Abstract
We recently demonstrated that the dorsal premotor cortex (PMd) of macaque monkeys houses mirror neurons (MirNs). MirNs, originally described in the ventral premotor cortex, are sensorimotor neurons that fire both when an animal performs an action and when the same animal observes another agent performing an identical or similar action. Here, we investigate whether a decoder trained on the activity induced during action-observation can reliably distinguish hand grips using the activity induced during action-execution. To this end, we use neuronal spiking cortical activity recorded from the PMd of macaques engaged in execution and observation of reaching-to-grasp actions. Decoding of actions is done by a Support Vector Machine with a linear kernel function using the activity occurred in 200 ms periods (starting 800 ms before and ending 1400 ms after movement onset) at various temporal combinations between observation and execution. Classification performance, using either the whole set of units (n=140) or subsets of units of various sizes, was at 50% regardless of the number of included units. To overcome this issue, a greedy-approach was employed: the unit with the best performance was initially selected and at each subsequent step a unit was added to those of the previous step so that to maximize the performance of the resulting population. Following this approach we reached cross-decoding accuracy above 90% using less than half of the population. Thus, a selected subset of units, that shares the same computations in the execution and observation, is sufficient for effective and consistent inter-condition decoding of grips. These results demonstrate that PMd-MirNs can provide a reliable signal for the development of grip decoders which could be trained even when the required registration of neural activity and behavior can’t be obtained due to the incapability of generating overt movements.