In slide #25~26, Linear alignment -> Feedback alignment
Presentation for ICML2019 reading pitch @ Kyoto 4th August 2019. Shuntaro Ohno introduced "Training Neural Networks with Local Error Signals" in Japanese.
25. さらに plausible に(pred)
H
pred-bpf と呼称
B (固定)
[Lillicrap et al. 2016] Lillicrap TP, Cownden D, Tweed DB, and Akerman CJ. Random synaptic feedback weights support
error backpropagation for deep learning. Nature Communications, 7:13276, 2016.
[Lillicrap et al. 2016 Fig. 1b&c]
Linear alignment に置換
37. CIFAR-100
50,000 training images (32x32), class: 100
Note that the CIFAR100 runs are less comparable to each other, because the sim and predsim runs had batches sampled
to have only 20 classes per batch during training, which we found to cause a higher training error, but lower test error.
WRN: WideResNet-40-10 → baseline
45. Although we did not get airplanes
from imitating birds
but from understanding the principles of
aerodynamics,
researching birds advances
aerodynamics.
46. 参考文献
1. Nøkland A and Eidnes LH. Training neural networks with local
error signals. arXiv, 1901.06656.
2. Bengio Y, Lee D, Bornschein J, Mesnard T, and Lin Z. Towards
biologically plausible deep learning. CoRR, abs/1407.7906,
2014. URL http://arxiv.org/abs/1812.11446.
3. Lillicrap TP, Cownden D, Tweed DB, and Akerman CJ. Random
synaptic feedback weights support error backpropagation for
deep learning. Nature Communications, 7:13276, 2016.
4. Sacramento J, Costa RP, Bengio Y, and Senn W. Dendritic
cortical microcircuits approximate the backpropagation
algorithm. CoRR, abs/1810.11393, 2018. URL http://dblp.uni-
trier.de/db/journals/corr/corr1810.html #abs-1810-11393.
5. Gadagkar V, Puzerey PA, Chen R, Baird-Daniel E, Farhang AR,
and Goldberg JH. Science, 354:1278-1282, 2016.
6. Moskovitz TH, Litwin-Kumar A, and Abbott L. Feedback
alignment in deep convolutional networks. CoRR, 12 2018.