Search

Choice deadlines are commonly imposed in decision-making research to incentivize speedy responses and sustained attention to the task settings. However, computational models of choice and response times routinely overlook this deadline, instead simply omitting trials past the deadline from further analysis. This choice is made under the implicit assumption that parameter estimation is not significantly affected by ignoring these omissions. Using new tools from likelihood-free inference, here we elucidate the degree to which omitting omissions, even in seemingly benign settings, can lead researchers astray. We explore the phenomenon using a Sequential Sampling Model (SSM) with collapsing boundaries as a test-bed.

Mice are increasingly used to study the neural circuitlevel basis of behavior, often with the ultimate goal to extrapolate these insights to humans. To generalize insights about neural functioning between species, it is crucial to first ensure correspondence in behavioral and cognitive strategy. Here, we analyzed decision-making behavior in both humans and mice, and identified the same cognitive strategy of history-dependent evidence accumulation. Specifically, individual differences in choice repetition were explained by a history dependent bias in the rate of evidence accumulation – rather than its starting point. Evidence integration over multiple temporal scales thus reflects a fundamental aspect of decision-making, conserved across mammalian species. These findings set the stage for linking the computations of decision-making to neural dynamics at the single-cell and population levels.

Computational cognitive modeling offers a principled inter-pretation of the functional demands of cognitive systems andaffords quantitative fits to behavioral/brain data. Typically,cognitive modelers are interested in the fit of a model withparameters estimated using maximum likelihood or Bayesianmethods. However, the set of models with known likeli-hoods is dramatically smaller than the set of plausible gen-erative models. For all but some standard models (e.g., thedrift-diffusion model), lack of closed-form likelihoods typi-cally prevents using traditional Bayesian inference methods.Employing likelihood-free methods is a workaround in prac-tice. However, the computational complexity of these methodsis a bottleneck, since it requires many simulations for each pro-posed parameter set in posterior sampling schemes. Here, wepropose a method that learns an approximate likelihood overthe parameter space of interest by encapsulation into a convo-lutional neural network, affording fast parallel posterior sam-pling downstream after a one-off simulation cost is incurredfor training.