The Geometry and Calculus of Losses

Authors: Robert C. Williamson, Zac Cranko

JMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical The paper contains no new algorithms and no experimental results. What it does contain is a new way of looking at loss functions which 1) illustrates the close connection between losses and norms and anti-norms; 2) presents the new idea of an antipolar loss; 3) develops a calculus for loss functions that allows multiple proper loss functions to be combined in a manner that the resulting loss is guaranteed proper; and 4) shows how the geometrical perspective can be used to design loss functions.
Researcher Affiliation Academia Robert C. Williamson EMAIL University of Tübingen and Tübingen AI Center, Germany; Zac Cranko EMAIL Sydney, Australia
Pseudocode No The paper contains no new algorithms and no experimental results. What it does contain is a new way of looking at loss functions...
Open Source Code No The paper contains no new algorithms and no experimental results.
Open Datasets No The paper contains no new algorithms and no experimental results.
Dataset Splits No The paper contains no new algorithms and no experimental results.
Hardware Specification No The paper contains no new algorithms and no experimental results.
Software Dependencies No The paper contains no new algorithms and no experimental results.
Experiment Setup No The paper contains no new algorithms and no experimental results.