Trying to replicate (partially) the result of Feedback Alignment Algorithm[1].
re1.py
is for the simple linear network.
re2.py
is for the nolinear network on MNIST dataset.
check2.py
is for the evaluation of the network on MNIST dataset.
Result in figs/e1.png
seems great.
Performance of figs/e2.png
matches the result in the paper also, whereas the change of angle is not so close to that of the paper.
[1] T. P. Lillicrap, D. Cownden, D. B. Tweed, and C. J. Akerman, “Random synaptic feedback weights support error backpropagation for deep learning,” Nature Communications, vol. 7, pp. 1–10, 1AD.