“…Actions may be generated by an inherently noisy system: synapses are unreliable ( Allen and Stevens, 1994 ), neurons generate variable spike trains ( Mainen and Sejnowski, 1995 ), and neural circuits may operate near the edge of chaos ( van Vreeswijk and Sompolinsky, 1996 ). On the other hand, neural networks, whether adaptive or hard-wired, have structures that shape population neural dynamics onto a low dimensional manifold, where nonrandom and ordered activity patterns emerge ( Ganguli et al, 2008 ; Harvey et al, 2012 ; Inagaki et al, 2018 ). Computational models have promised to provide a unified view of these observations ( Burak and Fiete, 2012 ; Brennan and Proekt, 2019 ; Mastrogiuseppe and Ostojic, 2018 ; Roberts et al, 2016 ), but a deep connection between theories and experiments remains to be established.…”