Skip to main content
. 2022 Apr 28;16:858377. doi: 10.3389/fnins.2022.858377

FIGURE 4.

FIGURE 4

(A) The mean accuracy for the three uNN decoders, replotted from Figure 3B with x-axis ticks realigned to match (B). (B) The computational requirements of the uNN decoders as measured by the number of batches required for training after disruptions are introduced. When zero channels are affected, no masking or retraining takes place, and the model would only receive unsupervised updates. The number of batches required is according to an early stopping criterion and is averaged over each of the 10 random initializations applied. Error bars represent 95% confidence intervals. (C) The number of batches required as a function of decoder accuracy.