The model executes a sequence of two reaches using a holistic strategy.
(A) Hand velocity during one of the reaches, with the corresponding hand trajectory shown in the inset. (B-C) We identified two 6-dimensional orthogonal sub-spaces, capturing 79% and 85% of total activity variance during single-reach preparation and movement respectively. (B) First principal component of the model activity for the 8 different reaches projected into the subspaces identified using preparatory (top) and movement-epoch (bottom) activity. (C) Occupancy (total variance captured across movements) of the orthogonalized preparatory and movement subspaces, in the model (top) and in monkey motor cortical activity (bottom; reproduced from Lara et al., 2018 for monkey Ax). We report mean ± s.e.m., where the error is computed by bootstrapping from the neural population as in Lara et al. (2018). We normalize each curve separately to have a maximum mean value of 1. To align the model and monkey temporally, we re-defined the model’s ‘movement onset’ time to be 120 ms after the model’s hand velocity crossed a threshold – this accounts for cortico-spinal delays and muscle inertia in the monkey. Consistent with Lara et al. (2018)’s monkey M1 recordings, preparatory subspace occupancy in the model peaks shortly before movement onset, rapidly dropping thereafter to give way to pronounced occupancy of the movement subspace. Conversely, there is little movement subspace occupancy during preparation. (D) Behavioural (top) and neural (middle) correlates of the delayed reach for one example of a double reach with an enforced pause of 0.6 s. The optimal strategy relies on preparatory inputs preceding each movement. (E) Same as (C), for double reaches. The onsets of the monkey’s two reaches are separately aligned to the model’s using the same convention as in (C). The preparatory subspace displays two clear peaks of occupancy. This double occupancy peak is also observed in monkey neural activity (bottom; reproduced from Zimnik and Churchland, 2021, with the first occupancy peak aligned to that of the model). (F) Same as (D), for compound reaches with no enforced pause in between. Even though the sequence could be viewed as a single long movement, the control strategy relies on two periods of preparation. Here, inputs before the second reach are used to reinject energy into the system after slowing down at the end of the first reach. (G) Even though no explicit delay period is enforced in-between reaches during the compound movement, the preparatory occupancy rises twice, before the first reach and once again before the second reach. This is similar to observations in neural data (bottom; reproduced from Zimnik and Churchland, 2021).