SIREN: Scalable isotropic recursive column multimodal neural architecture

Hubert Truchan, Alan Farag, Zahra Ahmadi

Abstract: Highlights•Introduces a multimodal isotropic architecture with <<math><mo is="true">&lt;</mo></math>18MB, <<math><mo is="true">&lt;</mo></math>4.5M params, 23ms/obs.•Scales to complex modalities via isotropic blocks, recurrent and column links.•Presents a 2D embedding that avoids compression, enabling flexible downsizing.•Achieves higher accuracy than SOTA, with recursion boosting semantic features.•Releases the Pentostreda dataset with five modalities for classification and regression.

External IDs:doi:10.1016/j.inffus.2025.103500