SENTRY: Selective Entropy Optimization via Committee Consistency for Unsupervised Domain Adaptation

Many existing approaches for unsupervised domain adaptation (UDA) focus on\nadapting under only data distribution shift and offer limited success under\nadditional cross-domain label distribution shift. Recent work based on\nself-training using target pseudo-labels has shown promise, but on challenging\nshifts pseudo-labels may be highly unreliable, and using them for self-training\nmay cause error accumulation and domain misalignment. We propose Selective\nEntropy Optimization via Committee Consistency (SENTRY), a UDA algorithm that\njudges the reliability of a target instance based on its predictive consistency\nunder a committee of random image transformations. Our algorithm then\nselectively minimizes predictive entropy to increase confidence on highly\nconsistent target instances, while maximizing predictive entropy to reduce\nconfidence on highly inconsistent ones. In combination with pseudo-label based\napproximate target class balancing, our approach leads to significant\nimprovements over the state-of-the-art on 27/31 domain shifts from standard UDA\nbenchmarks as well as benchmarks designed to stress-test adaptation under label\ndistribution shift.\n