Paper之DL之BP：《Understanding the difficulty of training deep feedforward neural networks》
sigmoid函数的test loss和training loss要经过很多轮数一直为0.5，后再有到0.1的差强人意的变化。
We hypothesize that this behavior is due to the combinationof random initialization and the fact that an hidden unitoutput of 0 corresponds to a saturated sigmoid. Note that deep networks with sigmoids but initialized from unsupervisedpre-training (e.g. from RBMs) do not suffer fromthis saturation behavior.
1、The normalization factor may therefore be important when initializing deep networks because of the multiplicative effect through layers, and we suggest the following initialization procedure to approximately satisfy our objectives of maintaining activation variances and back-propagated gradients variance as one moves up or down the network. We call it the normalized initialization
Activation values normalized histograms with hyperbolic tangent activation, with standard (top) vs normalized initialization (bottom). Top: 0-peak increases for higher layers.
Several conclusions can be drawn from these error curves:
(1)、The more classical neural networks with sigmoid or hyperbolic tangent units and standard initialization fare rather poorly, converging more slowly and apparently towards ultimately poorer local minima.
(2)、The softsign networks seem to be more robust to the initialization procedure than the tanh networks, presumably because of their gentler non-linearity.
(3)、For tanh networks, the proposed normalized initialization can be quite helpful, presumably because the layer-to-layer transformations maintain magnitudes of activations (flowing upward) and gradients (flowing backward).