第五十三周学习笔记

第五十三周学习笔记

论文阅读概述

• Densely Connected Convolutional Networks：This article introduces a novel CNN model DenseNet which distills the idea of short cut connection and extend it to connecting every two layers by short cut connection and use concat instead of sum which is used in ResNet, achieveing better performance with fewer parameters.
• Detecting Visual Relationships with Deep Relational Networks: This article introduces Deep Relational Network to which incorporate object feature, union region feature, object mask and conditional distribution of relationship to predict relationship and use linked units to fuse different type of information instead of directly concat them, achieving better performance than previous approaches.

医学图像分类部分实验结果

本周主要在进行caption信息提取，并写完带提取信息的模型训练代码，目前caption信息已经提取完毕，模型正在训练中，在提取信息之余，将之前跑的几个baseline模型结果列在这里

Name	acc sick	acc dire	备注
both_3_9_2	86.55%	86.07%	resnet50在最后fc特征上分类
capsule2	83.73%	83.74%	Capsule Net的分类结果
ssick	84.28%		resnet50仅在sick上分类的结果，作为baseline
ccg	79.94%		CNN+capsule+GCN，没有用提取的caption label的结果
ccg_p6416_d8	81.72%		前一个网络参数重调，为避免过拟合略微降低了模型表达能力的结果

本周小结

上周任务基本 未完成

下周目标

• 不少于5篇的论文阅读
• 整理先前阅读的论文
• 整理论文的书写方法
• 整理重要的引用文献
• 完成基本模型的运行任务，比对其原论文详细分析结果
• 研究策略梯度、CIDEr optimization和top-down model的细节