Thunderstruck: The ACDC model of flexible sequences and rhythms in recurrent neural circuits
Fig 8
A. Number of actions within a sequence as a function of feedforward and recurrent weights. The y- and x-axis represent G-A and recurrent weight values, respectively. As depicted, the G-A weight values control whether the model can produce six actions within a 1 sec temporal window, irrespective of the recurrent weight values. The color bar codes the number of actions that are produced within the sequence; yellow for a full sequence (six actions) and dark blue for no actions. B. Action sustainability. The heatmap reflects the sustainability measure magnitude (warm colors coding for higher values) as a function of A-N and G-A weight values. As depicted, action sustainability increases with decreasing values of A-N weights, over a large range of G-A weight values. C. Model regime as a function of the γE and γI parameters. The y- and x-axis represent γE and γI parameter values, respectively. As depicted, the model can reproduce fully and precisely a learnt action sequence within a broad range of parameter combination respecting the γE > γI inequality. Color bar is identical to A. D. RNN input overlap. The model can produce a full six action sequence up to 35% of input overlap to the RNN between contiguous actions; i.e. activating 35% of the previous and subsequent cluster in the RNN. Stronger overlap leads to a break in the sequence after 2 actions. Color bar is identical to A and C.