Accepted Paper: Adaptive truncated residual regression for fine-grained regression problems

Back to list of accepted papers


Hirotaka Hachiya (Center for Advanced Intelligence Project, RIKEN); Yamamoto Yu (Center for Advanced Intelligence Project, RIKEN); Kazuro HIrahara (Center for Advance Intelligence Project, RIKEN); Naonori Ueda (NTT)


Recently, anchor-based regression methods have been applied to challenging regression problems, e.g., object detection and distance estimation, and greatly improved those performances. The key idea of anchor-based regression is to reduce the variance of target variables by replacing the regression problem to the regression of the residual between selected anchors and the supervised signals. However, similar to an ordinary regression method, the anchor-based regression could face the difficulty on a fine-grained regression problem where the residual variables tend to be too small and complicated to accurately predict. To overcome this problem on the anchor-based regression, we propose to introduce an adaptive residual truncation in which the residual value is truncated and normalized, using adaptively tuned sigmoid function. Our proposed method, called ATR-Nets (Adaptive Truncated Residual-Networks) with an end-to-end deep-learning architecture could control the range of the target residual to be fitted so as to maximize the regression performance. Through experiments with toy-data and the system identification of earthquake asperity models, we show the effectiveness of our proposed method.