[Paper Review] Domain Independent SVM for Transfer Learning in Brain Decoding

Shuo Zhou, Wenwen Li+2 more|arXiv (Cornell University)|Mar 26, 2019

Domain Adaptation and Few-Shot Learning24 references6 citations

TL;DR

This paper proposes Domain Independent SVM (DI-SVM), a novel transfer learning method for brain condition decoding that minimizes both SVM empirical risk and dependence on domain-specific factors (e.g., experiment design, subject identity) using a simplified Hilbert-Schmidt Independence Criterion (HSIC). DI-SVM achieves over 24% higher accuracy than competing methods on multi-source transfer learning tasks, demonstrating its effectiveness in leveraging heterogeneous brain imaging data across subjects and experimental paradigms.

ABSTRACT

Brain imaging data are important in brain sciences yet expensive to obtain, with big volume (i.e., large p) but small sample size (i.e., small n). To tackle this problem, transfer learning is a promising direction that leverages source data to improve performance on related, target data. Most transfer learning methods focus on minimizing data distribution mismatch. However, a big challenge in brain imaging is the large domain discrepancies in cognitive experiment designs and subject-specific structures and functions. A recent transfer learning approach minimizes domain dependence to learn common features across domains, via the Hilbert-Schmidt Independence Criterion (HSIC). Inspired by this method, we propose a new Domain Independent Support Vector Machine (DI-SVM) for transfer learning in brain condition decoding. Specifically, DI-SVM simultaneously minimizes the SVM empirical risk and the dependence on domain information via a simplified HSIC. We use public data to construct 13 transfer learning tasks in brain decoding, including three interesting multi-source transfer tasks. Experiments show that DI-SVM's superior performance over eight competing methods on these tasks, particularly an improvement of more than 24% on multi-source transfer tasks.

Motivation & Objective

• To address the challenge of small sample size and high-dimensional brain imaging data in neuroimaging studies.
• To improve brain condition decoding performance by leveraging transfer learning across heterogeneous domains, such as different subjects and experimental designs.
• To develop a domain-independent classifier that generalizes across diverse brain imaging datasets by minimizing reliance on domain-specific features.
• To validate the effectiveness of domain independence through a comprehensive set of 13 transfer learning tasks derived from public neuroimaging repositories.

Proposed method

• DI-SVM jointly optimizes the standard SVM empirical risk and a simplified HSIC-based measure of dependence on auxiliary domain information (e.g., subject ID, experiment type).
• Domain information is encoded via one-hot vectors and used to compute a simplified HSIC that quantifies statistical dependence between features and domain labels.
• The method introduces a regularization term controlled by a hyperparameter λ that balances classification accuracy and domain independence.
• A linear or kernelized SVM is trained with the combined objective, enabling the model to learn features that are both discriminative and invariant to domain shifts.
• The approach is evaluated on 13 transfer learning tasks, including multi-source settings, using real fMRI data from OpenNeuro and HCP.
• Hyperparameter sensitivity analysis is performed on C (regularization) and λ (domain dependence weight), showing robustness and improved generalization when λ > 0.

Experimental results

Research questions

• RQ1 Can a transfer learning method that explicitly minimizes dependence on domain-specific factors improve brain condition decoding performance across heterogeneous datasets?
• RQ2 How does DI-SVM compare to state-of-the-art transfer learning methods in single-source and multi-source brain decoding tasks?
• RQ3 Does incorporating multiple source domains simultaneously lead to better performance than using a single source, especially when domain differences are large?
• RQ4 How sensitive is DI-SVM’s performance to key hyperparameters like C and λ in the presence of small sample sizes?
• RQ5 To what extent does domain independence—measured via HSIC—enhance generalization in brain decoding when training data come from different subjects and experimental designs?

Key findings

• DI-SVM achieved the highest overall accuracy across all 13 transfer learning tasks, outperforming eight competing methods, including TCA, JDA, and MIDA.
• On multi-source transfer tasks, DI-SVM showed a performance improvement of over 24% compared to the next-best method, highlighting its strength in leveraging multiple sources.
• The best-performing variant, DI-SVM with linear kernel, achieved 91.6% accuracy on the most challenging multi-source task (A&C → B), significantly outperforming the second-best method (MIDA) at 65.4%.
• DI-SVM was robust to hyperparameter changes, with classification accuracy remaining stable for λ > 0, and performance dropping significantly when λ = 0 (equivalent to standard SVM).
• HSIC-based methods, including MIDA and DI-SVM, consistently outperformed non-HSIC methods, confirming the value of domain dependence minimization in brain decoding.
• Source selection significantly impacted performance: multi-source transfer did not always improve accuracy, but DI-SVM still achieved the highest results across all configurations, proving its robustness.