Ctrl-TNDM: Decoding feedback-driven movement corrections from motor cortex neurons

Recent studies of motor control have shown that the neural population activity in motor cortical areas has a low-dimensional structure: a low number of latent dynamical factors that explain a large fraction of neural variability. It is unclear, however, whether this low-dimensional structure corresponds to task instructions or to behavioral output, since both variables are strongly correlated with one another.
To address this question, we analysed neural recordings from premotor and motor cortices of monkeys engaged in a center-out reaching task with a force field perturbation. The perturbation induced transient deviations from the desired trajectory which required correction in real-time.
While the instructed behavior (a reach towards the target) is known to be well captured by low-dimensional autonomous dynamical models, the uninstructed component includes online corrections to the hand trajectory that the monkey makes during each trial based on the visual and tactile feedback.
Using an RNN decoder, we confirmed that the information about uninstructed behavior can be decoded from neural activity. We next extracted latent dynamical factors using an unsupervised sequential autoencoder model with a controller RNN (LFADS) that can infer unobserved control inputs into the neuronal population. We found that the low-dimensional dynamical factors in LFADS represented instructed reach direction, but could not explain the uninstructed behavioral variability related to online movement corrections. We then modified the LFADS model following the ideas from the recently developed Targeted Neural Dynamical Modeling (TNDM) approach, which aims to align the latent dynamical factors with the observed behavior. Our new Ctrl-TNDM model captured almost all explainable behavioral variability, including uninstructed movement corrections.
This result suggests that the low-dimensional latent dynamics in neural activity can explain behavioral variability, but discovering the correct manifold and latent dynamics requires weak supervision with recorded behavior.