Fig. 5. RNNs can capture main behavioral results.

a Plotted are average estimates at the end of a simulated session of our learning task for N = 50 instances of the simulated RNNs and the average value of participants’ reward estimate vs. actual reward probabilities associated with each stimulus (each symbol represents one stimulus). Error bars represent s.e.m., and the dashed line is the identity. b The plot shows the time course of explained variance (R²) in RNNs’ estimates based on different GLMs. Error bars represent s.e.m. The solid line is the average of exponential fits to RNNs’ data, and the shaded areas indicate ±s.e.m. of the fit. c Time course of adopted learning strategies measured by fitting the RNNs’ output using a stepwise GLM. Plotted is the normalized weight of the informative feature, informative conjunction, and object (stimulus identity) on reward probability estimates. Error bars represent s.e.m. The solid line is the average of exponential fits to RNNs’ data, and the shaded areas indicate ±s.e.m. of the fit. d Plot shows differential response for the informative feature vs. differential response for the non-informative features in the trained RNNs. The inset shows the histogram of the difference between differential response of the informative and non-informative features. The dashed line shows the median values across all trained RNNs, and an asterisk indicates the median is significantly different from 0 (two-sided sign-rank test; P = 0.013). e Similar to d but for the informative and non-informative conjunctions (two-sided sign-rank test; P = 1.4 × 10⁻⁶). f Plot shows the difference between differential response for the informative and non-informative features vs. the difference between differential response for the informative and non-informative conjunctions. The inset shows the histogram of the difference between the aforementioned differences (two-sided sign-rank test; P = 0.018). Source data are provided as a Source Data file.