WEKO3
アイテム
SpikePropネットワークに対し入出力関係を滑らかにするウェイト・ディケイ導入効果の検討
http://hdl.handle.net/10076/12402
http://hdl.handle.net/10076/12402cdd2fc01-b244-4de6-b952-ee4b90c81555
| 名前 / ファイル | ライセンス | アクション |
|---|---|---|
2011Ｍ404.pdf (922.7 kB)
|
|
| Item type | 学位論文 / Thesis or Dissertation(1) | |||||||
|---|---|---|---|---|---|---|---|---|
| 公開日 | 2013-05-09 | |||||||
| タイトル | ||||||||
| タイトル | SpikePropネットワークに対し入出力関係を滑らかにするウェイト・ディケイ導入効果の検討 | |||||||
| 言語 | ja | |||||||
| 言語 | ||||||||
| 言語 | jpn | |||||||
| 資源タイプ | ||||||||
| 資源タイプ識別子 | http://purl.org/coar/resource_type/c_46ec | |||||||
| 資源タイプ | thesis | |||||||
| 著者 |
吉川, 雄也
× 吉川, 雄也
|
|||||||
| 抄録 | ||||||||
| 内容記述タイプ | Abstract | |||||||
| 内容記述 | Recently spiking neural networks (SNNs) attract many research attentions in the robot control field. SNNs represent information by spikes. They are good at temporally signal processing such as robot control. We focus on SpikeProp, which is a model of SNNs. It represents information by timing of spikes. Our final goal is to apply SpikeProp to robot control. SpikeProp is based on complex model, which is an integrate-firing model. It causes complex behaviors: outputs of SpikeProp is often sensitive to weights. It implies that the outputs are also sensitive to inputs. These sensitivities cause unexpected output in a noisy environment. So, our purpose is to reduce discontinuous output changes against a continuous input changes. To reduce discontinuous output changes, we introduce Weight Decay (WD), which improve input-output characteristic of conventional neural networks, which are constructed with non-spiking neuron such as sigmoidal neurons. WD exactly works to eliminating useless connections by reducing both the output error and a sum of squared weights. We assume that discontinuous output changes are caused by local maxima of activities. Since an activity is a sum of each response for all input stimuli, eliminating useless connections would reduce local maxima of activities. As a result, SpikeProp would resolve the discontinuity problem. We show the effectiveness of our method by some simple experiments: training the XOR-problem that is a famous benchmark problem. Without WD, SpikeProp show discontinuous output changes even near trained input patterns. By applying WD, discontinuous output changes are reduced around the trained input patterns. We show that the WD works well not only for conventional neural networks but also SpikeProp. In future, more improvements are required. In particular, it’s important to lessen weights which are related to local maxima of activity that degrade input-output characteristics. Improved method would reduce discontinuous output changes. Finally, we can apply SpikeProp to real problems such as robot control. | |||||||
| 内容記述 | ||||||||
| 内容記述タイプ | Other | |||||||
| 内容記述 | 三重大学大学院地域イノベーション学研究科博士前期課程地域イノベーション学専攻 | |||||||
| 内容記述 | ||||||||
| 内容記述タイプ | Other | |||||||
| 内容記述 | 4, 35 | |||||||
| 書誌情報 |
発行日 2011-01-01 |
|||||||
| フォーマット | ||||||||
| 内容記述タイプ | Other | |||||||
| 内容記述 | application/pdf | |||||||
| 著者版フラグ | ||||||||
| 出版タイプ | VoR | |||||||
| 出版タイプResource | http://purl.org/coar/version/c_970fb48d4fbd8a85 | |||||||
| 出版者 | ||||||||
| 出版者 | 三重大学 | |||||||
| 修士論文指導教員 | ||||||||
| 寄与者識別子Scheme | WEKO | |||||||
| 寄与者識別子 | 27057 | |||||||
| 姓名 | 高瀬, 治彦 | |||||||
| 言語 | ja | |||||||
| 資源タイプ（三重大） | ||||||||
| 値 | Master's Thesis / 修士論文 | |||||||
