Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Improve Unsupervised Domain Adaptation with Mixup Training (2001.00677v1)

Published 3 Jan 2020 in stat.ML, cs.CV, and cs.LG

Abstract: Unsupervised domain adaptation studies the problem of utilizing a relevant source domain with abundant labels to build predictive modeling for an unannotated target domain. Recent work observe that the popular adversarial approach of learning domain-invariant features is insufficient to achieve desirable target domain performance and thus introduce additional training constraints, e.g. cluster assumption. However, these approaches impose the constraints on source and target domains individually, ignoring the important interplay between them. In this work, we propose to enforce training constraints across domains using mixup formulation to directly address the generalization performance for target data. In order to tackle potentially huge domain discrepancy, we further propose a feature-level consistency regularizer to facilitate the inter-domain constraint. When adding intra-domain mixup and domain adversarial learning, our general framework significantly improves state-of-the-art performance on several important tasks from both image classification and human activity recognition.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (5)
  1. Shen Yan (47 papers)
  2. Huan Song (13 papers)
  3. Nanxiang Li (4 papers)
  4. Lincan Zou (2 papers)
  5. Liu Ren (57 papers)
Citations (211)

Summary

Improve Unsupervised Domain Adaptation with Mixup Training: An Analytical Overview

The paper investigates advancements in the area of unsupervised domain adaptation (UDA), specifically focusing on enhancing model performance in target domains lacking labeled data. This is achieved by leveraging the concept of mixup training, traditionally used for regularization in supervised learning, to enforce cross-domain training constraints. The authors propose a general framework termed Inter- and Intra-domain Mixup Training (IIMT), which significantly elevates performance on various benchmarks, including image classification and human activity recognition (HAR) tasks.

Core Contributions

Central to this work is the introduction of mixing training examples from both source and target domains to strengthen domain adaptation. Traditionally, UDA relies heavily on adversarial techniques for creating domain-invariant representations, typified by models like Domain-Adversarial Neural Networks (DANN). However, these often fall short as they do not encompass inter-domain constraints comprehensively. The mixup approach here involves interpolating training data across domains to bolster generalization to unannotated domains. Critical refinements include:

  1. Inter-domain mixup training: By incorporating target-domain predictions in mixup processes, a seamless blending of source and target features is achieved, offering additional supervision levels absent in existing adversarial methodologies.
  2. Consistency regularizer: A feature-level consistency regularizer is formulated to maintain linearity in the learned feature space, addressing potentially substantial discrepancies between domains.
  3. Intra-domain mixup training: This ensures local smoothness and mitigates rapid prediction shifts in feature space, similar to Virtual Adversarial Training (VAT).
  4. End-to-end training protocol: An advantageous synthesis of mixup with traditional adversarial training enhances domain adaptability without introducing complex additional learning objectives.

Experimental Validation

The IIMT framework outperforms existing methods across established benchmarks. On visual domain tasks involving datasets such as MNIST, CIFAR-10, and STL-10, IIMT consistently excels, achieving up to 8.1% improvement over previous state-of-the-art techniques in challenging settings like STL \rightarrow CIFAR-10 transitions. Meanwhile, for HAR tasks, IIMT showcases substantial improvements in cross-subject evaluations on the OPPORTUNITY dataset and environments like the WiFi activity dataset, suggesting robust real-world applicability.

Key numerical results highlight the ability of IIMT to deliver substantial accuracy improvements, evidencing the efficacy of inter and intra domain interactions tailored through mixup.

Implications and Future Scope

By bridging inter-domain gaps using mixup, this approach introduces a promising dimension for domain adaptation models. Beyond numerical advancements, it implies the potential for mixup-based strategies to mitigate data scarcity issues in settings where labeled data is unattainable or costly. The implications extend further into creating more generalized solutions applicable across varying shifts in data distribution, involving both time-series and spatial datasets.

Moving forward, the exploration of more sophisticated mixup formulations and potential integration with cutting-edge semi-supervised learning techniques could yield even more robust adaptation paradigms. Additionally, exploring domain-specific regularizations and the inclusion of auxiliary data could further amplify adaptation capabilities, especially as AI becomes more deeply integrated into everyday applications requiring adaptability in dynamically changing environments.