Figure - PMC

Skip to main content

An official website of the United States government

Here's how you know

Here's how you know

Official websites use .gov
A .gov website belongs to an official government organization in the United States.

Secure .gov websites use HTTPS
A lock ( ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites.

View full-text article in PMC

. 2019 Mar 2;68(6):876–895. doi: 10.1093/sysbio/syz014

Search in PMC
Search in PubMed
View in NLM Catalog
Add to search

© The Author(s) 2019. Published by Oxford University Press, on behalf of the Society of Systematic Biologists.

This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.

PMC Copyright notice

Figure 5. — Impact of preserving more features during feature extraction on identification accuracy for data sets D1 (a), D2 (b), D3 (c), and D4 (d). We used input images of size in all cases, and we pooled features from each of the last three convolutional blocks (c3–c5) into matrices of dimension , where F is the number of filters of the corresponding convolutional block and N was set to 1 (global average pooling), 2, 4, 7, 14, or 28. The dashed black line indicates the best performing model from previous experiments (Fig. 4b). For D2 and D4 we stopped at N = 14 because of the computational cost involved in computing the value for N = 28, and since the accuracy had already dropped significantly at N = 14. Because of the computational cost, we computed performance of fused feature matrices only up to N = 7 (D1 and D3) or N = 4 (D2 and D4). On data sets D1, D2, and D3, the optimum identification performance was somewhere between globally pooled and nonpooled features. For D4, global pooling yielded the best identification performance.

Inline graphic — Impact of preserving more features during feature extraction on identification accuracy for data sets D1 (a), D2 (b), D3 (c), and D4 (d). We used input images of size in all cases, and we pooled features from each of the last three convolutional blocks (c3–c5) into matrices of dimension , where F is the number of filters of the corresponding convolutional block and N was set to 1 (global average pooling), 2, 4, 7, 14, or 28. The dashed black line indicates the best performing model from previous experiments (Fig. 4b). For D2 and D4 we stopped at N = 14 because of the computational cost involved in computing the value for N = 28, and since the accuracy had already dropped significantly at N = 14. Because of the computational cost, we computed performance of fused feature matrices only up to N = 7 (D1 and D3) or N = 4 (D2 and D4). On data sets D1, D2, and D3, the optimum identification performance was somewhere between globally pooled and nonpooled features. For D4, global pooling yielded the best identification performance.