Neuron activities steer attractors approximation during memory processing in mice

Biological brains contain billions of neurons to carry out complex cognitive functions such as memory and decision making. Such system cannot be understood through the conventional “labeled line” model in neuroscience. To gain better understanding, we construct a general dynamic system framework in which the cognitive variables are represented in low-dimensional latent spaces. The low-dimensional nature of neuro-computation predicts that random noise in the system can propagate into the coding direction and affects the duration of information maintenance, which we demonstrated via large scale in vivo recordings and behavior experiments. Furthermore, the framework accurately estimates the neurons preferentially contributing to certain latent variable, opening the possibility for experimental manipulation. Importantly, in mice carrying out long-term or short-term memory tasks, we found that memory-related activities are organized into point attractor structures in the latent space. Memory processing is associated with activity patterns switching between attractor states. This switching can be mediated via intermediate slow manifolds, or external input that drives the approximation between attractors without modifying the connectivity of the network. Blockage of the state switching leads to failure in memory processing. Together, these data support the latent dimension computation as a governing principle in biological brains, which may have implications in accessing effective manipulations in biological and artificial neuro-networks.