| Presentation | 2019-12-12 Machine learning algorithms with quantized images and their influence Takayuki Osakabe, Yuma Kinoshita, Hitoshi Kiya, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Recently, appling quantized images to machine learning algorithmsis expected to enhance robustness against adversarial examples. However, quantizing data affects the performance of machine learning algorithms. In this paper, three quantized methods: linear quantization, lloyd-max quantization and error diffusion are applied to images respectively， and we consider the influence of the quantizationin some machine learning algorithms including deep learning for imageclassification. Experimental results show that we can get high classification accuracy even when low bits (1 or 2bit) images quantized by lloyd-max quantization are used in SVM, KNN and Logistic Regression. The results also demonstrate that we can obtain almost the same classificationaccuracy as that of baseline if we carefully choose a quantized method andthe number of bits under the use of each model. In deep learning with ResNet-20, the model gives high classification accuracyif both of training and test images are quantized by using an error diffusionalgorithm with the same number of bits. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | linear-quantization / lloyd-max quantization / error diffusion / machine learning / deep learning |
| Paper # | SIS2019-27 |
| Date of Issue | 2019-12-05 (SIS) |
| Conference Information | |
| Committee | SIS |
|---|---|
| Conference Date | 2019/12/12(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Okayama University of Science |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | Smart Personal Systems, etc. |
| Chair | Takayuki Nakachi(NTT) |
| Vice Chair | Noriaki Suetake(Yamaguchi Univ.) / Tomoaki Kimura(Kanagawa Inst. of Tech.) |
| Secretary | Noriaki Suetake(Tokyo Metropolitan Univ.) / Tomoaki Kimura(Kindai Univ.) |
| Assistant | Hideaki Misawa(National Inst. of Tech., Ube College) / Yukihiro Bandoh(NTT) |
| Paper Information | |
| Registration To | Technical Committee on Smart Info-Media Systems |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Machine learning algorithms with quantized images and their influence |
| Sub Title (in English) | |
| Keyword(1) | linear-quantization |
| Keyword(2) | lloyd-max quantization |
| Keyword(3) | error diffusion |
| Keyword(4) | machine learning |
| Keyword(5) | deep learning |
| 1st Author's Name | Takayuki Osakabe |
| 1st Author's Affiliation | Tokyo Metropolitan University(Tokyo Metro.Univ.) |
| 2nd Author's Name | Yuma Kinoshita |
| 2nd Author's Affiliation | Tokyo Metropolitan University(Tokyo Metro.Univ.) |
| 3rd Author's Name | Hitoshi Kiya |
| 3rd Author's Affiliation | Tokyo Metropolitan University(Tokyo Metro.Univ.) |
| Date | 2019-12-12 |
| Paper # | SIS2019-27 |
| Volume (vol) | vol.119 |
| Number (no) | SIS-335 |
| Page | pp.pp.23-28(SIS), |
| #Pages | 6 |
| Date of Issue | 2019-12-05 (SIS) |