www-ai.cs.tu-dortmund.de/LEHRE/FACHPROJEKT/WS1718/folien3.pdf
DeepLearning on FPGAs - Artifical Neuronal Networks: Image classification
be large enough to capture features, but small enough to be fast to compute. Usually we use 1× 1− 7× 7
Convolution tends to overfit, so regularization should be used
Deeper architectures usually perform [...] be large enough to capture features, but small enough to be fast to compute. Usually we use 1× 1− 7× 7
Convolution tends to overfit, so regularization should be used
Deeper architectures usually perform [...] = max(0,min(255, bkcc))
Thus: Assume appropriate mapping and use kc : Rr → R
DeepLearning on FPGAs 7
Image Representation: Making images smaller
Obviously: Images need to be smaller!
Merge a r × r grid …
