Intrinsic Properties Ensure Reliable Attractor Dynamics in Learned Neural Assemblies Embedded Within Noisy, Asynchronous Networks

Neural representations rely on the ability of neuronal assemblies to display organized spiking patterns, despite being embedded within noisy networks. These structured patterns arise from attractor dynamics due to activity reverberation promoted by learnt assembly connectivity. Yet, attractor dynamics have been assessed either under low-noise conditions or in highly idealized neuronal assemblies. Here, in a spiking recurrent neural network model displaying asynchronous irregular noise, we show that realistic spike timing-dependent plasticity (STDP) imposes either low controllability of attractor recall (when STDP is strong), low stability of attractor maintenance (when STDP is low), or both. Moreover, STDP-built attractors display low independence, i.e., they perturb activity in the surrounding network. These constraints may favor self-generated representation switches essential for cognitive flexibility but dampening cognitive reliability. We reveal, by contrast, that several biophysical mechanisms alleviate these issues through a common dynamical principle, protecting excitatory-driven attractorial dynamics from inhibitory-driven spontaneous fluctuations. Amongst biophysical determinants, intrinsic properties were most efficient to increase controllability, stability and independence of attractor dynamics. Specifically, spike-triggered calcium-activated conductances improved reliability by mitigating reliance on connectivity, even at low conductance levels, i.e., in the absence of cell-autonomous intrinsic bistability. Finally, we show that the mechanisms we identify operate over a large repertoire of static (e.g., Hebbian or ring) and dynamic (e.g., sequences) attractors, with uni- and bidirectional propagation. Altogether, these results pinpoint synaptic and intrinsic synergy as a generic principle to regulate attractor reliability, as a function of the cognitive demand.