Extended Data Fig. 2. Decoding algorithm.

The input Y_IN is E_N x 3 input matrix, where E_N is the number of electrodes (192) and 3 represents the most recent 3 50-ms bins. The output variable, $\hat{\mathit{v}}$, represents a normalized vector of each of d finger velocities. The actual decoded velocities were calculated by applying an empirically calculated mean value and gain value. Linear layers (W_T, W₁-W₃) included a learnable bias term except for the final linear layer, W₄, to reduce the magnitude of non-zero means. All instances of batchnorm, BN_A, were implemented with affine = True except for the final batchnorm, BN, where affine = False in an attempt to reduce the reliance of the decoding algorithm on an offset correction from the final batchnorm block. During training mode batchnorm layer, BN, did not correct for non-zero means or apply a mean correction to force the final linear layer, W₄, to learn an output signal with zero mean. BN, batchnorm; FC, fully connected; ReLU, rectified linear unit. Figure adapted from (Willsey et al. 2022)²³.