Fourier Vision-Language Fusion for Low-Complexity Multimodal Inference

The patent replaces the self-attention mechanism (a complex computational system with quadratic complexity) with a Fourier transform-based approach. The Fourier sublayer transforms the input embeddings into the frequency domain, performs simplified operations, and transforms back, achieving comparable integration accuracy with significantly reduced computational complexity that scales linearly rather than quadratically with input size.

The patent substitutes the parameter-heavy self-attention mechanism with a Fourier transform-based system that has far fewer trainable parameters. The Fourier sublayer uses fixed mathematical transforms (FFT and inverse FFT) combined with simple feed-forward networks, dramatically reducing the number of trainable parameters while preserving integration accuracy through the spectral domain processing.

The patent replaces the memory-intensive self-attention mechanism with a Fourier-based approach that processes embeddings in the frequency domain. This substitution allows the system to handle long text sequences and high-resolution images with linear rather than quadratic memory scaling, preventing memory exhaustion on devices with limited capacity.

The patent substitutes the computationally expensive self-attention operations with Fourier transform-based processing. The Fast Fourier Transform algorithm enables efficient frequency domain operations that are computationally lighter than quadratic self-attention, reducing processing time while maintaining integration performance through the spectral analysis of the embeddings.