References
Chen, Y., Xie, Y., Song, L., Chen, F., & Tang, T. (2020). A Survey of Accelerator Architectures for
Deep Neural Networks. Engineering, 6(3), 264–274.
https://doi.org/https://doi.org/10.1016/j.eng.2020.01.007
DNN training dataflow in PipeLayer [22]. Each arrow represents a data dependency. T:
logical cycle; L: the ground-truth label; d: the feature map; A: array; W: weight;
δ: error; W′: the reformed W
DNN may include
● convolutional layers
● fully connected layers (batched matrix multiplications),
● point-wise operation layers such as ReLU, sigmoid, max pooling,
● batch normalization.
Forward propagation (inference)
Backward propagation (training)
Matrix multiplications become transposed weight matrix
convolutions become rotated convolutional kernel,