Let's consider a classic feedforward neural network $F$ with input dimension $d$, output dimension $k$, $L$ layers $l_i$ with $m$ neurons each. ReLu activation.

This means that, given a point $x \in R^d$ its image $F(x) \in R^k$. Let's now assume i add some gaussian noise $\eta_i$ in EVERY hidden layer $l_i(x)$ at the same time, where the norm of this noise is 5% the norm of its layer computed on the point $x$. Let's call this new neural network $F_*$

I know that, empirically, neural networks are resistant to this kind of noise, especially on the first layers. How can i show this theoretically?

The question i'm trying to answer is the following:

After having injected this noise $\eta_i$ in every layer $l_i(x)$, how far the output $F_{*}(x)$ will be from the output of the original neural network $F(x)$?


Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Browse other questions tagged or ask your own question.