This New AI Architecture Subsumes CNNs, RNNs, and Transformers — and It's Surprisingly Simple

A new paper on arXiv introduces the Integral Transform Network (ITNet), which challenges the long-held assumption that CNNs, Transformers, and RNNs are fundamentally different architectural families. The authors — Ashim Dhor, Rasel Mondal, and Pin Yu Chen — prove that convolution, self-attention (including multi-head), and autoregressive recurrence (LSTM, GRU, S4, Mamba) are all special cases of a learnable integral transform. The core of ITNet is a small MLP-based kernel that models pairwise interactions depending on both positions and features, allowing the network to adapt its inductive bias purely from data. The paper shows ITNet is a universal approximator of continuous operators.

To make the architecture practical at scale, the team introduces three computational innovations: tiled kernel fusion for efficient memory access, importance-weighted Monte Carlo integration for fast kernel approximation, and learned low-rank factorization to reduce parameter count. Despite being a single unified architecture with a shared operator and lightweight modality-specific encoders, ITNet matches or exceeds specialized baselines across five diverse benchmarks: ImageNet-1K (image classification), GLUE (NLP), ModelNet40 (3D point clouds), VQA v2 (visual question answering), and NLVR2 (visual reasoning). This suggests that the fragmentation into separate architectural families is not a fundamental necessity but an artifact of incomplete mathematical formulations, and that a single learned interaction mechanism can adapt to any task given sufficient data.