Uncertainty Generation Meta-training for Cross-domain Few-shot Learning

Cross-domain few-shot learning aims to generalize source knowledge to an unknown target domain. However, some recent methods are restricted to additive perturbations, limiting their effectiveness. To overcome this problem, we propose an effective meta-uncertainty augmented learning method. A gradient-based uncertainty generation method is first introduced, which considers feature statistics as probabilistic representations following an unknown distribution, and utilizes task gradients to determine the range of probability variations, thereby constructing uncertainty feature distributions to mimic domain bias. Furthermore, a causally invariant information module is proposed to extract causal information to improve the homogeneity of the task gradient, which provides further guarantees for the construction of uncertainty boundaries. With these two modules, an asymptotic meta-learning optimization algorithm is presented to learn generalizable knowledge from uncertainty features while optimizing the boundary of the uncertainty distribution. Experimental results conducted on nine datasets show that the proposed method improves over the current state-of-the-art method by an average of 3.11%.