Fig. 5: Unsupervised pretraining accelerates subsequent task learning.
From: Unsupervised pretraining in biological neural networks
a, We trained three new cohorts of mice with or without a pretraining step of running through the virtual reality corridor without rewards for 10 days. All cohorts were trained in the task for 5 days. VRg, virtual reality gratings; VRn, virtual reality naturalistic stimuli. b, Task structure (sound cue was removed and rewards were delivered deterministically in the second half of the reward corridor). Mice had to lick to obtain the reward, except on the first day when rewards were delivered passively at the end of the corridor. Each mouse was trained on a separate combination of wall textures from among four stimuli (Extended Data Fig. 1a). c, Licking of example mice from each cohort on the first day of training with active rewards (grey background indicates grey corridor). d, Same as panel c but for the last day of training. e, Average lick responses across days for each cohort of mice (n = 11 VRn pretrained mice, n = 7 VRg pretrained mice and n = 5 no pretrained mice). f, Performance summary (difference in lick responses) across days for each cohort in panel e. g, Distribution of first licks across days for each cohort in panel e. h, Number of trials per day for each cohort in panel e. All data are mean ± s.e.m. *P < 0.05 and **P < 0.01 from two-sided t-tests.
