上午有AY关于"计算机视觉和深度学习综述性工作介绍"的talk,简单记录如下,有时间补充。
总体感受 :公司里比在学校小实验室瞎搞学得多得多
主题:Computational Cognition, Vision and learning & deep nets
印象最深的 The size of the dataset depends on the complexity of the task, the variability of the images and the classes.
一些概念记下来,需要后期补充:
1. Gauss 1800
2. conceptual model
3. fully conv deep net (Deeplab)
4. fully / weakly supervised
一些领域的介绍:
1. Occulusion detection
概念记录:
(1) occulusion boundary detect & bordeer-ownship
(2) 数据:Pascal
2. Human pose detection & semantic segmentation
(1) fully-connected CRF (conditional random fields)
3. Detecting symmetry axes
(1) Wei Shen st al., 2018
4. Siamese-Triplet Net: to learn similarity
(1) Tai Sing Lee collaboration, ICLR 2017
(2) 用于transfer learning
5. Text-captioning by m-RNN/LSTM
(1) Junhua Mao, ICLR 2015
(2) 介绍:
a. 从图中自动生成语言描述;
b. Text: m-RNN / image:CNN;
c. 结构大概是:
embedding 1-> embedding 2 -> RNN-> multimodel(这里有引入CNN)-> Softmax
(3) 会记住一些经常一起出现的词,举例:长颈鹿和树
(4) 现在这个task还是很难: 数据获取难/目前只在部分数据集上做成,迁移到别的数据集就不行
6. FELIX project:CT cancer detect
医疗数据,数据集一般都比较小
(1) multi-organ segmentation
(2) Dice similarity coefficient
(3) PDAL detection/segmentation
transfer learning
背景: 没有足够的annotated training data -> transfer learning
三个步骤:
1. 找closed problem B with enough training data
2. deep net on problem B
3. fine tune for A
举个例子:医疗上pain assessment from faces, 借用 face recognition
Feng Wang , ICIP 2017
一些记录:
(1) make this as a regression task
(2)CNN-> dense layer -> regression loss/ center loss 这两个loss分别代表什么?
(3)avoid overfitting: dense layer 神经元少一点
few shot learning
(1) key idea: relate the activations to the weights
(2) learn the mapping
这篇是乔思远的论文,本科同学,赞叹大神。
- deep net and random forest