Regular expressions for decoding of neural network outputs

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Neural networks trained with Connectionist Temporal Classification (CTC) are effective for tasks like handwritten text recognition.
• Decoding the output of CTC models can be computationally intensive, limiting real-world applications.

Purpose of the Study:

• To develop a computationally efficient decoder for CTC-trained neural networks.
• To leverage regular expressions and finite automata for improved handwritten text recognition.

Main Methods:

• Utilized regular expressions to model expected writing structures.
• Constructed a decoder based on the corresponding finite automata.
• Performed theoretical analysis to identify and eliminate redundant calculations.
• Introduced an approximation technique to significantly accelerate the decoding process.

Main Results:

• The proposed decoder demonstrates high efficiency compared to existing methods.
• An approximation strategy yields substantial speed-up in decoding.
• The approximation's failure (when regex doesn't match ground truth) is often inconsequential due to probability underestimation.

Conclusions:

• The developed decoder offers a practical and efficient solution for CTC-based handwritten text recognition.
• The method's efficiency and accuracy make it suitable for diverse applications, including information retrieval and full text recognition.