Table 2 Phone error rates of the CNN for various types of pooling functions
From: Phone recognition with hierarchical convolutional deep maxout networks
Network type and training method | Frame error on developmental set (%) | Phone error | |
|---|---|---|---|
Developmental set (%) | Core test (%) | ||
ReLU | 34.7 | 16.0 | 18.8 |
Maxout | 34.3 | 15.7 | 18.3 |
Maxout, DPT | 34.2 | 15.8 | 18.1 |
2-norm, DPT | 35.5 | 16.3 | 18.9 |
Maxout, hybrid DPT | 33.4 | 15.6 | 18.0 |