Learning from Data (Fall 2022)

[kundtx]

http://8.129.175.102/lfd2022fall-poster-session/33.html

[Prof-Greatfellow]

G1 Haizhou Liu: Good job on the flower classification project. A few questions: 1) how large is the dataset? If it it is collected by a team of 3, I would estimate <1000, which might not be sufficient for CV problems? 2) Since ANNs (DenseNet) seem to be the "only" accurate approach, it might be better to focus more on the ANN family rather than the SVM family?

[Xinzhe-Ni]

@Prof-Greatfellow G1 Haizhou Liu: Good job on the flower classification project. A few questions: 1) how large is the dataset? If it it is collected by a team of 3, I would estimate <1000, which might not be sufficient for CV problems? 2) Since ANNs (DenseNet) seem to be the "only" accurate approach, it might be better to focus more on the ANN family rather than the SVM family?

G33 Xinzhe Ni: Thanks for the questions. (1) We also ask some friends for help, so we have more than 1000 images each class. It is enough for classification in our views. (2) It is right that we should focus on more DL methods. However, because of the COVID-19, we cannot complete our whole plan. We will try other DL methods in the future. Thanks again!

[HuazhangYing]

Can you briefly describe the differences between TSNE, PCA, and ISOMAP in dimensionality reduction results?

[Yiwendyh]

@HuazhangYing Can you briefly describe the differences between TSNE, PCA, and ISOMAP in dimensionality reduction results?

G33 Yiheng Dai: After the dimensionality reduction of PCA method, the distance within classes is small, and the distance between classes is large. Although the tSNE method has a small distance within a class, the distance between classes is also small Very small; Although the ISOMAP method has a large distance between classes, the distance within a class is also large. Therefore, in the case of different concerns, all three algorithms are Each has its own advantages.

[Xinzhe-Ni]

@HuazhangYing Can you briefly describe the differences between TSNE, PCA, and ISOMAP in dimensionality reduction results?

Thanks for the question. Through our three visualization results, tSNE, PCA and ISOMAP are all not proper for our dataset. It is also because the selected numbers are too big (i.e. 500 each class). However, we can also get some meaningful results especially with ISOMAP which we do not put it on the poster. For example, we can know which two classes are easy to discriminate and which two classes are hard. Thanks again!