Deep Ordinal Regression using Optimal Transport Loss and Unimodal Output Probabilities

Abstract

We propose a framework for deep ordinal regression, based on unimodal output distribution and optimal transport loss. Despite being seemingly appropriate, in many recent works the unimodality requirement is either absent, or implemented using soft targets, which do not guarantee unimodal outputs at inference. In addition, we argue that the standard maximum likelihood objective is not suitable for ordinal regression problems, and that optimal transport is better suited for this task, as it naturally captures the order of the classes. Inspired by the well-known Proportional Odds model, we propose to modify its design by using an architectural mechanism which guarantees that the model output distribution will be unimodal. We empirically analyze the different components of our propose approach and demonstrate their contribution to the performance of the model. Experimental results on three real-world datasets demonstrate that our proposed approach performs on par with several recently proposed deep learning approaches for deep ordinal regression with unimodal output probabilities, while having guarantee on the output unimodality. In addition, we demonstrate that the level of prediction uncertainty of the model correlates with its accuracy.

Multi-Task Convolutional Neural Network for Image Aesthetic Assessment

“…Reddy et al (2020) [35] proposed a multi-task network based on EfficientNet [42] for the same purpose, along with a visualization technique and activation maps generated using Gradient-weighted Class Activation Mapping (Grad-CAM) [37] More recently, another image dataset, namely the Explainable Visual Aesthetics (EVA) [14], has been released, which includes overall aesthetic scores and attribute scores. Although there are a few studies that have used this dataset for aesthetics research, their models only predict overall aesthetic scores [6], [25], [26], [38]. Therefore, our study is the first multi-task neural network that can make predictions on the EVA dataset.…”

Controlling Class Layout for Deep Ordinal Classification via Constrained Proxies Learning

“…Traditional ordinal classification methods (Frank and Hall 2001;Chu, Ghahramani, and Williams 2005;Cardoso and da Costa 2007;Lin and Li 2012) mainly work on handcrafted features, which is labor-intensive and time-consuming. Recently, with the great progress brought by deep neural networks, several deep ordinal classification methods have been proposed (Liu, Kong, and Goh 2018;Diaz and Marathe 2019;Shaham and Svirsky 2020;Li et al 2021) and show superior performance than traditional methods (Liu, Kong, and Goh 2018). Such performance gain is mainly attributed to neural networks' strong capability of representation learning.…”