www-ai.cs.tu-dortmund.de/LEHRE/FACHPROJEKT/WS1617/folien4.pdf
DeepLearning on FPGAs - Artifical Neuronal Networks: Image classification
remarks So far: We assumed 1 color channel - what about 3 channels? Idea 1: Merge color channels into single value
Average: (ri,j + gi,j + bi,j) /3
Lightness: (max (ri,j , gi,j , bi,j)−min (ri,j , gi,j [...] should be large enough to capture features, but small enough to be fast to compute. Usually we use 3× 3− 7× 7
Convolution tends to overfit, so regularization should be used
Deeper architectures usually [...] j
( 1− f (L)j
)
derivative of activation function
derivative of loss function
DeepLearning on FPGAs 3
Backpropagation for sigmoid activation / RMSE loss
Gradient step:
w (l) i,j = w
(l) i,j − α · δ
(l) …
