Fig. 3: Identifying the stable states during handwriting.
From: Human motor cortex encodes complex handwriting through a sequence of stable neural states
a, Neural activity recorded during attempted handwriting is divided into non-overlapped stable states, which are classified using the TFC algorithm. Each colour denotes a specific state. b,c, The model encoding loss decreases when more states are included, and the tipping point occurs around 10 states. A total of 18 sessions, each containing 30 characters, were included (threefold cross-validation, each fold containing 10 characters with 3 repetitions, with non-overlapped characters across folds, M ± s.d.). Both the encoding loss (b) and the loss decrease (c) are given (n = 54, 18 sessions each containing 3 folds). d, With a threshold of loss decrease <10−5, we found that the mean number of states was around 10 (mean = 9.732, n = 54). e, Evaluation with the Bayesian information criterion (BIC) also indicated an optimal model number around 10 (the lowest mean BIC was obtained at the state number of 12 and 8 for training and test data, respectively, n = 54, M ± s.d.). f, Pairwise neural activity encoding loss between states. Each matrix entry (i, j) indicates the prediction error of neural activity in state j for the model trained on state i. A model can only well predict the neural responses during the state it is trained on. g, Visualization of encoding models learned by TFC with three non-overlapping character sets. The first two principal components of the linear mapping matrix were plotted. The encoding functions were mostly similar despite different character sets. h, State prediction performance of diverse kinematic parameters and neurons. Neural signals can reliably predict the states while kinematic parameters cannot (threefold cross-validation with no overlapping characters, chance level = 0.1). All box plots depict the median (horizontal line inside the box), 25th and 75th percentiles (boxes), and minimum and maximum values (whiskers) Source data.
