Presentation | 2022-10-25 Relationship between the Defects in Learning Programs and the Model Distortion on the Convolutional Neural Networks Takumi Tsuchiya, Kozo Okano, Shinpei Ogata, Shin Nakajima, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | In recent years, the quality issue of machine learning software has become an important concern. When considering the quality of machine learning software, the standard evaluation method is to examine the performance of a deep learning model against a test dataset. In this paper, we assume that flaws in training programs, which is the root cause, appear as distortion of weights in trained CNN models, and evaluated the distortion of weights by a new proposed index using distribution of active neurons and NC (Neuron Coverage), where NC indicates the coverage of active neurons in the model. Using these indices, we conducted comparative experiments on probably correct programs and five defective programs in two datasets and five models and found that the indices were different between correct programs and defective programs, indicating that the indices are effective as an indicator of weight distortion. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | CNN / Neuron Coverage / Learning Program |
Paper # | SS2022-26,DC2022-32 |
Date of Issue | 2022-10-18 (SS, DC) |
Conference Information | |
Committee | DC / SS |
---|---|
Conference Date | 2022/10/25(1days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | |
Chair | Tatsuhiro Tsuchiya(Osaka Univ.) / Kozo Okano(Shinshu Univ.) |
Vice Chair | Toshinori Hosokawa(Nihon Univ.) / Yoshiki Higo(Osaka Univ.) |
Secretary | Toshinori Hosokawa(Nihon Univ.) / Yoshiki Higo(Chiba Univ.) |
Assistant | / Shinsuke Matsumoto(Osaka Univ.) |
Paper Information | |
Registration To | Technical Committee on Dependable Computing / Technical Committee on Software Science |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Relationship between the Defects in Learning Programs and the Model Distortion on the Convolutional Neural Networks |
Sub Title (in English) | |
Keyword(1) | CNN |
Keyword(2) | Neuron Coverage |
Keyword(3) | Learning Program |
1st Author's Name | Takumi Tsuchiya |
1st Author's Affiliation | Shinshu University(Shinshu Univ.) |
2nd Author's Name | Kozo Okano |
2nd Author's Affiliation | Shinshu University(Shinshu Univ.) |
3rd Author's Name | Shinpei Ogata |
3rd Author's Affiliation | Shinshu University(Shinshu Univ.) |
4th Author's Name | Shin Nakajima |
4th Author's Affiliation | National Institute of Informatics(NII) |
Date | 2022-10-25 |
Paper # | SS2022-26,DC2022-32 |
Volume (vol) | vol.122 |
Number (no) | SS-225,DC-226 |
Page | pp.pp.23-28(SS), pp.23-28(DC), |
#Pages | 6 |
Date of Issue | 2022-10-18 (SS, DC) |