Commit 41748b21 authored by David Peter's avatar David Peter
Browse files

Update citation

parent 72a72ec0
No preview for this file type
......@@ -492,4 +492,12 @@
journal = {Coursera, video lectures},
title = {Neural networks for machine learning},
year = {2012},
}
\ No newline at end of file
}
@article{Roth2020,
title={{R}esource-{E}fficient {N}eural {N}etworks for {E}mbedded {S}ystems},
author={Wolfgang Roth and G{\"u}nther Schindler and Matthias Zohrer and Lukas Pfeifenberger and Robert Peharz and Sebastian Tschiatschek and Holger Froning and F. Pernkopf and Zoubin Ghahramani},
journal={ArXiv},
year={2020},
volume={abs/2001.03048}
}
......@@ -30,7 +30,7 @@ Figure~\ref{fig:ste_quant} illustrates weight quantization using the \gls{ste} o
\begin{figure}
\centerline{\includegraphics[width=0.7\linewidth]{\pwd/plots/ste.pdf}}
\caption{Weight quantization using the STE on a typical convolutional layer (without batch normalization). Red boxes have zero gradients whereas green boxes have non-zero gradients. Weight updates are performed to the blue circle. During the forward pass, the weight tensor $\mathbf{W}^l$ is quantized to obtain the $k$ bit weight tensor $\mathbf{W}^l_q$ used in the convolution. The activation function is then applied to the output of the convolution $\mathbf{a}^{l+1}$ to obtain $\mathbf{x}^{l+1}$ which is the input tensor of the subsequent layer. During the backward pass, the \gls{ste} replaces the derivative of the quantizer by the derivative of the identity function \textit{id}. Figure from \cite{Roth19}.}
\caption{Weight quantization using the STE on a typical convolutional layer (without batch normalization). Red boxes have zero gradients whereas green boxes have non-zero gradients. Weight updates are performed to the blue circle. During the forward pass, the weight tensor $\mathbf{W}^l$ is quantized to obtain the $k$ bit weight tensor $\mathbf{W}^l_q$ used in the convolution. The activation function is then applied to the output of the convolution $\mathbf{a}^{l+1}$ to obtain $\mathbf{x}^{l+1}$ which is the input tensor of the subsequent layer. During the backward pass, the \gls{ste} replaces the derivative of the quantizer by the derivative of the identity function \textit{id}. Figure from \cite{Roth2020}.}
\label{fig:ste_quant}
\end{figure}
......
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment