YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information

The authors of this paper identify a fundamental limitation in deep learning: as input data passes through deep neural networks, it undergoes spatial transformations that lead to significant information loss. This phenomenon, known as the information bottleneck, causes models to lose critical data required for accurate predictions, resulting in biased gradient flows and poor model convergence.

To address this, the authors introduce Programmable Gradient Information (PGI). PGI is an auxiliary supervision framework that ensures the model retains essential information throughout the training process. It works by creating an auxiliary reversible branch that provides reliable gradient information to the main network, allowing the model to learn more effectively without incurring additional costs during inference.

Additionally, the authors propose the Generalized Efficient Layer Aggregation Network (GELAN). This new, lightweight architecture is designed based on the principles of gradient path planning. GELAN demonstrates that it is possible to achieve state-of-the-art performance using only conventional convolution operators, outperforming more complex models that rely on depth-wise convolution.

• Overcoming Information Bottlenecks: By using PGI, the authors successfully mitigate the information loss that typically occurs in deep networks, ensuring that even lightweight models receive reliable gradients during training.
• Superior Efficiency: The GELAN architecture achieves better parameter utilization and accuracy than existing state-of-the-art models, proving that high-performance object detection does not require overly complex or computationally expensive operations.
• Performance Gains: When combined, PGI and GELAN form YOLOv9. Experimental results on the MS COCO dataset show that YOLOv9 outperforms previous real-time object detectors, including those pre-trained on massive datasets, even when trained from scratch.
• Inference Efficiency: Because PGI is only utilized during the training phase, the resulting YOLOv9 model maintains high speed and efficiency during inference, as the auxiliary branches are discarded.