Fig. 9.

Comparison of prediction results between the deep models for multi-task learning with partial transfer learning and the sparse coding features for the 10 terms in stages 13–17. The x-axis shows the 10 terms. The y-axis corresponds to a subset of 50 images in stages 13–17 with the largest numbers of annotated terms. The gene names and the FlyExpress image IDs in parentheses are displayed. The prediction results of different methods compared with the ground truth are distinguished by different colors. The white entries correspond to predictions agreed upon by these two methods, while non-white entries were used to denote different types of disagreements. Specifically, the green and blue entries correspond to correct predictions by the multi-task learning features but incorrect predictions by the sparse coding features. Green and blue indicate positive and negative samples, respectively, in the ground truth. Similarly, the red and pink entries correspond to incorrect predictions by the multi-task learning features but correct predictions by the sparse coding features. Red and pink indicate positive and negative samples, respectively, in the ground truth.