A Unified Framework of Surrogate Loss by Refactoring and Interpolation
This work addresses the need for automated loss design in deep learning, offering a generalizable solution that reduces manual effort, though it appears incremental as it builds on existing interpolation and refactoring techniques.
The paper tackles the problem of manually designing task-specific surrogate losses for training deep networks by introducing UniLoss, a unified framework that refactors performance metrics into differentiable steps and uses interpolation to generate surrogate losses, achieving comparable performance on three tasks and four datasets.
We introduce UniLoss, a unified framework to generate surrogate losses for training deep networks with gradient descent, reducing the amount of manual design of task-specific surrogate losses. Our key observation is that in many cases, evaluating a model with a performance metric on a batch of examples can be refactored into four steps: from input to real-valued scores, from scores to comparisons of pairs of scores, from comparisons to binary variables, and from binary variables to the final performance metric. Using this refactoring we generate differentiable approximations for each non-differentiable step through interpolation. Using UniLoss, we can optimize for different tasks and metrics using one unified framework, achieving comparable performance compared with task-specific losses. We validate the effectiveness of UniLoss on three tasks and four datasets. Code is available at https://github.com/princeton-vl/uniloss.