ITNet Unifies AI Architectures: A Learnable Integral Transform for All Models

For decades, major neural network architectures like convolutional networks, recurrent networks, and transformers have been considered mathematically distinct, each embodying specific inductive biases such as locality, sequential memory, and content-dependent interactions. This research challenges that fragmentation by proposing that these diverse architectures are, in fact, incomplete views of a single, underlying mathematical object: a learnable integral transform. The Integral Transform Network (ITNet) is introduced as a unified architecture built around this concept. At its core is a learnable kernel, implemented as a small neural network (specifically, an MLP), which models pairwise interactions between data points. This design allows ITNet to adapt its processing behavior directly from the data, rather than relying on pre-defined architectural biases. The paper demonstrates that convolution, self-attention (including multi-head), and various forms of autoregressive recurrence (such as LSTM, GRU, S4, and Mamba) can all emerge as special cases of ITNet under appropriate parameterizations. This establishes ITNet as a universal approximator of continuous operators. To ensure practical applicability, the authors developed efficient computational techniques like tiled kernel fusion, importance-weighted Monte Carlo integration, and learned low-rank factorization. Benchmarking shows that a single ITNet architecture, with a shared operator and lightweight modality-specific encoders, performs comparably to or better than specialized baselines across diverse tasks including ImageNet-1K, GLUE, ModelNet40, VQA v2, and NLVR2.