Figure 6. Population decoding analysis of action recognition task (ART) vs. trajectory decoding task (TDT) models.

(A) Population decoding of speed (light) and direction (dark) for spatial-temporal models for the ART- and TDT-trained models. The faint line shows the $R^2$ score for an individual model; the dark one the mean over all instantiations ($N=5$). (B) Population decoding of end-effector position (X and Y coordinates) for spatial-temporal models. The faint line shows the $R^2$ score for an individual model; the dark one the mean over all instantiations ($N=5$). (C) Same as (A) but for spatiotemporal models. (D) Same as (B) but for spatiotemporal models. (E) Same as (A) but for long short-term memory (LSTM) models. (F) Same as (B) but for LSTM models.

Figure 6—figure supplement 1. Analysis of population decoding for action recognition task (ART)-trained and untrained models.

(A) Population decoding of speed (light) and direction (dark) for the ART-trained and untrained for spatial-temporal models (left), spatiotemporal (middle), and long short-term memory (LSTM) (right) models. The faint line shows the $R^2$ score for an individual model; the dark one the mean over all instantiations ($N=5$). (B) Population decoding of end-effector position (X and Y coordinates) for spatial-temporal models. The faint line shows the $R^2$ score for an individual model; the dark one the mean over all instantiations ($N=5$).

Figure 6—figure supplement 2. Analysis of population decoding for trajectory decoding task (TDT)-trained and untrained models.

(A) Population decoding of speed (light) and direction (dark) for the TDT-trained and untrained for spatial-temporal models (left), spatiotemporal (middle), and long short-term memory (LSTM) (right) models. The faint line shows the $R^2$ score for an individual model; the dark one the mean over all instantiations ($N=5$). (B) Population decoding of end-effector position (X and Y coordinates) for spatial-temporal models (left), spatiotemporal (middle), and LSTM (right) models. The faint line shows the $R^2$ score for an individual model; the dark one the mean over all instantiations ($N=5$).

Figure 6—figure supplement 3. Results for the position and velocity trajectory decoding task (TDT-PV).

(A) For an example instantiation, the fraction of neurons that are tuned for a particular feature (${\displaystyle R^2>0.2}$ on the relevant encoding model). Model architectures: (left) spatial-temporal, (middle) spatiotemporal, (right) long short-term memory (LSTM). Tested features were direction tuning, speed tuning, velocity tuning, Cartesian and polar position tuning, acceleration tuning, and label tuning (328/5446 scores excluded for TDT-PV-trained spatial-temporal model, 140/3262 for spatiotemporal, and 1150/9142 for LSTM; see Methods). (B) The means of 90% quantiles over all five model instantiations of models trained on action recognition task (ART) and TDT-PV are shown for direction tuning (dark) and position tuning (light). 95% confidence intervals are shown over instantiations ($N=5$). (C) Population decoding of direction (dashed) and Cartesian coordinates (solid; mean over individually computed scores for X and Y directions taken) for the ART-trained and TDT-PV-trained for spatial-temporal models (left), spatiotemporal (middle), and LSTM (right) models. The faint line shows the $R^2$ score for an individual model; the dark one the mean over all instantiations ($N=5$). (D) For quantifying uniformity, we calculated the total absolute deviation from the corresponding uniform distribution over the bins in the histogram (red line in inset) for the spatial-temporal model (left), the spatiotemporal model (middle), and the LSTM model (right). Normalized absolute deviation from uniform distribution for preferred directions per instantiation is shown ($N=5$, faint lines) for TDT-PV-trained and untrained models as well as mean and 95% confidence intervals over instantiations (solid line; $N=5$).