DODO: Discrete OCR Diffusion Models

Quick Overview

The DODO (Discrete OCR Diffusion) model successfully addresses the speed-accuracy trade-off in large-scale document processing by introducing a novel block-discrete approach that allows for parallel processing, achieving a 3x speed advantage over standard autoregressive models while maintaining high accuracy, especially when compared to models like DALL-E or Stable Diffusion which struggle with rigid, sequential tasks like OCR.

Key Points: DODO (Discrete OCR Diffusion) research from Technion and Amazon Web Services proposes a block-discrete method to solve the speed/accuracy trade-off in large-scale document processing. The DODO architecture successfully avoids the primary bottleneck of sequential autoregressive models by processing document blocks in parallel, resulting in a 3x speed multiplier over the baseline. The model achieves a Normalized Edit Distance (NED) score of 0.066 on OCR tasks, placing it in the high-precision realm, compared to the baseline autoregressive model's score of 0.093. DODO uses a discrete diffusion approach, which is typically associated with image generation, but applies it to OCR by breaking the document into blocks and processing them simultaneously. The inherent rigidity of OCR tasks (requiring strict sequential processing for elements like invoice numbers) makes them historically poor fits for general diffusion models, which DODO overcomes. The technique allows for caching of processed blocks, preventing recomputation of the history and enabling faster throughput, which is crucial for enterprise-scale operations. The paper frames the DODO approach as moving AI processing from an era of 'Typing AI' to 'Stamping AI' for document understanding.

Context: This video discusses a research paper titled 'DODO: Discrete OCR Diffusion Models,' a joint effort between the Technion Israel Institute of Technology and Amazon Web Services. The paper introduces a new architecture designed to improve the speed and accuracy of Optical Character Recognition (OCR) on large documents, specifically addressing the inherent sequential processing limitations found in traditional autoregressive models.