Multi-Task Learning Using Uncertainty to Weight Losses for Scene Geometry and Semantics 阅读

Posted 2021-08-04 一杯敬朝阳一杯敬月光

摘要

作者表明，他们发现多任务学习的效果很大程度依赖各损失的相对权重， In this paper we make the observation that the performance of such systems is strongly dependent on the relative weighting between each task’s loss.将这些权重作为模型的超参调节所花费的代价是巨大的。文中提出依据各任务的同方差不确定性对损失进行加权，weighs multiple loss functions by considering the homoscedastic uncertainty of each task。这样就可以同时学习分类和回归任务中的不同单位或不同尺度的数据，This allows us to simultaneously learn various quantities with different units or scales in both classification and regression settings。

引言

多任务学习旨在通过共享的表示学习多个目标来提高效率和预测的准确性，Multi-task learning aims to improve learning efficiency and prediction accuracy by learning multiple objectives from a shared representation。实验场景是视觉场景理解visual scene understanding，场景理解算法必须同时理解场景的几何和语义，Scene understanding algorithms must understand both the geometry and semantics of the scene at the same time，这意味着需要同时学习具有不同单位和尺度的各种分类和回归任务，joint learning of various regression and classification tasks with different units and scales。这边还交代了多任务学习的一个优点：多个任务整合到一起，可以减少计算量，实时性更好，Combining all tasks into a single model reduces computation and allows these systems to run in real-time。在这之前的工作，多任务学习的损失函数要么使用简单的加和，要么人工调优。作者在实验中观察到每个任务最优权重取决于度量的尺度（例如，米，厘米，毫米），归根结底取决于任务噪声的大小（感觉上是每个任务相对权重的意思，其他任务输出的尺度不同，影响自然不同），We observe that the optimal weighting of each task is dependent on the measurement scale (e.g. meters, centimetres or millimetres) and ultimately the magnitude of the task’s noise。

未完待续