Google Retrains Computer Vision With Reward-Based Criteria

According to AI TIMES, Google proposed, through U.S. Patent No. 12731392, a method for retraining computer vision models so they are not trained only to match ground-truth data but are instead rewarded for actual work performance. The patent, introduced on September 14, focuses on fine-tuning after pretraining with reinforcement learning so that the model directly targets evaluation criteria that matter in the field. The key idea is that accurately mimicking labels in training data and performing well in real-world environments can be different things. Accordingly, it opens a path to linking training with criteria that better reflect on-the-ground performance across tasks such as object detection, image segmentation, and colorization. In particular, it notes that a reward function does not have to be differentiable, arguing that human judgments or composite performance assessments can also be pulled into training objectives. The article interprets this as a sign that the alignment trend in generative AI is spreading to computer vision and physical AI, and says that in the future, competitiveness may depend less on the size of the model itself than on how well goals are designed as rewards.