Implicit Bayes Adaptation: A Collaborative Transport Approach

Bo Jiang, Hamid Krim, Tianfu Wu, Derya Cansever

Introduction

Unsupervised Domain Adaptation (UDA) is a critical frontier in machine learning, addressing the challenge of "domain shift"—where a model trained on a source dataset fails to generalize to a target dataset due to differing underlying distributions. While Optimal Transport (OT) has emerged as a powerful framework for aligning these distributions by minimizing the cost of moving source data onto target data, standard OT approaches often struggle with high-dimensional, non-linear data structures. This paper introduces Implicit Bayes Adaptation, a novel framework that enhances OT by incorporating manifold geometry and a collaborative learning mechanism.

The Limitations of Standard Optimal Transport

Standard OT methods for UDA typically suffer from two primary drawbacks:

  1. Euclidean Bias: Most OT formulations assume data resides in a flat Euclidean space. However, high-dimensional data, such as images, often lies on complex, non-linear submanifolds. Ignoring this geometry leads to suboptimal transport plans.
  2. Label Mismatching: Without semantic constraints, OT may minimize geometric distance at the expense of class integrity—for example, mapping a source "1" to a target "7" simply because they are spatially close in the feature space.

Methodology

To address these issues, the authors propose two key innovations designed to respect the intrinsic structure of the data and ensure semantic consistency.

1. Geodesic-Aware Cost Function

Rather than relying on standard squared Euclidean distance ($L_2$ norm), the proposed method utilizes a refined metric inspired by Riemannian geometry. By incorporating a term based on the exponential map, the cost function better approximates the "geodesic" distance—the shortest path along the data manifold. This ensures that the transport plan respects the non-linear shape of the latent representations. Additionally, the authors "soften" cluster boundaries within the cost matrix to prevent mathematical singularities, facilitating smoother optimization and better convergence.

2. Collaborative Optimal Transport

The "Implicit Bayes" component of the framework involves a dual-stream alignment process. The model generates two distinct transport plans:

  • Plan A ($\gamma_1$): Based on labels predicted by the current neural network classifier.
  • Plan B ($\gamma_2$): Based on labels estimated by a $k$-nearest neighbor ($k$-NN) algorithm, which serves as a "Bayesian prior."

The training objective minimizes the Kullback–Leibler (KL) divergence between these two plans. By forcing the neural network's predictions to align with the local geometric consistency of the $k$-NN prior, the model reduces uncertainty and avoids poor local minima. The final loss function balances source classification accuracy, the collaborative OT loss, and target entropy regularization.

Experimental Results

The researchers evaluated the framework on standard UDA benchmarks, demonstrating superior performance over state-of-the-art methods such as ADDA, Fast OT, and DeepJDOT.

Digit Classification

On the challenging SVHN $\to$ MNIST task, which involves adapting from noisy real-world house numbers to clean handwritten digits, the model achieved 93.1% accuracy. This outperforms DeepJDOT (92.7%) and significantly exceeds ADDA (86.2%).

Office-31

In the Office-31 dataset, which features 31 categories across Amazon, DSLR, and Webcam domains, the model excelled in "hard" transfers. Specifically, in the DSLR $\to$ Amazon (D $\to$ A) task, it reached 71.5% accuracy, a substantial improvement over the 67.1% achieved by previous leading methods.

Discussion and Insights

The success of Implicit Bayes Adaptation highlights the importance of manifold awareness in domain adaptation. By accounting for the latent submanifolds where data representations reside, the model avoids the "gross mismatches" common in Euclidean-based transport. Furthermore, the collaborative $k$-NN mechanism acts as a robust sanity check, stabilizing the training process and preserving multi-class structures across domains.

Limitations and Future Work

Despite its performance, the framework introduces certain complexities:

  • Hyperparameter Sensitivity: The loss function requires tuning three $\lambda$ weights and two geometric parameters ($\alpha, \beta$), which may vary across datasets.
  • Computational Overhead: Maintaining a memory bank for $k$-NN and computing dual transport plans increases the computational footprint compared to simpler OT variants.
  • Scalability: Future research is needed to determine how the manifold reconstruction and collaborative framework scale to extremely large datasets containing millions of images.

In conclusion, Implicit Bayes Adaptation provides a more geometrically and semantically grounded approach to UDA, offering a robust path forward for adapting models to complex, real-world distribution shifts.

Summary updated

BibTeX
@inproceedings{jiang2023implicit,
  author = {Bo Jiang and Hamid Krim and Tianfu Wu and Derya Cansever},
  title = {{Implicit Bayes Adaptation: A Collaborative Transport Approach}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  doi = {10.1109/icassp49357.2023.10096599},
}
Figure