model deployment, and empower next‑generation automated learning technologies.

Looking ahead, ViTAEv2 is expected to drive progress in a range of visual tasks such as semantic segmentation, object detection, pose estimation, and video object segmentation, while future research will focus on improving performance and reducing training and inference costs through better training methods and model architecture design.