Assisting Main Task Learning by Heterogeneous Auxiliary Tasks with Applications to Skin Cancer Screening

Ning Situ, Xiaojing Yuan, George ZouridakisAISTATS 2011

Abstract

This paper introduces a novel Multitask Learning (MTL) framework, Concurrent Main and Heterogeneous Auxiliary task learning (CMHA), designed to leverage "nonoperational" features—expert knowledge available during training but absent during deployment. Unlike traditional MTL models that assume homogeneous input spaces, CMHA accommodates auxiliary tasks with unique features. By modeling the main task as a weighted combination of these auxiliary learners, the framework significantly improves diagnostic accuracy, as demonstrated by a 9.42% improvement in skin cancer screening AUC over baseline models.

Introduction: The Challenge of Nonoperational Features

In specialized domains like medicine, experts utilize high-level semantic concepts to reach conclusions. For instance, a dermatologist identifies specific patterns such as "blue-whitish veils" or "irregular dots" to diagnose melanoma. While these features are highly informative, they are often "nonoperational": in real-world deployment, an expert is not available to label every new image with these complex features.

The CMHA framework leverages expert-identified local and global patterns as auxiliary tasks to improve the main melanoma diagnosis task.

Traditional Multitask Learning (MTL) attempts to solve this by learning related tasks simultaneously. However, most MTL models are homogeneous, assuming all tasks share the same input space. In reality, identifying a "color" pattern requires different low-level data than identifying a "geometric" structure. This paper addresses the challenge of heterogeneous auxiliary tasks, where each task possesses its own unique input space and features.

Methodology: The CMHA Framework

The authors propose the CMHA (Concurrent Main and Heterogeneous Auxiliary task learning) model, which incorporates several key innovations:

  • Weighted Ensemble Approach: The main task classifier is modeled as a weighted average of the outputs from various auxiliary task learners, rather than a simple concatenation of features.
  • Convex Optimization: The learning process is formulated as a convex optimization problem, ensuring a globally optimal solution and avoiding the local minima common in complex neural networks.
  • Task Relatedness and Regularization: The model includes a regularization term that penalizes the "distance" between the main task and auxiliary tasks. This allows the model to automatically assign higher weights to relevant tasks and ignore irrelevant or noisy ones.
  • Theoretical Guarantees: The work provides mathematical "true risk bounds," offering formal guarantees on the model's performance on unseen data.

Experimental Evaluation

The researchers evaluated CMHA across two distinct domains to test its performance and generalizability:

  1. Skin Cancer Screening: Utilizing 360 dermoscopy images (90 melanoma, 270 benign), the main task was melanoma detection. Auxiliary tasks included identifying one global pattern and three local dermoscopic features.
  2. Music Genre Recognition: Using the CAL500 dataset (502 songs), the model predicted "electronica" or "alternative" genres using "electric guitar" presence as an auxiliary task.

Key Results and Insights

The CMHA method demonstrated superior performance across benchmarks. In the skin cancer screening domain, it achieved an Area Under the Curve (AUC) of 83.43%, representing a 9.42% improvement over the baseline single-task model.

Quantitative comparison showing CMHA's significant AUC improvement over standard MTL and kernel learning methods.

Key Insights:

  • Labels as Features: Expert labels for sub-features act as independent learning tasks that guide the main learner.
  • Concurrent Learning: Learning auxiliary and main tasks simultaneously allows auxiliary classifiers to adapt specifically to the main task's requirements, outperforming sequential "two-step" methods.
  • Robustness to Noise: When "junk" auxiliary tasks were intentionally added, the CMHA model successfully assigned them near-zero weights, maintaining performance integrity.

Conclusion

By treating expert knowledge as heterogeneous auxiliary tasks, the CMHA framework effectively bridges the gap between training-time expertise and deployment-time constraints. The model's ability to handle noise and its efficiency in avoiding the "curse of dimensionality" make it a versatile tool for complex multi-feature classification problems beyond medical imaging. Future work will explore incorporating unlabeled data through semi-supervised techniques and refining hyperparameter selection processes.

Summary updated

BibTeX
@article{situ2011assisting,
  author = {Ning Situ and Xiaojing Yuan and George Zouridakis},
  title = {{Assisting Main Task Learning by Heterogeneous Auxiliary Tasks with Applications to Skin Cancer Screening}},
  journal = {AISTATS},
  year = {2011},
}
Figure