Column: Google’s Computer Vision Reinforcement Learning Patent — Learning Rewards, Not Just Answers

AI Times, a Korean AI publication, has published Managing Partner Yongduck Kim's column on artificial intelligence patents.

In this column, we look at reinforcement learning-based fine-tuning technology applied to the computer vision field, focusing on Google's U.S. Patent No. 12,731,392. So far, computer vision AI has been learning with a focus on how accurately it predicts correct answer data prepared by humans. However, in actual industrial settings, it is not enough to get the right answer from the training data. This is because how effectively AI achieves its goals in actual work is more important.

Google's patent notes this difference. The method is to first build a pre-trained computer vision model, then set an important performance indicator in actual work as ‘Task Reward’ and fine-tune the model again through reinforcement learning. For example, in object detection, actual performance indicators such as recall or average precision (mAP) can be used as compensation, and we propose that this method can also be applied to various computer vision tasks such as image segmentation and image colorization.

The column specifically focuses on the differences between correct answer data and actual work performance. In the existing method, the important goal was for the model to produce results similar to the correct answers of the training data, but by applying reinforcement learning, the goals that are evaluated as important in the actual system can be directly reflected in learning. In addition, it is differentiated from existing computer vision learning methods in that it does not necessarily use a differentiable loss function, but can also use evaluation criteria that are complex or difficult to calculate directly as compensation.

These technologies may have greater significance in areas where AI makes judgments and acts in real environments, such as autonomous driving, robots, smart factories, and medical imaging. Especially in the era of physical AI, it may become more important to reliably perform actual tasks based on recognition results rather than simply accurately recognizing objects in images. Google's latest patent can be said to be an example of the expansion of reinforcement learning technology, which has been developed around generative AI, into the areas of computer vision and physical AI.

Read the column to explore developments in AI training and how patents protect these technologies.

The full article can be found at the link below.