Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hybrid Losses for Hierarchical Embedding Learning

Haokun Tian, Stefan Lattner, Brian McFee, Charalampos Saitis

TL;DR

The paper addresses how traditional cross-entropy training fails to capture label similarity by exploiting a tree-structured hierarchy of fine-grained labels. It introduces a hybrid multi-task framework that combines a generalised triplet loss with various hierarchy-aware classification losses to shape the embedding space and enable generalisation to unseen classes. Key contributions include the definition of Leaf, Binary, and Per-Level losses, a tree-based triplet sampling strategy, and novel evaluation metrics that assess how well embeddings reflect the hierarchy and predict unseen classes, demonstrated on the OrchideaSOL dataset. The findings show that hierarchy-aware, multi-task hybrids (notably Per-Level alone or in combination with Triplet) improve classification, retrieval, embedding structure, and unseen-class generalisation, with potential applicability to other hierarchical labeling tasks and larger multi-label datasets like AudioSet.

Abstract

In traditional supervised learning, the cross-entropy loss treats all incorrect predictions equally, ignoring the relevance or proximity of wrong labels to the correct answer. By leveraging a tree hierarchy for fine-grained labels, we investigate hybrid losses, such as generalised triplet and cross-entropy losses, to enforce similarity between labels within a multi-task learning framework. We propose metrics to evaluate the embedding space structure and assess the model's ability to generalise to unseen classes, that is, to infer similar classes for data belonging to unseen categories. Our experiments on OrchideaSOL, a four-level hierarchical instrument sound dataset with nearly 200 detailed categories, demonstrate that the proposed hybrid losses outperform previous works in classification, retrieval, embedding space structure, and generalisation.

Hybrid Losses for Hierarchical Embedding Learning

TL;DR

The paper addresses how traditional cross-entropy training fails to capture label similarity by exploiting a tree-structured hierarchy of fine-grained labels. It introduces a hybrid multi-task framework that combines a generalised triplet loss with various hierarchy-aware classification losses to shape the embedding space and enable generalisation to unseen classes. Key contributions include the definition of Leaf, Binary, and Per-Level losses, a tree-based triplet sampling strategy, and novel evaluation metrics that assess how well embeddings reflect the hierarchy and predict unseen classes, demonstrated on the OrchideaSOL dataset. The findings show that hierarchy-aware, multi-task hybrids (notably Per-Level alone or in combination with Triplet) improve classification, retrieval, embedding structure, and unseen-class generalisation, with potential applicability to other hierarchical labeling tasks and larger multi-label datasets like AudioSet.

Abstract

In traditional supervised learning, the cross-entropy loss treats all incorrect predictions equally, ignoring the relevance or proximity of wrong labels to the correct answer. By leveraging a tree hierarchy for fine-grained labels, we investigate hybrid losses, such as generalised triplet and cross-entropy losses, to enforce similarity between labels within a multi-task learning framework. We propose metrics to evaluate the embedding space structure and assess the model's ability to generalise to unseen classes, that is, to infer similar classes for data belonging to unseen categories. Our experiments on OrchideaSOL, a four-level hierarchical instrument sound dataset with nearly 200 detailed categories, demonstrate that the proposed hybrid losses outperform previous works in classification, retrieval, embedding space structure, and generalisation.
Paper Structure (13 sections, 8 equations, 2 figures, 1 table)

This paper contains 13 sections, 8 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Triplet Learning
  4. Sampling Triplets from the Tree
  5. Hierarchy-Aware Losses
  6. Experiments and Evaluation
  7. Dataset
  8. Data Split
  9. Training
  10. Evaluation
  11. Classification
  12. Retrieval
  13. Results and Discussion

Figures (2)

  • Figure 1: Visualisation of how targets are chosen from a sample tree for different hierarchy-aware losses. Left: shading indicates binary classification on the node. Right: multi-class classification is performed within each group of nodes sharing the same colour.
  • Figure 2: Overview of the model pipeline.