Multi-Domain Adversarial Learning

TL;DR

This paper introduces MuLANN, a multi-domain adversarial learning method that improves model performance across diverse datasets with experimental bias, especially in bioimaging, by leveraging a new theoretical bound and semi-supervised loss.

Contribution

It proposes a novel multi-domain adversarial learning framework with theoretical risk bounds, a new semi-supervised loss, and demonstrates improved results on benchmarks and a new bioimage dataset.

Findings

Improves state-of-the-art on two image benchmarks.

Validates approach on a new bioimage dataset.

Provides theoretical bounds on domain risk.

Abstract

Multi-domain learning (MDL) aims at obtaining a model with minimal average risk across multiple domains. Our empirical motivation is automated microscopy data, where cultured cells are imaged after being exposed to known and unknown chemical perturbations, and each dataset displays significant experimental bias. This paper presents a multi-domain adversarial learning approach, MuLANN, to leverage multiple datasets with overlapping but distinct class sets, in a semi-supervised setting. Our contributions include: i) a bound on the average- and worst-domain risk in MDL, obtained using the H-divergence; ii) a new loss to accommodate semi-supervised multi-domain learning and domain adaptation; iii) the experimental validation of the approach, improving on the state of the art on two standard image benchmarks, and a novel bioimage dataset, Cell.