Orthogonal preparatory and movement subspaces in monkey, mouse, and model

T-C Kao, M Sadabadi, and G Hennequin
COSYNE, 2018  

Abstract


In what dynamical regime does the motor cortex operate? Population recordings in monkey motor cortex during delayed reaching tasks have revealed a surprising relationship between preparatory and movement-related activity trajectories: they unfold in orthogonal subspaces (Elsayed et al., 2016). Specifically, the activity subspace that captures most of the across-movement variance in neural activity during movement preparation captures little variance during movement execution – and vice versa.

To begin to understand whether this effect is a useful diagnostic feature of the collective dynamics of motor cortical areas and not a mere epiphenomenon, we examined the relationship between preparatory and movement-related activity in rodents during a delayed tactile discrimination task (data courtesy of Karel Svoboda). Mice were trained to lick left or lick right depending on the location of a pole presented during a brief «sampling» epoch. Mice had to wait for a certain duration before licking, presumably using this delay period to prepare their motor output. We analysed recordings made in the anterior lateral motor cortex (ALM) using the same population analysis techniques as previously applied to monkey data. We found the same effect: a strong decoupling of ALM delay and movement-related activities, in the eight mice that we analysed.

Next, we show that this effect is naturally captured by the dynamics of recurrent networks with strong and intricate excitatory (E) connections stabilised by detailed inhibitory (I) feedback. In previous work (Hennequin et al, 2014), we have shown that such networks produce naturalistic transients following initialisation in specific states characterized by broken E/I balance. Upon movement onset, activity rotates away from these «sensitive preparatory states» as it grows bigger in amplitude during the movement epoch, rapidly entering an orthogonal subspace where excitation and inhibition re-balance. This is a highly robust feature of this class of models.

Icons by