Backpropを全部記述せよ
Lossをl, f, g で書き換える
線形で簡単に考える場合
ヤコビアンで考える
Taking advantage of the above, rewrite the gradient in a form that highlights the exponentially decaying or exploding effects known as gradient vanishing and gradient explosion. What condition on the eigenvalues of per-time step Jacobians is guaranteed to give rise to gradient vanishing?