Accumulation noise is important to modeling depression since it serves as a protective factor against negative information processing biases. Assuming the network has been overtrained on some negative stimulus, the network, without noise may still be able to drift toward identification of the stimulus when it is present (i.e., before the stimulus duration) but once the stimulus is removed, affective accumulation tends to move toward the negative affective valence via recurrent activation in the affective-semantic loop. With no noise, the stimulus would often be evaluated as negative at that point, regardless of its true affective valence, because accumulation of that affective valence effectively ceases. With noise, accumulation still occurs on the nonactivated but correct affective valence. Often this noise, combined with the accumulation which had occurred during the stimulus presentation, is enough to create a post-stimulus-duration identification even after extensive training on stimuli with a negative affective valence.
One caveat to the use of accumulation noise is that it has the potential to dwarf small effects because of the variance it induces into responses. Simulations run without noise, not reported in this thesis, did conform to the human data even better than those with noise, though the variability in response times was much smaller.