Figure 5. Longer reaction times lead to faster learning and higher instantaneous reward rates.

(a) Schematic of experiment and hypothesized results. Previously trained animals were randomly divided into two groups: could only respond above (blue, $n=7$) or below (black, $n=5$) their individual mean reaction times for the previously trained stimulus and the new stimulus. Subjects responding above their individual mean reaction times were predicted to learn faster, reach a higher instantaneous reward rate sooner and accumulate more total reward. (b) Mean and individual reaction times before and after the reaction time restriction in rats. The mean reaction time for subjects randomly chosen to respond above their individual mean reaction times (blue, $n=7$) was not significantly different to those randomly chosen to respond below their individual means (black, $n=5$) before the restriction (Wilcoxon rank-sum test p > 0.05), but were significant after the restriction (Wilcoxon rank-sum test p < 0.05). Errors represent 95% confidence intervals. (c) In the model a long (blue) and a short (black) target decision time were set through a control feedback loop on the threshold, ${\displaystyle \frac{d}{dt}z(t)=\gamma (D{T}_{targ}-DT(t))}$ with parameter ${\displaystyle \gamma =0.01}$. (d) Mean accuracy ±95% confidence interval across sessions for rats required to respond above (blue, $n=7$) or below (black, $n=5$) their individual mean reaction times for a previously trained stimulus. Both groups had initial accuracy below chance because rats assume a response mapping based on an internal assessment of similarity of new stimuli to previously trained stimuli. To counteract this tendency and ensure learning, we chose the response mapping for new stimuli that contradicted the rats’ mapping assumption, having the effect of below-chance accuracy at first. * denotes p < 0.05 in two-sample independent $t$-test. Inset: accuracy change (slope of linear fit to accuracy across sessions to both groups, units: fraction per session). * denotes p < 0.05 in a Wilcoxon rank-sum test. (e) Mean inferred signal-to-noise ratio (SNR), (f) mean, $iRR$ and (g) mean cumulative reward across task engagement time for new stimulus pair for animals in each group. (h) Accuracy, (i) SNR, (j) $iRR$, and (k) cumulative reward across task engagement time for long (blue) and short (black) target decision times in the linear drift-diffusion model (LDDM).