| Presentation | 2022-07-29 Fault Localization for RNNs Based on Probabilistic Automata and n-grams Yuta Ishimoto, Masanari Kondo, Naoyasu Ubayashi, Yasutaka Kamei, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | If deep learning models misbehave, serious accidents may occur.Previous studies have proposed approaches to overcome such misbehavior by detecting and modifying the responsible faulty parts (e.g., neurons of the network) in deep learning models.However, such approaches are not applicable to deep learning models that have internal states that change dynamically based on the input data, for example, recurrent neural networks (RNNs). To detect misbehavior RNNs, we propose PAFL, a new fault localization approach for application to RNNs.PAFL enables developers to detect faulty parts even in RNNs by computing suspiciousness scores with fault localization using $n$-grams.Furthermore, by using this suspiciousness score, PAFL can extract data strongly associated with RNN misbehavior.Compared to the random approach, PAFL can extract data that are statistically significantly more strongly associated with misbehavior.Specifically, in 83% of all experimental settings for two difficult datasets (i.e., RTMR and IMDB), PAFL can extract data that is difficult to predict of RNNs than randomly extracted data.Our experimental results show that PAFL is useful as a fault localization method for RNNs. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | deep learning / recurrent neural network / fault localization / probabilistic automaton / n-gram |
| Paper # | SS2022-10,KBSE2022-20 |
| Date of Issue | 2022-07-21 (SS, KBSE) |
| Conference Information | |
| Committee | SS / IPSJ-SE / KBSE |
|---|---|
| Conference Date | 2022/7/28(3days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Hokkaido-Jichiro-Kaikan (Sapporo) |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | Kozo Okano(Shinshu Univ.) / 鷲崎 弘宜(早稲田大学) / Takuya Saruwatari(NTT Data) |
| Vice Chair | Yoshiki Higo(Osaka Univ.) / / Yoshinori Tanabe(Tsurumi Univ.) |
| Secretary | Yoshiki Higo(Shinshu Univ.) / (Tokyo Inst. of Tech.) / Yoshinori Tanabe |
| Assistant | Shinsuke Matsumoto(Osaka Univ.) / 伊原 彰紀(和歌山大学) / 小川 秀人(日立製作所) / 竹内 広宜(武蔵大学) / 徳本 晋(富士通) / 伏田 享平(NTT株式会社) / 福田 浩章(芝浦工業大学) / 横川 智教(岡山県立大学) / Yoshitaka Aoki(BIPROGY) / Hiroki Horita(Ibaraki Univ.) |
| Paper Information | |
| Registration To | Technical Committee on Software Science / Special Interest Group on Software Engineering / Technical Committee on Knowledge-Based Software Engineering |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Fault Localization for RNNs Based on Probabilistic Automata and n-grams |
| Sub Title (in English) | |
| Keyword(1) | deep learning |
| Keyword(2) | recurrent neural network |
| Keyword(3) | fault localization |
| Keyword(4) | probabilistic automaton |
| Keyword(5) | n-gram |
| 1st Author's Name | Yuta Ishimoto |
| 1st Author's Affiliation | Kyushu University(Kyushu Univ.) |
| 2nd Author's Name | Masanari Kondo |
| 2nd Author's Affiliation | Kyushu University(Kyushu Univ.) |
| 3rd Author's Name | Naoyasu Ubayashi |
| 3rd Author's Affiliation | Kyushu University(Kyushu Univ.) |
| 4th Author's Name | Yasutaka Kamei |
| 4th Author's Affiliation | Kyushu University(Kyushu Univ.) |
| Date | 2022-07-29 |
| Paper # | SS2022-10,KBSE2022-20 |
| Volume (vol) | vol.122 |
| Number (no) | SS-138,KBSE-139 |
| Page | pp.pp.55-60(SS), pp.55-60(KBSE), |
| #Pages | 6 |
| Date of Issue | 2022-07-21 (SS, KBSE) |