Tianyi Cloud won the championship in the International AI Summit Large Model Challenge

On June 7, the first large model challenge (CVPR 2023 Workshop on Foundation Model: 1st foundation model challenge) held by the top international artificial intelligence conference CVPR 2023 came to an end. This competition attracted well-known universities and researchers from around the world. 1024 contestants from well-known companies. After two months of fierce competition, the Tianyi Cloud AI team (team name CTRL) performed well in the multi-task large model track and won the championship of this competition.

(Picture source: Photo Network)

The CVPR conference is an international academic conference on computer vision and pattern recognition hosted by IEEE. It contains the latest research results and technological developments in this field. It is one of the three top conferences on computer vision in the world.

The traditional visual model production process usually uses a single task and starts training from scratch, and each task cannot learn from each other. Due to limited single-task data, the actual effect of the model is too dependent on the task data distribution, and the generalization effect for different scenarios is usually poor.

In recent years, big data pre-training technology has developed rapidly. By using large amounts of data to learn general knowledge and transferring it to downstream tasks, it essentially achieves mutual learning between different tasks. The pre-trained model based on massive data has good knowledge completeness and can still achieve good results even if a small amount of data is used for fine-tuning in downstream tasks. However, the model production process based on pre-training and downstream task fine-tuning requires separately training models for each task, which consumes a lot of resources in research and development. In contrast, the multi-task training scheme trains a powerful general model using data from multiple tasks, which can be directly applied to handle multiple tasks, thereby effectively improving model productivity and generalization capabilities.

In this competition, contestants need to use a single model to simultaneously complete the joint training of three representative tasks: classification, detection and segmentation in traffic scenes. Tianyi Cloud AI team relied on its rich experience in algorithm development in model design and selected a pre-trained model with only 60% of the parameters of the second place, achieving higher accuracy with fewer parameters.

In order to solve the problem of slow convergence caused by inconsistent loss functions and gradients of each branch in multi-task training, the Tianyi Cloud AI team adopted the method of loss equalization and gradient scale unification to balance the loss functions of each task branch and make The gradient has a consistent scale, thereby improving the training efficiency and convergence speed of the model. In addition, the Tianyi Cloud AI team also uses carefully designed task-specific feature pyramids and attention mechanisms to enable each branch task to utilize features in the backbone network that are more effective for its own tasks, further improving the accuracy and performance of the overall model.

Through the above model design and training strategies, Tianyi Cloud AI team achieved excellent results in the competition, fully demonstrating its deep accumulation and continuous innovation capabilities in the fields of image, audio and multi-modality. In the future, Tianyi Cloud will continue to innovate and explore in the vast field of artificial intelligence, benefit more users with more advanced technology and excellent results, and provide support for the digital development of thousands of industries.

The above is the detailed content of Tianyi Cloud won the championship in the International AI Summit Large Model Challenge. For more information, please follow other related articles on the PHP Chinese website!