@ There is a bit of glossing over detail on part of the subject I see a number of confused people posting on Stack Overflow or Datascience Stack Exchange. Namely that you don't backpropagate the *error* value per se, but the gradient of the error with respect to a current parameter. This is made more confusing to many software devs implementing back propagation because usual design of neural nets is to cleverly combine the loss function and the output layer transform, so that the derivative is numerically equal to the error (specifically only at the pre-transform stage of the output layer). It really matters to understand the difference though because in the general case it is not true, and there are developers "cargo culting" in apparently magic manipulations of the error because they don't understand this small difference.(00:04:40 - 00:05:29)
Backpropagation in 5 Minutes (tutorial)
Siraj Raval
※本サイトに掲載されているチャンネル情報や動画情報はYouTube公式のAPIを使って取得・表示しています。
Timetable
動画タイムテーブル
動画数:471件