WEKO3
アイテム
Numerical Simulations of Optoelectronic Deep Neural Network Using Trainable Activation Function
http://hdl.handle.net/10228/0002001767
http://hdl.handle.net/10228/0002001767e05f245d-919f-4e61-8bb4-1a23b476be1a
| 名前 / ファイル | ライセンス | アクション |
|---|---|---|
neuro_108.pdf (886.9 KB)
|
|
| アイテムタイプ | 共通アイテムタイプ(1) | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 公開日 | 2025-07-11 | |||||||||||||
| タイトル | ||||||||||||||
| タイトル | Numerical Simulations of Optoelectronic Deep Neural Network Using Trainable Activation Function | |||||||||||||
| 言語 | en | |||||||||||||
| 著者 |
Takatsu, Taichi
× Takatsu, Taichi
× Tomioka, Rio
× 高林, 正典
WEKO
35482
|
|||||||||||||
| 抄録 | ||||||||||||||
| 内容記述タイプ | Abstract | |||||||||||||
| 内容記述 | Optoelectronic deep neural network (OE-DNN) is one implementation of neural network (NN) hardware in which an electronic computing is combined with an optical computing to achieve a large-scale and flexible NNs. The optical computing part realizes the inter-node interactions with high spatial parallelism using spatial light modulator (SLM) and image sensor, whereas the electronic computing part realizes the other computing of NNs which are difficult to realize optically such as applying nonlinear activation function and normalization. However, an activation function commonly used in the conventional NNs such as rectified linear unit (ReLU) and hyperbolic tangent (tanh) functions sometimes decrease the accuracy of NNs when the input values to such activation functions are small and/or non-negative values which often occur in OE-DNNs. Therefore, the activation function needs to be specialized for OE-DNNs. In this paper, we propose to apply a trainable activation function [1], which can optimize the shape of the activation function for input values. Specifically, the coefficients involved in the activation function are included in the training process. We numerically compare the classification accuracy of input images by the model using activation function with and without trainable parameters. It should be noted that only the results of the image classification using a complex-valued OE-DNN are shown in this paper, though the application of the trainable activation function may not be limited to the complex-valued NNs. |
|||||||||||||
| 言語 | en | |||||||||||||
| 備考 | ||||||||||||||
| 内容記述タイプ | Other | |||||||||||||
| 内容記述 | International Symposium on Imaging, Sensing, and Optical Memory 2024, ISOM’24, October 20-23, 2024, Arcrea HIMEJI, Himeji, Hyogo, Japan | |||||||||||||
| 言語 | en | |||||||||||||
| 書誌情報 |
en : International Symposium on Imaging, Sensing and Optical Memory (ISOM '24) Technical Digest p. Tu-E-44, 発行日 2024-10 |
|||||||||||||
| 出版社 | ||||||||||||||
| 出版者 | 日本光学会 | |||||||||||||
| 言語 | ja | |||||||||||||
| 言語 | ||||||||||||||
| 言語 | eng | |||||||||||||
| 資源タイプ | ||||||||||||||
| 資源タイプ識別子 | http://purl.org/coar/resource_type/c_5794 | |||||||||||||
| 資源タイプ | conference paper | |||||||||||||
| 出版タイプ | ||||||||||||||
| 出版タイプ | AM | |||||||||||||
| 出版タイプResource | http://purl.org/coar/version/c_ab4af688f83e57aa | |||||||||||||
| 会議記述 | ||||||||||||||
| 会議名 | International Symposium on Imaging, Sensing, and Optical Memory 2024, ISOM’24 | |||||||||||||
| 言語 | en | |||||||||||||
| 開始年 | 2024 | |||||||||||||
| 開始月 | 10 | |||||||||||||
| 開始日 | 20 | |||||||||||||
| 終了年 | 2024 | |||||||||||||
| 終了月 | 10 | |||||||||||||
| 終了日 | 23 | |||||||||||||
| 開催会場 | Arcrea HIMEJI | |||||||||||||
| 言語 | en | |||||||||||||
| 開催地 | Hyogo | |||||||||||||
| 言語 | en | |||||||||||||
| 開催国 | JPN | |||||||||||||
| 研究者情報 | ||||||||||||||
| URL | https://hyokadb02.jimu.kyutech.ac.jp/html/100000508_ja.html | |||||||||||||
| 連携ID | ||||||||||||||
| 値 | 14664 | |||||||||||||
