Figure 4 : Training the system with varying numbers of examples.

From: Automated home-cage behavioural phenotyping of mice

Figure 4

For this leave-one-video-out experiment, the system accuracy is computed as a function of the amount of video data (in minutes/video) used for training the system. For each leave-one-out trial, the system is trained on a representative set of videos and tested on the full length of the video left out. A representative set consisting of x 1-min segments is manually selected such that the total time of each of the eight behaviours is roughly the same. (a) Average accuracy and standard error across the 12 leave-one-out runs on the 'full database'. (b) Average accuracy and standard error across the 12 runs on the 'full database'. (c) Average accuracy and standard error across the 13 leave-one-out runs on the 'wheel-interaction database'. We also perform the training/testing on the same video: a representative set of video segments is selected from the first 30 min of each video for training; testing is carried out on the remaining of the video from the 30th minute to the end of the same video. System accuracy is computed as a function of the amount of video data (in minute) used for training. (d) Average accuracy and standard error across the 13 runs on the 'wheel-interaction database'.