5/31(水)に第2回学生ゼミが開かれました。
引き続き,『深層学習 第二版』の輪読をすすめています.
今回は,『第3章確率的勾配降下法』の前半部分について,B4の嶋野さんから発表がありました.
(SGD・汎化性能と過剰適合・正則化・学習率の選定と制御など)
バッチ学習・ミニバッチ学習,モメンタム・ネステロフの加速勾配法,学習率の選定や制御などについて,議論の深ぼりもありました.
次週の学生ゼミでは,第3章の後半を扱います.
(SGDの改良・層出力の正規化・重みの初期化など)
——–
The second student seminar was held on May 31 (Wed.).
Following the 1st student seminar, we have been reading “Deep Learning, Second Edition” by turns.
This time, Mr. Shimano (B4) gave a presentation on the first half of “Chapter 3: Stochastic Gradient Descent Method.
(SGD, generalization performance and over-fitting, regularization, learning rate selection and control)
There was a deep discussion on batch learning, mini-batch learning, Momentum-Nesterov’s accelerated gradient method, learning rate selection and control, and so on.
Next week’s student seminar will cover the second half of Chapter 3.
(SGD improvement, normalization of layer outputs, initialization of weights, etc.)
投稿者 | 諸田