Improving Pseudo-Multiview Learning via Dual-Sampler Bidirectional Feature Flow

Pseudo-multiview learning (PML) improves classification by generating complementary internal representations and fusing their evidence using subjective logic. However, conventional PML incurs substantial computational costs as the number of pseudo-views increases. Multiscale grid-based PML alleviates this issue by generating lower-resolution pseudo-views through downsampling and integrating cross-scale features. Nevertheless, its unidirectional top-down feature flow is vulnerable to noise in high-resolution branches, which can propagate to lower-resolution branches and degrade classification performance. This paper proposes a noise-resilient multiscale PML framework. The proposed dual-sampler module enables each branch to aggregate complementary features from both higher- and lower-resolution branches, thereby establishing bidirectional cross-scale information flow. In addition, a noise-aware spatial-channel attention module jointly models feature discriminativeness and reliability, suppressing corrupted spatial regions and feature channels during cross-scale fusion. Experimental results on real-world datasets from diverse domains, including BreakHis at 40×, 100×, 200×, and 400× magnifications, Oxford-IIIT Pet, and Chest X-ray, demonstrate the superiority of the proposed method over baseline models in classification accuracy, convergence stability, and robustness to cross-scale Gaussian noise.