The neocortex is composed of spiking neuronal units interconnected in a sparse, recurrent network. Neuronal networks exhibit spiking activity that transforms sensory inputs into appropriate behavioral outputs. In this study, we train biologically realistic spiking neural network (SNN) models to identify the architectural changes which enable task-appropriate computations. Specifically, we employ a binary state change detection task, where each state is defined by motion energy. This task mirrors behavioral paradigms that mice perform in the lab. SNNs are composed of excitatory and inhibitory units randomly interconnected with connection likelihoods and strengths matched to observations from mouse neocortex. Following training, we discover that SNNs selectively adjust firing rates depending on state, and that excitatory and inhibitory connectivity between input and recurrent layers change in accordance with this rate modulation. Input channels that exhibit bias to one specific motion energy input develop stronger connections to recurrent excitatory units during training, while channels that exhibit bias to the other input develop stronger connections to inhibitory units. Furthermore, recurrent inhibitory units which were tuned to one input strengthened their connections to recurrent units of the opposite tuning. The convergence of trained network models on the specific pattern of cross-tuned inhibition serves as further validation of the fundamental role played by interneurons and their connections in shaping dynamics and information processing within the neocortical circuits.

The stability over time of the relationship between neural activity and a given external variable remains a matter of dispute in many brain areas. Whether neural representations are stable or changing (“drifting”) has massive implications for *how* those representations could be read out by downstream neural populations and incorporated into neural computations driving perception, behavior, and memory. Notably, studies of stability of neural coding for movements have generally focused on stereotyped, overtrained behaviors. The study of behaviors which exhibit higher trial-to-trial variance may enhance our understanding of the neural control of naturalistic and unconstrained behaviors. Such behaviors also allow us to move beyond trial-averaged analyses to single-trial analyses, such as instantaneous decoding, to quantify the stability of neural coding for a wider range of kinematic variables. Moreover, it is far from clear whether the laminar position of a motor cortical neuron corresponds to the stability of the single-trial mapping between its activity and kinematic variables over days.  We have collected an extensive neural and behavioral dataset spanning 24 consecutive days while mice learn a difficult skilled reach-to-grasp task which minimizes automaticity. We adapted the Whishaw single-pellet reaching task to be freely moving and self-initiated, and increased task difficulty by requiring greater precision in paw placement and spatiotemporally coordinated digit control. These steps evoked high variance in forelimb, paw, and digit movements and minimize automaticity even after the task is well-learned. We study representational drift in motor cortical coding using fine-grained behavioral descriptions of paw and digit kinematics  and the calcium fluorescence activity of large populations of neurons (~300 per field of view) recorded across cortical layers and tracked across days in unrestrained mice. Encoding models of single trials will enable insights into trial-to-trial and day-to-day stability/drift in the motor cortical representation of a skilled motor behavior.