Fig. 4: Spiking neural networks to decode phantom movements from the intraneural recordings.
From: Decoding phantom limb movements from intraneural recordings
A L is the loss (e.g., cross-entropy between membrane potentials and target class), η is the learning rate, ut is the membrane potential at time t, ϑ is the spike threshold, and H is the Heaviside function. B Comparison of the SNN-based decoders’ accuracy in both participants. Performance using the signal standard encodingr, threshold spikes and double encoding are reported. Two-sided Paired-t-tests *p < 0.05 (Signal enc-thr spikes, p = 0.003 for S1, p = 0.03 for S2. Double enc-thr spikes, p = 0.0012 for S1, p = 0.001 for S2. Signal enc- double enc, p = 0.18 for S1, p = 0.06 for S2. n = 540 for S1 and n = 720 for S2 (5-fold cross validation). C Comparison of the decoders’ accuracy in both participants. SNN spiking neural network, SVM support vector machine, MLP multilayer perceptron. Error bars indicate mean and STD. Chance levels indicated with dashed lines. Two-sided paired-t-tests *p < 0.05 (p = 0.04 for SNN vs. MLP, p = 0.0029 for SNN vs. SVM for S1; p = 0.006 for SNN vs. MLP, p = 0.018 for SNN vs. SVM for S2. SVM vs. MLP p = 0.44 for S1, p = 0.55 for S2). n = 540 for S1 and n = 720 for S2 (5-fold cross validation). Confusion matrix of the SNN performance, using the LIF encoding in S1 for all movement types.
