FusionFlow: A New Paradigm for Cross-Domain Reasoning with Lightweight Transformer-GRU Fusion
Year: 2025
Type: dissertation
Abstract: This paper unveils FusionFlow, a groundbreaking paradigm architecture that redefines AI by enabling efficientcross-domain reasoning between vision and text. Developed by a solo researcher, FusionFlow achieves90% to 95% token accuracy (via beam search) and a BLEU score of 0.3443 on the KITTI dataset and Persona Chat plus Dayli Dialog,while requiring only 0.76GB of model memory and 1.46GB during inference—over 100× lessthan giants like GPT-4V. With an inference time of 4 seconds for 100 samples (batch size = 10) on anNVIDIA GeForce RTX 3050 with 8GB VRAM, this lightweight design leverages Conditional LayerNormalization and adaptive fusion for real-time edge deployment. FusionFlow slashes computationalcosts by 90-99% and energy use by 10-100×, offering a sustainable, inclusive path to advanced AI inautonomous systems, IoT, and beyond. This is a triumph of independent innovation over industry titans.
Author: Furlan, Rafael
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6968302582
Cites:
Cited by:
Related to: