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

**Authors:** Chien-Yao Wang, I-Hau Yeh, Hong-Yuan Mark Liao  
**arXiv:** [2402.13616](https://arxiv.org/abs/2402.13616)  
**Date:** 2024-02-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.13616v1/

## Key points

- Current focus in deep learning: designing objective functions to optimize model predictions and gather sufficient information for accurate results
- Issue of information loss during feature extraction and transformation in deep networks
- Introduction of programmable gradient information (PGI) and reversible functions for data transmission through deep networks
- PGI ensures complete input information for target tasks, enabling accurate calculation of objective functions and reliable updates to network weights
- Proposal of Generalized Efficient Layer Aggregation Network (GELAN) based on gradient path planning principles
- Experimental results showing GELAN with PGI outperforming existing methods, utilizing conventional convolution operators more efficiently than state-of-the-art techniques relying on depth-wise convolution
- Versatility of PGI across various model sizes, facilitating superior results even in train-from-scratch models compared to pre-trained ones
- Potential of PGI and GELAN in enhancing deep learning methodologies by addressing data loss issues within deep networks through innovative approaches

## Summary

In the realm of deep learning, there is a current focus on designing objective functions to optimize model predictions and gather sufficient information for accurate results. However, existing methods often overlook the loss of information during feature extraction and transformation in deep networks. To address this issue, this paper introduces the concept of programmable gradient information (PGI) and reversible functions for data transmission through deep networks. PGI ensures complete input information for target tasks, enabling accurate calculation of objective functions and reliable updates to network weights. Additionally, a novel lightweight network architecture called Generalized Efficient Layer Aggregation Network (GELAN) is proposed based on gradient path planning principles. Experimental results on MS COCO dataset-based object detection tasks demonstrate that GELAN with PGI outperforms existing methods by utilizing conventional convolution operators more efficiently than state-of-the-art techniques relying on depth-wise convolution. Notably, PGI's versatility allows its application across various model sizes – from lightweight to large – facilitating superior results even in train-from-scratch models compared to those pre-trained on extensive datasets. This study showcases the potential of PGI and GELAN in enhancing deep learning methodologies and highlights the importance of addressing data loss issues within deep networks through innovative approaches for improved model performance and parameter utilization.