Improved ResNet + XGBoost

[frozenNoob]

Experimental results demonstrate that the proposed improvements to ResNet led to an approximate 1% increase in accuracy on the CIFAR-100 dataset. Furthermore, by integrating the XGBoost algorithm—thus combining deep learning with ensemble learning—accuracy improved by an additional ~1%, achieving a maximum accuracy of 66.63% under the Improved ResNet + XGBoost configuration.
Although these results were attained through extensive parameter tuning, the overall findings indicate that the model’s performance was indeed enhanced under certain conditions. Due to constraints in computational resources and experimental costs, further exploration was not pursued. Nevertheless, the implementation provides at least two concrete innovative approaches that may serve as a reference for beginners. While both ideas were conceptually adapted from prior research and integrated into this work, all corresponding code was developed independently by myself based on the original code provided, as no direct code implementations from the referenced studies were found. Achieving such promising results under these circumstances is, in itself, worthy of dissemination.

实验结果表明，对 ResNet 的改进使 CIFAR-100 数据集上的准确率提升了约 1%。进一步引入 XGBoost 算法（即深度学习与集成学习的结合）后，准确率又提升了约 1%，在 “Improved ResNet + XGBoost” 配置下取得了 66.63% 的最高准确率。
虽然这些结果是在多次调参后获得的，但整体来看，模型在特定条件下的性能确实得到了提升。由于计算资源和实验成本的限制，本研究未能进行更深入的探索。然而，该实现至少提供了两种具有参考价值的创新方法，可供初学者借鉴。尽管这两个创新点在概念上参考了已有研究并进行了融合，但所有相关代码均在原始代码的基础上由本人独立编写，且未找到对应论文的现成代码实现。在这样的条件下取得较为理想的结果，本身就是一件值得分享的事情。