Traffic Sign Classification with Reinforcement Learning from Human Feedback

Reinforcement learning from human feedback (RLHF) can handle complex tasks such as natural language processing, game playing, robotics, and recommendation systems. The successful application of RLHF in these fields has inspired our exploration of image classification. In this paper, we explore using RLHF to fine-tune the weights of the LeNet-5 network to improve the accuracy of traffic sign classification models. Traditional image classification models only partially utilize human expert knowledge during training. RLHF provides a novel approach to optimize the model weights by introducing human feedback, making the model perform better than before. Our experiments use the classic LeNet-5 network to perform preliminary training on the traffic sign classification task. We then design an RLHF-based fine-tuning process where the LeNet-5 model adjusts its weights based on human feedback. We precisely quantify human feedback through a reward function, and then the agent is trained using the reward values generated by the reward function. Subsequently, the weights of the model are adjusted based on the actions taken by the agent. Experimental results show that the performance metrics of the LeN et-5 network fine-tuned by the RLHF method are higher than the original LeN et-5 model in the traffic sign classification task. This result demonstrates the importance of human feedback in the model optimization process and shows the potential of RLHF in improving image classification.