Activation functions filter internal Gaussian noise more effectively when injected before them, yielding higher accuracy especially with additive noise, while noise after activation accumulates and degrades performance more in earlier layers.
The noise influences were intro- duced separately into the 2nd (blue curves), 3rd (orange curves) and 4th layer (green curves)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.NE 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Internal noise in deep neural networks: interplay of depth, neuron number, and noise injection step
Activation functions filter internal Gaussian noise more effectively when injected before them, yielding higher accuracy especially with additive noise, while noise after activation accumulates and degrades performance more in earlier layers.